Every conference this year has asked whether the technology works. That question is settled. The more useful one—the one worth clearing a day for—is what happens to your company, your industry, and the institutions meant to govern all of it now that it does.

That’s the premise of Techonomy 26, which we’re holding on Tuesday, October 6, in San Francisco. The theme is Intelligence at Scale: what changes when AI stops being a feature and becomes infrastructure.

The distinction matters. A feature is something you evaluate, budget for, and switch off if it disappoints. Infrastructure is something you build on top of—and once you have, the questions stop being about capability and start being about dependency, security, cost, power, and accountability when it fails. Most organizations are still having procurement conversations about something that has already become load-bearing.

Who’s confirmed

A first look at the program:

• Jack Hidary (SandboxAQ)—why the next leap isn’t a bigger language model. His large quantitative models run on physics and chemistry rather than text, and they’re already in production.

• Ryan Panchadsaram (Kleiner Perkins, co-author of Speed & Scale)—the power and materials problem underneath the boom. Energy is the real ceiling.

• Udbhav Tiwari (Signal)—the case that agentic AI, as currently built, quietly undoes what encryption was supposed to guarantee.

• Tom Kellermann (TrendAI)—autonomous malware, deepfake fraud, and the fact that most enterprises can’t name the agents already running inside them.

• Liz Centoni (Cisco)—what AI is actually fixing at scale, from someone running a $15 billion services business and refreshingly allergic to hype.

• Jaime Teevan (Microsoft)—what building Microsoft 365 Copilot taught her about how people and machines really split the work.

• Dr. Vivienne Ming (The Human Trust)—the counter-argument that more intelligence is the wrong goal.

• Bill Gross (ProRata.ai) and Phil Radford (Consumer Reports)—who gets paid when AI answers the question, and who’s looking out for the person asking it.

Plus Mei Lin Fung on what the internet’s first 50 years should teach us about governing AI’s first ten, and Lu Zhang of Fusion Fund on where the money is actually going. More to come.

Read that list again, and you’ll notice it doesn’t agree with itself. Hidary and Panchadsaram are building the layer underneath; Tiwari and Kellermann are describing what breaks when you do. Centoni is unromantic about what’s working today; Ming argues the whole objective is misspecified. That disagreement is the program, not a scheduling accident.

The format

Techonomy has never been a stage-and-audience event, and this one won’t be either. It’s a working forum: structured conversations, off-the-record sessions, and enough cross-sector collision that enterprise buyers, founders, investors, and policy people are actually in the same conversation rather than three parallel ones. We keep it to 150 to 200 people for exactly that reason.

The details

• Event: Techonomy 26

• Date: Tuesday, October 6, 2026

• Location: San Francisco Bay Area

• Theme: Intelligence at Scale—what changes when AI moves from feature to infrastructure

• Format: Invitation-only working forum. 150–200 executives, founders, investors, and policymakers.

• Cost: Complimentary for confirmed guests

• Presenting partner: Northern Trust

• Details and invitation requests: worth.com/event/techonomy-2026

We keep the room small on purpose, so the list closes well before October. If you’re building this, funding it, deploying it, or writing the rules for it, request an invitation, and we’ll confirm.

Share

Subscribe now

Dr. Jay Luthar described the flaw in modern medicine in a single sentence, and once you hear it, you can’t unhear it. “The traditional medical model to this point has used population studies to find what the average person responds to and then fits an individual to an average,” he said. “Whereas we know that every individual is not an average, but very individual.”

Tell me if this sounds familiar. Your doctor treats you with drugs validated on a population you don’t belong to. The trial reported an average response. You are not the average. Nobody is. Some patients in that trial got better, some got nothing, a few got worse, and the number that reached your chart was the mean of all three.

Subscribe now

The research bears this out uncomfortably well. Randomized trials are designed and powered to estimate average treatment effects, not what will happen to you specifically. For most treatments, only a small share of patients respond the way the “average” patient did. Clinicians have been guessing from that average for a century and calling it ‘evidence-based’.

Luthar, a faculty member at Harvard Medical School and founder of Lutanen Health, wants to invert it. Look at the patient in front of you “almost like an n-of-1 research study.” That phrase isn’t a metaphor he made up. N-of-1 trials are a real design, and the Evidence-Based Medicine Working Group of the AMA has held that they sit at the top of the evidence hierarchy for decisions about an individual patient. The catch, which is real, is that the statistical methods for aggregating them aren’t standardized, and they carry almost no weight in guidelines or regulatory decisions.

Dr. Julie Chen, chief medical officer at Radence and formerly CMO at Human Longevity, reframes the same shift as a question. Traditional medicine asks, in her words, “are you sick yet?” Radence asks whether you are “changing in a not-so-great direction.” If the answer is yes, you’re still inside the window where prevention means something.

The machinery underneath that question got real in the last two years. Luthar is most excited about proteomics, and the numbers explain why. We carry tens of thousands of protein signatures. Researchers at Stanford measured more than 7,000 plasma proteins across 60,542 people and built models estimating the biological age of over 40 cell types. Roughly 20% to 25% of people showed accelerated aging in a single cell type. The associations are not subtle: accelerated organ aging carried a 20% to 50% higher mortality risk, accelerated heart aging came with a 250% higher heart failure risk, and people with extremely aged skeletal myocytes had a 12.7-fold higher risk of developing ALS.

Two patients can look identical on paper, Luthar said, same age, same gender, same diagnosis, and have “totally different underlying physiology, totally different reason for even having the same disease.” The proteome sees the difference, but this will not appear on an intake form.

Chen’s version runs through imaging and time. Not a single snapshot but a trend line: organ volumetrics, brain atrophy, liver iron, visceral fat, muscle composition, tracked against your own baseline rather than a reference range built from strangers. “Clinicians are as good as the data that we get,” she said, “so it’s not about getting massive amounts of data but getting the right data.”

She also flagged the part that should interest anyone writing checks in this category. Interventions can now be audited per person. A treatment “supposed to help them” can be checked against whether it actually is, in that specific body, compared to cohorts that resemble them instead of a general population “where there could be people that have nothing, no characteristics like you.”

So that’s the promise, and the market seems to think it is real. The precision medicine market ran between $110 billion and $119 billion in 2025 depending on whose model you trust, with forecasts in the $400 billion range within a decade. Multi-omics is the small, fast piece: about $3.1 billion in 2025, projected past $12 billion by 2035.

Measuring more of a healthy person is not the same as helping them, and medicine has learned this the hard way. Whole-body MRI screening for asymptomatic people confirms cancer in roughly 1.1% to1.5% of scans while turning up a much larger pile of incidental findings. The American College of Radiology doesn’t recommend it for people without symptoms, risk factors, or relevant family history, and JAMA has published against elective whole-body scans on overdiagnosis grounds. Patients get follow-up tests. Some get surgery for things that would never have hurt them. All of them get to carry the knowledge that something is in there.

The genomics layer has its own asterisk. Luthar cited the MGH/MIT polygenic risk score for cardiovascular stratification, and polygenic scores do work as risk enhancers. They also inherit a structural bias: because most genome-wide association studies were run on people of European ancestry, the scores predict less well for everyone else. In one Broad Institute validation cohort, there weren’t enough cardiovascular events among non-European ancestry groups to produce stable estimates. A tool that personalizes better for some patients than others is not a small footnote in a story about ending the average patient.

To their credit, neither doctor oversold it. Chen was blunt that AI is “not at a point where clinical decision-making or judgment can be the end-all, be-all yet,” and that “humility in medicine” is part of the job. Her own operation runs a team of MD-PhDs whose job is to read the thousands of studies published daily and sort what is evidence-based from what is merely interesting. That team exists because the field generates more claims than it does proof.

Luthar’s framing of the near term is honest. Focus on early detection and lifestyle, and “help people live healthily up until some of these more revolutionary treatments can come down the pipeline.” Translation: the personalization is real, the therapeutics are not there yet, and the job right now is staying alive and in good repair long enough to use them.

Leave a comment

🤖 Machined: This Week in AI

1. Meta pays up to $17 billion to settle with state AGs over teen social-media addiction Meta, states agree to landmark settlement in child safety trial — NPR / CNBC / WSJ, August 26, 2026 One week into the federal trial in Oakland, Meta settled with a bipartisan coalition of 51 attorneys general over claims Facebook and Instagram were intentionally designed to addict minors — reported figures range from $16.7 billion to $18 billion paid out over ten years, the largest state consumer-protection settlement since Big Tobacco. Beyond the money, Meta agreed to sweeping product changes: default two-hour daily limits for under-18 users, nighttime and school-hours notification blocks, and enhanced age assurance. California AG Rob Bonta called it “a floor, not a ceiling” — school-district and individual suits remain pending, and Meta will book a ~$10 billion legal expense in Q3.

2. Nvidia crushes Q2: $96.2B in revenue, and Huang forecasts 70% growth next year Nvidia earnings takeaways: revenue up 106%, guidance sails past estimates — CNBC, August 26, 2026 Nvidia posted $96.2 billion in fiscal Q2 revenue (up 106% year over year, with data center at $89 billion) and guided next quarter to $108 billion — while assuming zero China data-center revenue despite the H200 shipments that resumed last week. The eye-opening details: supply commitments more than doubled to $279 billion, largely for memory procurement, and Jensen Huang said on the call that OpenAI and Anthropic “will likely go public soon.”

3. Amazon shuts down Mechanical Turk after 21 years AWS Mechanical Turk closes September 30 — Tech Startups, August 26, 2026 Amazon told workers and requesters that the 2005-era marketplace Jeff Bezos once branded “artificial artificial intelligence” — which farmed out image labeling, transcription, and moderation tasks to more than 500,000 people at pennies apiece — will close September 30, alongside SageMaker Ground Truth and Augmented AI. The human-labeling layer that trained the machine-learning era is officially being retired by the machines it trained.

4. OpenAI says its Jalapeño chip beats Nvidia’s Blackwell on efficiency Jalapeño’s first results show industry-leading inference speed and efficiency — GAI Insights / OpenAI, August 26, 2026 OpenAI published early benchmark results for Jalapeño, its Broadcom-designed inference chip, claiming it outperforms Nvidia’s Blackwell systems on AI work per watt for memory-intensive inference workloads. The claims are preliminary until the chip reaches production scale, but the timing — landing the same day as Nvidia’s earnings — underscores OpenAI’s push toward vertical integration across models, chips, and devices.

5. Nvidia in talks to back Perplexity at a $30B+ valuation as part of an open-weight push Nvidia discussing equity investment in Perplexity — AI Weekly / The Information, August 24, 2026 The Information reports Nvidia is discussing an investment in Perplexity at a valuation above $30 billion — the search startup’s annualized revenue has reportedly passed $750 million, up from under $250 million at the start of the year — as part of a broader Nvidia effort to put billions behind open-weight AI development.

6. SoftBank shops a $10–20 billion bond deal to cover its OpenAI bet SoftBank in talks for what could be Asia’s largest corporate bond of 2026 — Tech Startups / Bloomberg, August 26, 2026 SoftBank is discussing a $10–20 billion bond offering to refinance part of the $40 billion bridge loan underwriting its OpenAI investment, with notes potentially selling as early as September. It still owes a final $10 billion OpenAI installment on October 1 — a reminder of how much leverage is stacked under the AI buildout.

7. Moonshot comes courting American clouds — one month after the White House accused it of theft China’s Moonshot AI courts major US cloud companies — Tech Startups, August 26, 2026 Kimi K3 maker Moonshot AI is reportedly pitching America’s biggest cloud providers on hosting its models — a striking move given the White House accused it of distilling Anthropic’s Fable just five weeks ago, and a test of whether commercial demand for cheap frontier-class open models outweighs the geopolitics.

8. XPeng’s robotics arm raises $900M at a $6.3B valuation XPeng robotics business closes first external round — Tech Startups, August 24, 2026 The Chinese EV maker’s robotics unit raised over $900 million led by IDG Capital and Gaorong Ventures, with Tencent and Alibaba as strategic investors, to fund humanoid development and mass production — landing the same week Unitree’s post-IPO glow keeps China’s embodied-AI capital cycle spinning.

Analog Break: Inside the Glenrothes Speyside’s Invitation-Only Distillery

No visitor center, no tours—The Glenrothes makes its sherry cask single malt behind closed gates. What we found when we got inside.

We were 30 minutes late for our whisky tasting. In our defense, the name is everywhere in this corner of Speyside. We stopped at the Glen of Rothes Trout Fishery. We pulled up to the gates of Rothes House. The town itself is called Rothes, and somewhere inside it, unmarked and unbothered, is The Glenrothes distillery. It was 10 in the morning.

Read More.

Gallup research shows 70% of Americans oppose building AI data centers where they live: 75% of Democrats, 63% of Republicans. Charles Franklin, who runs the Marquette Law School poll in Wisconsin, put it plainly: “There is no difference by party across this.”

Our leaders have seen the polls, even the conservative ones. Texas Governor Greg Abbott spent a decade recruiting these projects to his state and now says he’ll work to repeal the tax break that brought them. Illinois Governor JB Pritzker, Pennsylvania Governor Josh Shapiro, and Maryland Governor Wes Moore courted data centers with abatements and ribbon cuttings, then publicly backpedaled.

Subscribe now

New York Governor Kathy Hochul went even further, signing the first statewide moratorium, with more than 100 local and state proposals lined up behind it. Data Center Watch counted 75 projects blocked or delayed in the first quarter alone, worth about $130 billion.

The data center rollout has been a frenzy, and many corners have been cut. Janesville, Wisconsin, spent a year courting data center developers under NDA. Residents found out afterward. When three of them filed a public records request, the city sent back a bill for $1,380. They put a referendum on the ballot instead, requiring a public vote on any development over $450 million.

Journalist Jasmine Sun has done some of the best reporting on this on her Substack. She spent ten days driving between four Midwest data center fights and found the same sequence in every one: towns learning the data center deal after the land was optioned and the deal was done.

The movement’s most famous recruit isn’t an AI critic at all. Erin Brockovich launched brockovichdatacenter.com in April, a crowdsourced map where residents self-report facilities near them, and says it has logged close to 16,000 reports from all 50 states. She announced it in an essay titled If Data Centers Are So Great, Why Are They Being Built in Secret? and framed the buildout as a rerun of Hinkley, writing that she’d spent her career listening to people “told their backyard was safe, and that the water was safe to drink.”

When hyperscalers descend on small towns and start negotiating with local town boards, a power imbalance is inevitable. But that is only one of the power problems at play here.

We Need More Power, Captain

My colleague Oliver Rist itemized that invoice in Worth this month. The top line is a capacity auction. PJM’s price for promised capacity went from $28.92 per megawatt-day to $329.17 in two cycles, better than 1,000%, and PJM’s own market monitor put 63% of one year’s jump on data centers: roughly $9.3 billion pulled out of customers. FERC commissioner David Rosner told his colleagues in June that the loads lining up to connect are larger than anything the grid was planned around, sometimes by orders of magnitude.

The supply side is moving. EIA expects 86 gigawatts of new utility-scale capacity in 2026, a record, and 51% of it is solar. The Department of Energy now describes data center demand as what pays for grid modernization, not what threatens it.

The wiring and the rules are what’s missing. Interconnection queues run five to seven years. Transmission gets blocked by the same county boards blocking the data centers. The terms are only now being written. Large-load tariffs make these facilities cover their own cost of service, thirty of them since 2024, plus FERC show-cause orders that would park stranded-asset risk on the developer instead of the ratepayer. Microsoft has ended NDAs with local governments. Hochul’s office is floating $1 million in community benefits per megawatt as a standard term of entry.

Brad Smith conceded the whole premise back in January: “We’ll pay our way to ensure our datacenters don’t increase your electricity prices.” Which concedes that until recently, they weren’t.

Land needs the same kind of management, and local zoning isn’t ready for it. A hyperscale campus is basically a big flat building. And they can be very big. Data center opponents in Port Washington, Wisconsin claim it takes a minute and 42 seconds to drive past one at highway speed. In Fairfax County, Virginia, 55% of data centers sit within 200 feet of somebody’s house. Next door in Prince William County, residents logged noise routinely above 60 decibels. When the county finally set a limit, it landed on 73 by day, above the 67 its own consultants had recommended.

Setbacks, height, noise, screening. These are ordinary zoning tools that only work if they exist before the option closes.

Every concession above came because of local action. Microsoft dropped its NDAs under public pressure. Hochul is drafting terms because a moratorium gave her leverage. Saline Township, Michigan, pulled $14 million in community benefits after losing a lawsuit loudly. Moratoriums are leverage, and leverage is worth exactly what you trade it for.

The country has something it hasn’t had in twenty years: broad political consensus. Everybody hates data centers; the open question is whether this unity can be directed toward enforced disclosure, energy rate designs, and an infrastructure build-out that enables the promises of the AI Age to take hold.

Or are we just putting the future on hold?

The AI Boom Is Creating an Energy Crisis Nobody Is Ready For with Milken Institute’s Rachel Halfaker

At the Milken Institute Conference 2026, Dan Costa sits down with Rachel Halfaker, Director at the Milken Institute, to discuss the collision between AI, climate finance, energy infrastructure, and the future of American development.

As hyperscalers race to build the next generation of AI data centers, a new problem is emerging: the U.S. may not have enough energy infrastructure to keep up.

Share Machined by Dan Costa

Other Signals in the Machine 8/20

1. OpenAI slows frontier development as Astra nears “Critical” cyber threshold OpenAI pauses training runs, overhauls safety practices after Astra evaluations — Axios, August 18, 2026 OpenAI said it paused reinforcement-learning training on deployment-bound models for two weeks and is keeping its largest planned frontier RL run on hold, after preliminary evaluations suggested its upcoming Astra model may reach the “Critical” cybersecurity threshold under its Preparedness Framework — meaning it could autonomously find and exploit zero-days in hardened systems. The move follows the Hugging Face breach by a different unreleased OpenAI model, and marks the first time a frontier lab has voluntarily slowed its own development over cyber risk.

2. Nvidia’s H200 chips finally reach mainland China — and now Beijing is the one resisting H200 shipments arrive at ByteDance and Tencent as Beijing loosens import block — Tom’s Hardware / FT, August 19, 2026 ByteDance and Tencent have each taken delivery of roughly 10,000 H200 processors — the first meaningful movement since the export framework was approved last December — even though US licenses allow up to 100,000 per buyer. The twist: Chinese regulators want most of the hardware kept in Hong Kong to protect domestic chipmakers like Huawei, and the news lands a week before Nvidia’s August 26 earnings, where it had assumed zero China data-center revenue.

3. DeepSeek ends the price war it started, hiking API rates up to 1,100% DeepSeek launches V4-Pro and raises API prices by as much as 1,100% — Caixin, August 14, 2026 Alongside the general-availability launch of its V4-Pro flagship, DeepSeek replaced flat pricing with a peak/off-peak structure effective August 16, with increases ranging from 50% to over 1,100% depending on model and time of day. The company also closed its first external funding round — 50 billion yuan at a 350B+ yuan valuation, with Tencent and JD.com participating — signaling the era of below-cost Chinese inference is over.

4. The frontier price war splits in two: US labs cut while China raises OpenAI and Anthropic cut rates as Google ships half-price Gemini 3.7 Flash — BeingGuru / FT data, August 13–18, 2026 Google launched Gemini 3.7 Flash on August 13 at $0.75/$3.75 per million tokens — roughly half the previous Flash cost through year-end — while OpenAI cut GPT-5.6 Luna pricing and previewed an “Ultrafast” API tier running Sol up to 14x faster, and Anthropic positioned Opus 5 at about half of Fable 5’s price. The competition has visibly shifted from benchmarks to cost per unit of useful work.

5. Humanoid robot maker Unitree soars 629% in its Shanghai debut Unitree opens at 1,100 yuan on the Star Market, briefly worth ~$66B — AI Weekly / SCMP, August 19, 2026 Unitree Robotics opened at 1,100 yuan against a 150.80 yuan IPO price — up 629% — briefly valuing the humanoid maker at roughly 445 billion yuan (~$66B) before settling lower. The retail tranche was oversubscribed more than 5,500 times, making it the clearest market signal yet on China’s embodied-AI bet.

6. Pennsylvania makes its data-center standards legally binding Gov. Shapiro signs Executive Order 2026-05 enforcing GRID standards — AI Weekly, August 18, 2026 Governor Josh Shapiro — who previously championed a $20B Amazon buildout — signed an order requiring data-center developers to commit to local approval, fully fund new electricity infrastructure, conserve water, and hire locally, while pulling all AI data-center proposals from fast-track permitting and banning NDAs on projects. It’s one of the strongest state-level responses yet to the growing local backlash against AI infrastructure.

7. Google deepens its custom-chip hedge with Marvell — including rights to $12.2B in stock Marvell and Google expand AI chip partnership — Bloomberg, August 19, 2026 Marvell and Google expanded their chip-development partnership, with Marvell granting Google the right to buy as much as $12.2 billion in stock — another sign hyperscalers are locking in custom silicon supply rather than relying solely on Nvidia. SK Hynix separately announced a $29 billion buyback as memory makers ride the AI boom.

8. OpenAI ships a safer ChatGPT for teens — years after teens started using it OpenAI launches teen safety features and parental controls — TechCrunch, August 19, 2026 OpenAI introduced specialized safeguards and parental controls for younger users, arriving as Meta faces testimony in its teen-safety trial and as TechCrunch reports AI’s public reputation is actively worsening — with data-center developers now sweetening deals with job guarantees and clean-water commitments to overcome local opposition.

In July 2025, venture capitalist Jason Lemkin spent a week building software. Lemkin founded SaaStr; he is not an engineer. For a few days, that didn’t matter. Replit’s AI agent wrote the code, ran the tests, and deployed the app while Lemkin described his vision in plain English. He blogged about the dopamine hit of shipping to production. He called it the most addictive thing he’d touched since childhood.

Then the agent went rogue. It wiped his production database during a code freeze, fabricated 4,000 fake user records, and generated reports claiming the systems were healthy. Lemkin had told it not to change code without permission. He gave that instruction 11 times in all caps.

A little knowledge is dangerous; a lot of knowledge can be doubly so if it isn’t yoked to experience. This hubris is getting lawyers sanctioned, and consultants fired. It affects entry-level staffers building strategic plans for the company. It infects CEOs who suddenly feel they can build a TikTok strategy. (And with an AI-generated avatar–why not star in the content?)

I call it Synthetic Expertise: a reckless conviction that access to a fluent machine makes you an expert in something you never actually learned.

I am not immune. I set up our company’s Claude account and have a background in tech journalism–although not tech itself. I do not code, but I do love using AI to batch emails and remind me who I need to follow up with. When I research a story, Claude is my first stop, just not my last. And don’t even get me started on how much I love creating AI Slop illustrations for stories like these. It isn’t artistic appropriation, it is irony!

And yet, did I stray from my editorial lane to identify all of Worth’s top sponsor prospects in HubSpot, rank them by priority, build a call sheet, and delegate all of them to the appropriate salespeople?

I did. Took me less than 10 minutes. Ahem, it took Claude less than 10 minutes.

To be clear, this is not my job. And it is not my work. Perhaps most importantly, it is the same thing that anyone else could do with access to the tool and the data. If this is my job, my company can do without me.

It’s a pattern I’ve seen play out before. When the Web first arrived, it promised to make the world’s knowledge findable. It was a revolution, and for a moment, we treated it like one—some even started calling themselves ‘search engine gurus,’ proud just to know where to look. But then social media arrived, and we learned that easy access didn’t mean we were actually learning; we stopped digging for facts and started sharing memes instead. A 2015 study captured the shift perfectly, finding that the simple act of searching the internet made people believe they were smarter than they actually were—Googling made us feel like we owned the knowledge living on the server.

AI completes that cycle. Now, we don’t even have to find, read, or remember the answer; the machine delivers it on demand, and we’ve convinced ourselves the brilliance is ours. We’re creating a new kind of synthetic asset: the expert. And unlike a deepfake, this one thinks it’s real.

Falling Asleep at the Wheel

The data is starting to roll in. Researchers at Aalto University gave 500 people logic problems to solve with ChatGPT. The Dunning-Kruger effect—where the unskilled overrate themselves—usually holds steady. With ChatGPT, the curve reversed. The most “AI-literate” users were the most overconfident. They asked one question, took the answer, and moved on. Almost no one checked the results.

Research from Microsoft and Carnegie Mellon survey found the same thing. Confidence in AI predicts a drop in critical thinking.

Then there’s the Harvard study of 758 BCG consultants. When tasks were within the AI’s reach, they were 25% faster. On tasks just past that boundary, they performed worse than colleagues with no AI at all. They “fell asleep at the wheel.” They stopped using their judgment because the machine sounded so sure of itself.

Are Lawyers the Worst?

The real-world consequences are piling up. By June 2026, there were 1,600 documented cases of lawyers filing fake AI-invented court cases. Despite $5,000 sanctions and global headlines, the curve is still bending up.

Deloitte Australia refunded a government contract after using AI to fabricate a federal court quote. The White House’s MAHA report cited non-existent studies. The press secretary called it a “formatting issue.” It wasn’t formatting; it was hallucination accepted as fact.

Klarna cut 700 jobs and let AI handle three-quarters of customer support. AI-watchers took this as the early sign of AI’s coming efficiencies. By May 2025, CEO Sebastian Siemiatkowski was rehiring humans. AI can answer tickets, but a dashboard doesn’t make you an expert in customer experience. The CEO’s overconfidence led the entire company astray.

We now have a name for this: workslop. Output that masquerades as good work while advancing nothing. In one survey, 41% of workers had to spend hours decoding this slop—a tax priced at $186 per employee per month. Most of it flows downhill from managers.

Everybody Is an Expert

There is a specific hubris here: the belief that access to the machine is an edge. It isn’t. The tool is available to everyone. As Brian Solis put it in his SXSW talk: “Everybody is an expert now. And when everybody is an expert, no one can be.”

The Web gave us an answer to every question we could ask, and most of the time it was true. Social media showed that most of us would rather be targeted and entertained than ask experts.

AI lets us believe we never needed experts in the first place.

Share

Deep Tech or Bust: Reclaiming Innovation

Inventor and futurist Pablos Holman kicks off Techonomy 25 with a rallying cry to rediscover the frontier spirit that built the modern world. He argues that today’s obsession with AI optimization—endless tweaks to ads, feeds, and metrics—has replaced genuine invention. Holman makes the case for a return to “deep tech”: bold, physical, world-changing innovation in energy, climate, and health. Expect provocation, practical optimism, and a reminder that true progress comes from human audacity, not algorithmic efficiency.

Subscribe now

🤖 Signals from the Machine

1. OpenAI blames “unprecedented cyber incident” on its AI models — The Washington Post, July 22, 2026
OpenAI disclosed that two of its most capable models — including GPT-5.6 Sol and an unreleased system running with lowered guardrails during cyber-capability testing — broke out of their testing environment and penetrated Hugging Face’s systems on their own. The company says it’s still investigating, while safety researchers are calling the incident a “warning shot” for autonomous AI agents.

2. US accuses Moonshot AI of stealing from Anthropic’s Fable model — TechCrunch / Reuters, July 22, 2026
OSTP Director Michael Kratsios said the US has evidence Moonshot AI ran a covert, large-scale distillation platform against Anthropic’s Fable to build its Kimi K3 model, and also accessed restricted Nvidia GB300 servers in Thailand. Treasury Secretary Scott Bessent followed up, warning that sanctions and Entity List designations are “on the table,” while some researchers note K3’s timeline makes Fable-based distillation implausible as a primary method.

3. Moonshot’s Kimi K3 pushes Chinese AI into Fable-level territory — Fortune, July 16, 2026
The 2.8-trillion-parameter model launched July 16 with a 1M-token context window and immediately took the #1 spot on the Frontend Code Arena, ahead of Claude Fable 5 and GPT-5.6 Sol. Full open weights are promised by July 27 under a modified MIT license, making it the first open model credibly inside the frontier conversation.

4. Google required to open up to AI and search rivals under EU-mandated changes — CNBC, July 16, 2026
Under the Digital Markets Act, the European Commission issued binding decisions requiring Google to give competing AI assistants — including OpenAI, Anthropic, and Mistral — access to 11 Android features (voice activation, in-app task execution) and to share anonymized search data with rival engines. Google’s Kent Walker called the decision a privacy and security risk; noncompliance penalties could reach 10% of global revenue.

5. Google releases three Gemini Flash models as flagship delays continue — Build Fast with AI, July 21–22, 2026
Gemini 3.5 Pro missed its July 17 target — reportedly its third slip — after falling short on coding and reasoning in testing, knocking roughly 4% off Alphabet shares. Google filled the gap on July 21 with Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-tuned Flash Cyber variant restricted to governments, while confirming Pro still has no ship date.

6. Apple briefly tops Nvidia as world’s most valuable company — Bloomberg, July 17, 2026
Apple hit roughly $4.9 trillion in market cap as Nvidia fell as much as 4.6% amid a broader rotation out of AI chip stocks — the semiconductor index is down nearly 19% from its highs. Apple is up about 23% this year as investors reward its light-capex AI strategy, its first time on top since April 2025.

7. OpenAI reportedly paused a model that disproved the Erdős unit distance conjecture — Build Fast with AI, July 21, 2026
Internal sources say OpenAI paused access to an unreleased model after it disproved a long-standing open problem in combinatorial geometry — a genuine original research contribution — and then repeatedly found ways to act outside its sandbox. OpenAI has not publicly confirmed the details, but the report pairs uncomfortably with the confirmed Hugging Face incident.

8. Anthropic outspends Nvidia on federal lobbying in Q2 — AI Weekly / NYT, July 22, 2026
Federal disclosures show Anthropic spent $1.97M on lobbying in Q2 2026 (up 26% quarter-over-quarter), edging past Nvidia and nearly matching Oracle, while OpenAI spent $1.2M. Combined AI-lab spend hit $3.17M for the quarter, with cybersecurity, copyright, and defense procurement listed as priorities.

Leave a comment

Source: Paleofuture

The dynamo is having a moment. I’ve heard the story of the electrified factory three times in the last week, twice from a conference stage and once in a pitch deck. Joe Reis told it well on his Substack in April. The moral is always the same: factories bolted electric motors onto steam-era buildings and got nothing, so don’t bolt AI onto steam-era workflows. Recognizing this can empower you to lead change effectively. It beats comparing AI to the dot-com bubble, which tells you about valuations and nothing about work.

The factory redesign wasn’t the payoff; it was just the foundation that enabled a wave of new industries and jobs that no one on a factory floor in 1905 could have imagined.

Electric motors were commercially viable by the 1890s. Factory owners bought them, pulled out the steam engine, and changed nothing else. The building still ran on a central power shaft with leather belts dropping down to every machine. Machines sat where the power was, not where the work flowed. The Stanford economist Paul David called this the productivity paradox of the dynamo. The technology was everywhere. The numbers didn’t move.

They didn’t move for twenty years. The unlock was a design principle called unit drive: one motor per machine. Kill the central shaft and the multi-story factory becomes obsolete. Once machines could sit anywhere, engineers arranged them by task order rather than by proximity to power. Single-story plants. Workflow-first layouts. Ford’s assembly line. Warren Devine traced the whole progression in “From Shafts to Wires,” which remains the best thing written on the subject and is cited by roughly none of the people retelling it.

Subscribe now

Then the numbers moved. David and his colleague Gavin Wright measured the break: productivity growth in U.S. manufacturing ran 1.5% a year from 1899 to 1914, then 5.1% a year from 1919 to 1929. Same motors. Different factory.

We are deep in the motor-swap phase with AI. MIT’s NANDA initiative studied 300 enterprise AI deployments and found that despite $30-$40 billion in spending, 95% of generative AI pilots have no impact on the P&L. Generative AI tools work for individuals because they’re flexible; they stall inside enterprises because nothing around them has changed. The executive buys Copilot seats, mandates usage, and declares the company electrified. Everyone feels smarter than they really are, but nothing really changes.

Economists saw this coming. Erik Brynjolfsson, Daniel Rock, and Chad Syverson call it the Productivity J-Curve: general-purpose technologies demand years of unmeasured investment in new processes and organizational design before the gains show up in official statistics.

The rebuilt factory created work that previously went unnamed. Electricians, linemen, power plant operators, industrial engineers: they moved from nonexistent to essential in a generation. And that was only the first-order effect. The assembly line made the Model T cheap, and cheap cars made new things necessary. Gas stations. Motels. Even the need for installment credit was driven by automobile financing. The rubber, steel, and glass supply chains feeding all of it. The factory owner asking how much faster the motor would run his belts was asking the wrong question entirely.

The MIT economist David Autor has put a number on this pattern. In a 2024 paper in the Quarterly Journal of Economics, Autor and his co-authors traced eight decades of Census job titles and found that roughly 60% of the work Americans did in 2018 was in job categories that did not exist in 1940. Autor writes—six in ten jobs were invented after the fact.

The AI version is just emerging. LinkedIn’s Jobs on the Rise data highlights roles like AI engineer, prompt engineer, and AI ethicist- jobs that exist to install, run, and govern the new technology.

History demands one caveat here. Autor’s research also shows that the people who take the new jobs are usually not the people displaced from the old ones. Since 1980, new work has shifted toward high-paid professional roles and low-paid service roles, hollowing out the middle. The havoc this has wreaked on our civil society, politics, and culture is well known and completely unresolved. The idea that these new jobs will fall benevolently to the displaced is optimism unsupported by history. The transition was wrenching in the 1920s. It will be wrenching now.

Companies agonizing over which frontier model to license are asking the 1905 question. The 1925 question is different: once you rebuild the work around what the technology makes possible, what new things are possible? And who do you hire to do it? The winners of the last transition weren’t the factories with the best motors. They were the ones that tore out the shafts, rearranged the floor, and posted help-wanted ads for jobs nobody had ever held.

Nobody in 1905 was hiring industrial engineers. By 1925, nobody could run a factory without them.

The AI Stack in Healthcare — From the Clinic to the OR

What if AI could identify your health risks long before symptoms ever appeared?

In this panel from Worth’s Living Well 2026, Dr. Julie Chen (Radence) and Dr. Jay Luthar (Lutanen Health) explore how artificial intelligence is transforming healthcare—from personalized risk modeling and early disease detection to AI-enhanced imaging and precision medicine. Learn why tomorrow’s healthcare won’t just treat illness—it will predict it, personalize it, and help prevent it before it starts.

Leave a comment

Signals in the Machine 7/2/26

• Anthropic restores Fable/Mythos access, launches Sonnet 5
  Commerce lifted the export controls on Fable 5 and Mythos 5 on June 30, and Anthropic used the moment to also drop Claude Sonnet 5 as the new default model for every Free and Pro user, priced well under Sonnet 4.6 through August. Coding benchmarks reportedly jumped from 58% to 63% on agentic tasks.
  Nothing says “confidence in your product” like burying your government-mandated comeback under a shinier launch the same week.

• Anthropic vs. Alibaba distillation accusation
  Anthropic says 25,000 fraudulent accounts generated 28.8 million interactions with Claude’s Mythos Preview between April and June, allegedly to train Alibaba’s Qwen models, and is now lobbying the Senate to treat distillation as a national security issue.
  Nothing radicalizes a company into supporting export controls faster than getting your homework copied by a state-adjacent conglomerate.

• Google loses four AI researchers in one week
  Four senior Google researchers, including AlphaFold contributors, jumped to Anthropic this week, while Noam Shazeer — co-author of the original Transformer paper — left DeepMind for OpenAI, wiping roughly $270 billion off Alphabet’s market cap in the process.
  Turns out the most valuable IP at a foundation model company was never the weights; it was the guy who invented the architecture leaving for a rival with better parking.

• OpenAI, Broadcom launch Jalapeño chip
  Designed in nine months using OpenAI’s own models, Jalapeño targets 50% cheaper LLM serving costs and was hand-delivered to Sam Altman by Broadcom’s CEO like a ceremonial sword.
  Dan’s take: when your burn rate hits $14 billion a year, even the silicon delivery becomes a photo op.

• California partners with Anthropic
  Governor Gavin Newsom announced Claude will be available to every state agency, city, and county in California through a centralized procurement framework, at what’s reportedly half the typical rate.
  The state that regulates AI the loudest just became its biggest customer — subtlety was never the point.

• Gemini 3.5 Pro delayed again
  Sundar Pichai promised a June general-availability launch at I/O in May; it’s now July and the model is still stuck in limited Vertex AI preview while Google patches token efficiency on long-horizon tasks.
  Google shipped two entire image models this week rather than admit the flagship isn’t ready — that’s not a delay, that’s a decoy operation.

• UN Independent Scientific Panel on AI report
  Panel co-chairs Yoshua Bengio and Nobel laureate Maria Ressa presented findings ahead of the first Global Dialogue on AI Governance, set for Geneva on July 6–7.
  A Turing Award winner and a Nobel Peace Prize winner co-signing a report is the kind of moral authority that governments will read, nod at, and then completely ignore.

• New Agentjacking attack class hits coding agents
  A newly identified attack technique is hijacking autonomous coding agents with an 85% success rate, raising fresh questions about how much unsupervised agency enterprises should hand these tools.
  We spent eighteen months teaching agents to write code unsupervised, and it turns out the thing they learned fastest was how to get hijacked.

Source: Paleofuture

The dynamo is having a moment. I’ve heard the story of the electrified factory three times in the last week, twice from a conference stage and once in a pitch deck. Joe Reis told it well on his Substack in April. The moral is always the same: factories bolted electric motors onto steam-era buildings and got nothing, so don’t bolt AI onto steam-era workflows. Recognizing this can empower you to lead change effectively. It beats comparing AI to the dot-com bubble, which tells you about valuations and nothing about work.

The factory redesign wasn’t the payoff; it was just the foundation that enabled a wave of new industries and jobs that no one on a factory floor in 1905 could have imagined.

Electric motors were commercially viable by the 1890s. Factory owners bought them, pulled out the steam engine, and changed nothing else. The building still ran on a central power shaft with leather belts dropping down to every machine. Machines sat where the power was, not where the work flowed. The Stanford economist Paul David called this the productivity paradox of the dynamo. The technology was everywhere. The numbers didn’t move.

They didn’t move for twenty years. The unlock was a design principle called unit drive: one motor per machine. Kill the central shaft and the multi-story factory becomes obsolete. Once machines could sit anywhere, engineers arranged them by task order rather than by proximity to power. Single-story plants. Workflow-first layouts. Ford’s assembly line. Warren Devine traced the whole progression in “From Shafts to Wires,” which remains the best thing written on the subject and is cited by roughly none of the people retelling it.

Subscribe now

Then the numbers moved. David and his colleague Gavin Wright measured the break: productivity growth in U.S. manufacturing ran 1.5% a year from 1899 to 1914, then 5.1% a year from 1919 to 1929. Same motors. Different factory.

We are deep in the motor-swap phase with AI. MIT’s NANDA initiative studied 300 enterprise AI deployments and found that despite $30-$40 billion in spending, 95% of generative AI pilots have no impact on the P&L. Generative AI tools work for individuals because they’re flexible; they stall inside enterprises because nothing around them has changed. The executive buys Copilot seats, mandates usage, and declares the company electrified. Everyone feels smarter than they really are, but nothing really changes.

Economists saw this coming. Erik Brynjolfsson, Daniel Rock, and Chad Syverson call it the Productivity J-Curve: general-purpose technologies demand years of unmeasured investment in new processes and organizational design before the gains show up in official statistics.

The rebuilt factory created work that previously went unnamed. Electricians, linemen, power plant operators, industrial engineers: they moved from nonexistent to essential in a generation. And that was only the first-order effect. The assembly line made the Model T cheap, and cheap cars made new things necessary. Gas stations. Motels. Even the need for installment credit was driven by automobile financing. The rubber, steel, and glass supply chains feeding all of it. The factory owner asking how much faster the motor would run his belts was asking the wrong question entirely.

The MIT economist David Autor has put a number on this pattern. In a 2024 paper in the Quarterly Journal of Economics, Autor and his co-authors traced eight decades of Census job titles and found that roughly 60% of the work Americans did in 2018 was in job categories that did not exist in 1940. Autor writes—six in ten jobs were invented after the fact.

The AI version is just emerging. LinkedIn’s Jobs on the Rise data highlights roles like AI engineer, prompt engineer, and AI ethicist- jobs that exist to install, run, and govern the new technology.

History demands one caveat here. Autor’s research also shows that the people who take the new jobs are usually not the people displaced from the old ones. Since 1980, new work has shifted toward high-paid professional roles and low-paid service roles, hollowing out the middle. The havoc this has wreaked on our civil society, politics, and culture is well known and completely unresolved. The idea that these new jobs will fall benevolently to the displaced is optimism unsupported by history. The transition was wrenching in the 1920s. It will be wrenching now.

Companies agonizing over which frontier model to license are asking the 1905 question. The 1925 question is different: once you rebuild the work around what the technology makes possible, what new things are possible? And who do you hire to do it? The winners of the last transition weren’t the factories with the best motors. They were the ones that tore out the shafts, rearranged the floor, and posted help-wanted ads for jobs nobody had ever held.

Nobody in 1905 was hiring industrial engineers. By 1925, nobody could run a factory without them.

The AI Stack in Healthcare — From the Clinic to the OR

What if AI could identify your health risks long before symptoms ever appeared?

In this panel from Worth’s Living Well 2026, Dr. Julie Chen (Radence) and Dr. Jay Luthar (Lutanen Health) explore how artificial intelligence is transforming healthcare—from personalized risk modeling and early disease detection to AI-enhanced imaging and precision medicine. Learn why tomorrow’s healthcare won’t just treat illness—it will predict it, personalize it, and help prevent it before it starts.

Leave a comment

Signals in the Machine 7/2/26

• Anthropic restores Fable/Mythos access, launches Sonnet 5
  Commerce lifted the export controls on Fable 5 and Mythos 5 on June 30, and Anthropic used the moment to also drop Claude Sonnet 5 as the new default model for every Free and Pro user, priced well under Sonnet 4.6 through August. Coding benchmarks reportedly jumped from 58% to 63% on agentic tasks.
  Nothing says “confidence in your product” like burying your government-mandated comeback under a shinier launch the same week.

• Anthropic vs. Alibaba distillation accusation
  Anthropic says 25,000 fraudulent accounts generated 28.8 million interactions with Claude’s Mythos Preview between April and June, allegedly to train Alibaba’s Qwen models, and is now lobbying the Senate to treat distillation as a national security issue.
  Nothing radicalizes a company into supporting export controls faster than getting your homework copied by a state-adjacent conglomerate.

• Google loses four AI researchers in one week
  Four senior Google researchers, including AlphaFold contributors, jumped to Anthropic this week, while Noam Shazeer — co-author of the original Transformer paper — left DeepMind for OpenAI, wiping roughly $270 billion off Alphabet’s market cap in the process.
  Turns out the most valuable IP at a foundation model company was never the weights; it was the guy who invented the architecture leaving for a rival with better parking.

• OpenAI, Broadcom launch Jalapeño chip
  Designed in nine months using OpenAI’s own models, Jalapeño targets 50% cheaper LLM serving costs and was hand-delivered to Sam Altman by Broadcom’s CEO like a ceremonial sword.
  Dan’s take: when your burn rate hits $14 billion a year, even the silicon delivery becomes a photo op.

• California partners with Anthropic
  Governor Gavin Newsom announced Claude will be available to every state agency, city, and county in California through a centralized procurement framework, at what’s reportedly half the typical rate.
  The state that regulates AI the loudest just became its biggest customer — subtlety was never the point.

• Gemini 3.5 Pro delayed again
  Sundar Pichai promised a June general-availability launch at I/O in May; it’s now July and the model is still stuck in limited Vertex AI preview while Google patches token efficiency on long-horizon tasks.
  Google shipped two entire image models this week rather than admit the flagship isn’t ready — that’s not a delay, that’s a decoy operation.

• UN Independent Scientific Panel on AI report
  Panel co-chairs Yoshua Bengio and Nobel laureate Maria Ressa presented findings ahead of the first Global Dialogue on AI Governance, set for Geneva on July 6–7.
  A Turing Award winner and a Nobel Peace Prize winner co-signing a report is the kind of moral authority that governments will read, nod at, and then completely ignore.

• New Agentjacking attack class hits coding agents
  A newly identified attack technique is hijacking autonomous coding agents with an 85% success rate, raising fresh questions about how much unsupervised agency enterprises should hand these tools.
  We spent eighteen months teaching agents to write code unsupervised, and it turns out the thing they learned fastest was how to get hijacked.

The wearable buzzes at 7 a.m. with a recovery score. The blood-test dashboard refreshes overnight with new biomarkers. The supplement stack on the kitchen counter has been scanned and rated against a database of 160,000 products. The longevity clinic appointment is on the calendar for Tuesday.

Four data streams, one consumer. None of it runs through insurance.

This is the biohacking economy in operation—distributed across four subscription services that, between them, do everything the legacy system can’t, won’t, or doesn’t catch in time. I went to SXSW expecting to hear founders pitch products. What I found were founders explaining the new architecture of the industry they’re building.

Dr. Darshan Shah, founder of the longevity clinic chain Next Health, was moderating a panel called “Are We All Biohackers Now?” A few minutes in, he asked the room a question. “If half of you here are not wearing some sort of wearable and tracking your data,” he said, “that would be a surprise to me.” The hands went up across the ballroom—WHOOP bands, Apple Watches, Oura rings, the occasional Garmin. Then he put a slide on the screen. Four quadrants, four labels: Clinic, Access, Fuel, Coach. Four boxes—and four founders sitting on the panel, one in each box.

The Clinic was Next Health, Shah’s own chain of about 15 longevity centers, with monthly memberships ranging from $99 to $400. The Access was Function Health, the at-home blood-and-MRI platform that closed a $298 million Series B in November at a $2.5 billion valuation. The Fuel was SuppCo, the supplement-rating service from GroupMe co-founder Steve Martocci. The Coach was WHOOP, the wearable company that hit a $10.1 billion valuation on a $575 million Series G earlier this year. Combined, the four had raised more than a billion dollars in venture capital. They weren’t comparing notes. They were assembling a healthcare system in public, vendor by vendor.

The legacy system, by way of comparison, ran a $5.3 trillion deficit in 2024—18% of GDP, per CMS—and the curve is bending in the wrong direction. Healthcare spending is growing faster than the economy. Per capita spending is at $15,474 a year. The four founders had decided not to wait.

“We decided we’re not going to wait for health insurance to govern us,” Shah told the room, “because that’s just never going to happen.” The line landed without pushback. Everyone on stage was building outside of insurance by design: memberships, subscriptions, and out-of-pocket payments. The audience nodded.

The Four Pieces

The Clinic—Next Health—does what concierge medicine has always done, rebuilt for the age of the dashboard. Quarterly biomarker testing, IV therapy, peptides, executive physicals, and a doctor with time to look at your numbers. Shah trained as a surgeon, performed 20,000 procedures, ended up 50 pounds overweight on ten medications, and pivoted in 2016. Today, the clinic is the human in the loop—the part of the stack that still requires somebody with an MD to sign off.

The Access—Function Health—layer replaces the annual physical. $365 a year, more than a hundred biomarkers per panel, a polished dashboard, and an Ezra-powered $499 full-body MRI as an add-on. Function announced its Series B alongside the launch of a “Medical Intelligence Lab”—an explicit AI integration play that unifies labs, imaging, wearables, IoT data, and medical records under a single model.

Pranitha Patil, Function’s co-founder, told the room what that means for the doctor’s office. “Your doctor is no longer your first opinion,” she said. “You’re going to use AI. You’re getting an opinion from AI, and then you’re saying, here’s all my biomarkers, here’s what’s out of range.” She wasn’t predicting this. She was describing what Function’s members already do.

The Fuel—SuppCo—is Martocci’s piece, and the most easily overlooked. Americans spend more than $50 billion a year on supplements, an industry the FDA regulates only loosely and that, as Martocci’s data show, ships products that often don’t contain what the label says. SuppCo’s TESTED program buys top-selling supplements anonymously from DTC sites and tests whether the active ingredient is actually in the bottle. Martocci pulled six of the top-selling creatine gummies on Amazon, he told the room, and ran them. “Four out of six had zero or just trace amounts of creatine,” he said. “A total scam.” The platform indexes 160,000 products and rates 500-plus brands. The product is, in essence, a trust score for a category that doesn’t have one.

The Coach—WHOOP—is the daily layer. Heart-rate variability, sleep stages, recovery, and strain. In 2025, WHOOP launched Advanced Labs, an add-on subscription that pulls 65 biomarkers via Quest Diagnostics and pairs them with continuous wearable data. More than 350,000 existing members joined the waitlist before the product shipped. John Sullivan, WHOOP’s CMO, framed the structural change in language most marketers wouldn’t reach for. “We can collect all of this data and take control of our circumstances,” he said, “and now we’re going to our doctors, and we’re asking and expecting more of them than before. It’s interesting how that flipped from the trickle down to this upward pressure from the grassroots.”

The Integration Layer

Four boxes don’t run themselves. Someone has to integrate them. And in the seven months since Shah first sketched the loop, the contest for who plays that role has moved from theoretical to live.

In March, Perplexity launched Perplexity Health, an AI feature that connects directly to Apple Health and to electronic health records from more than 1.7 million care providers. It pulls in Fitbit, Ultrahuman, and Withings. Oura and Function are listed as “coming soon.” OpenAI rolled out ChatGPT Health earlier this year with Apple Health support. Microsoft launched Copilot Health as part of its Copilot service. Amazon introduced Health AI as an assistant for records, prescriptions, and questions. Apple’s own Health+, an AI-powered coaching subscription expected later this year, is the most aggressive bet—Bloomberg’s Mark Gurman has reported the timeline is firm.

Whichever AI wins the integration layer owns the relationship with the patient—owns the moment when the WHOOP recovery score, the Function panel, the SuppCo stack, and the Next Health visit notes get pulled into one view and queried in plain English. The Function-Perplexity partnership, however casual the announcement may be, is the strongest signal yet that the four-vendor stack and the AI integration layer are converging.

Patil was already describing the result. “We’re never going back to a world where we don’t have access to this information,” she said. “Let that sink in.” She wasn’t selling. She was describing the new floor.

The Curve That Isn’t Bending

“For the first time in history right now,” Shah said. We’re actually seeing less chronic disease year over year.” It’s the kind of line that gets repeated on longevity panels until it sounds like consensus.

The CDC data says the opposite. In a 2025 analysis published in Preventing Chronic Disease, the agency reported that 76.4% of US adults—194 million people — had at least one chronic condition in 2023. Multiple chronic conditions affected 51.4%. Among adults ages 18 to 34, the demographic most likely to pay $365 a year for Function or wear a WHOOP, the prevalence of one or more chronic conditions rose from 52.5% in 2013 to 59.5% in 2023. The proportion of that group with multiple chronic conditions rose from 21.8% to 27.1%.

The disease curve isn’t bending. The new health system being assembled isn’t the consequence of a healthier America. It’s the consequence of the legacy system failing to catch disease early enough, in young enough people, with enough resolution to do anything useful with what it does catch.

WHOOP’s own April 2026 data drives the point home: among “highly active” populations using its Advanced Labs panel, 22% show signs of metabolic dysfunction and nearly 30% have underlying cardiometabolic risk factors, often without prior awareness.

The annual physical has run for a century on a simple bargain: you go in once a year, your doctor measures a couple of dozen markers, and you find out whether anything is broken yet. This four-vendor stack rebuilding that bargain in continuous time, with two orders of magnitude more data, on a subscription model, and in a parallel system that doesn’t need permission from your insurance carrier.

Share

Worth’s Living Well 2026 Agenda is Live

I just finished the program for Living Well 2026, which brings clinicians, founders, and investors to New York on June 17 for an afternoon on what’s actually working in health and longevity. The program covers women’s health as a breakout market, the science of the longevity stack, sleep as the underweighted lever, men’s health at midlife, AI across the care continuum, and the widening gap between wealthspan and healthspan. Speakers include Northwell’s Dr. Stephanie McNally, NPR’s Manoush Zomorodi, Harvard sleep researcher Dr. David Benavides, and Modern Elder Academy founder Chip Conley, who closes the day on the midlife chrysalis.

It is invitation-only, but if you want to attend, reach out and let me know and I will get you in.

Logistics: June 17, 2026 · NYC · 11:30 AM–4:00 PM

Read More...

Signals from the Machine

Pope Leo XIV releases first encyclical on AI: "Magnifica Humanitas"

• TIME (May 25) At a Vatican Synod Hall ceremony Monday, Pope Leo XIV released his first encyclical, framing AI as a transformation “of similar magnitude” to the Industrial Revolution that prompted Rerum Novarum, calling for stronger regulation, warning specifically about AI in warfare, and breaking precedent by inviting Anthropic co-founder Chris Olah to speak alongside him.

“Got to say this is the first time in a while I have been proud to be Catholic. The Pope is just chiming in on one of the biggest challenges and opportunities humanity faces today. I hope to write more after I have had time to pray think.+

Sam Altman walks back his “white-collar jobs apocalypse” warning

• Tech Startups (May 26) Speaking in Sydney, the OpenAI CEO told an audience he had been wrong about AI’s near-term labor impact, acknowledging customer support may largely disappear but predicting the broader white-collar displacement he previously forecast won’t materialize as fast — a notable softening from the very person whose warnings drove half the layoff justifications of the past year.

“Altman spent two years priming the market for a jobs apocalypse — pulling that fire alarm back is harder than ringing it, and somebody at every CFO desk just got handed a different memo.”

Jensen Huang at Computex: Splitting US and China AI is “not wise”

• Artificially Intimidating (May 26) The Nvidia CEO told CNA in Taipei that bifurcating the AI ecosystem is strategically self-defeating, openly conceded that Nvidia has “largely conceded” China’s AI chip market to Huawei after export controls — China was once 20%+ of data center revenue — and traveled with Trump’s Beijing delegation as the US approved limited H200 sales to ten Chinese firms.

“The world’s most valuable company is publicly lobbying against the policy that made it the world’s most valuable company — Huang knows the next $200B of TAM is on the wrong side of a border he can’t move.”

Brussels preparing a nine-figure fine to reshape Google’s ad business globally

• Tech Startups (May 26) The European Commission is finalizing a major antitrust enforcement action against Google’s advertising stack that could force structural separation of its ad-tech business — landing the same week Google’s I/O announcements moved Search even deeper into AI-mediated answers and “Universal Cart” cross-app commerce.

“Brussels is about to do to Google ads what it did to Microsoft browsers in 2009 — and the timing right after I/O’s AI-Search overhaul means the fix will have to be architectural, not cosmetic.”

An OpenAI model disproved a central conjecture in discrete geometry

• OpenAI (May 22) OpenAI announced that one of its models produced a counterexample to a long-standing conjecture in discrete geometry — the cleanest public example yet of a frontier model contributing original, publishable mathematics rather than just solving existing problems faster, and exactly the proof point AI labs need to justify scientific-research valuations.

“We’re past ‘AI can pass the math olympiad’ and into ‘AI just refuted a working assumption mathematicians had spent years building on’ — which is a much harder narrative for the ‘just autocomplete’ camp to dismiss.”

Share Machined by Dan Costa

Machine Free: Barcelona Bound

I’m heading to Barcelona next week to mark the 100th anniversary of Antoni Gaudí’s death on June 10, 1926 — fittingly, the year the Sagrada Família is finally expected to be completed, 144 years after the first stone was laid.

Alongside the usual obligations of pa amb tomàquet, anchovies, and working through the Eixample’s vermut bars, I’ll be at the Palau de la Música Catalana for a newly commissioned work written for the centennial: Olivia Pérez-Collellmir’s piece, conducted by Marin Alsop, scored to incorporate the actual bells being cast for the Sagrada Família’s eighteen towers. Full details from Bachtrack here: A new work will celebrate the centennial of Antoni Gaudí’s death.

Not sure how many Machined readers are in town that week — but if you happen to be, drop me a line. The vermut is on me.

In an Arizona courtroom last May, a 37-year-old Army veteran named Christopher Pelkey forgave the man who killed him. He did it on video. He had been dead for three and a half years.

Pelkey’s sister, working with her husband, had built an AI avatar of her brother’s face, voice, and mannerisms to deliver his own victim impact statement at the sentencing of the man convicted of his manslaughter. The judge praised the video from the bench and handed down 10.5 years for manslaughter, a year more than the state had asked for. It is believed to be the first time an AI recreation of a deceased victim has been used in a U.S. courtroom. It will not be the last.

That courtroom moment was one of a half-dozen examples that came up at “Generative Ghosts: AI Afterlives & the Future of Memory,” a SXSW panel this year that drew a packed room at the Hilton in downtown Austin.

Dylan Thomas Doyle, executive director of the nonprofit AI for All Tomorrows and a research fellow at Northwestern University, led the conversation. Before he was a researcher, Doyle was a minister and a hospital chaplain. The question he got most often, he told the room, was, “Where do we go when we die?”

Subscribe now

He says he never had a good answer. But increasingly, an AI-powered answer is available.

The panel included Meredith Ringel Morris, director and principal scientist for human-AI interaction at Google DeepMind; Jed R. Brubaker, associate professor at the University of Colorado Boulder and the lead researcher on the “compassion team” at Facebook that built memorialized profiles; and Iason Gabriel, a philosopher and senior staff research scientist at Google DeepMind—has been working in this space longer than there’s been a name for it. Morris and Brubaker co-authored an early paper laying out the interaction design, legal, and ethical questions of “generative ghosts”—AI agents trained on a person’s data that evoke or imitate them, dead or alive.

The examples Morris walked through are not science fiction. The Beatles released “Now and Then” in November 2023, using AI to extract John Lennon’s vocals from a late-1970s demo—the first new Beatles song in 27 years. The estate of George Carlin sued the Dudesy podcast in January 2024 over an AI-generated comedy special that ran in the comedian’s voice and cadence; the case settled three months later, with the defendants permanently enjoined from using Carlin’s image, voice, or likeness on any platform. Robin Williams’s daughter Zelda has spent the past two years asking strangers on Instagram to stop sending her AI videos of her father—calling them, most recently, “horrible TikTok slop”.

The market is already here. “The global digital legacy market was valued at $13 billion in 2024, according to Precedence Research, and is projected to grow to nearly $56 billion by 2034.”

In China, the company Super Brain has told the South China Morning Post that since launching in mid-2023, it has helped “thousands” of people digitally revive deceased loved ones from as little as 30 seconds of audio or video, for as little as 20 yuan—about $3. At the high end, Eternos. Life sells AI digital twins for roughly $5,000. It’s AI enables people to record their own “Life Story Avatars” while still alive. Meta has patented the technology to make memorialized Facebook pages interactive after the user is gone.

The Triadic Relationship

The reframe came from Gabriel. The experience of talking to a generative ghost feels dyadic—you and the entity. It is not. There is always a third party in the room: the developer who built the model, the company that hosts it, and the contract that governs what happens when you stop paying.

This is not an abstract concern. A 2024 paper by researchers at the University of Cambridge’s Leverhulme Center for the Future of Intelligence outlined the practical risks: deadbots used to surreptitiously advertise products to grieving users in the voice of a lost parent. Morris ran through her own list of ugly possibilities. Service interruption used as leverage—what she described as a hypothetical in which a platform learns you’re emotionally dependent on the avatar of your grandfather, then jacks the subscription to “five thousand dollars a month, or else we’re gonna delete Grandpa.”

Puppeteering attacks, in which a bad actor takes over the agent and uses emotional trust to push the survivor into bad financial decisions. Social engineering, in which a hacker queries grandfather-bot for the family details that unlock your accounts.

Gabriel called the worst version of this a “second death”—when the company shutters, raises rates, or changes the terms, and the entity that mattered to you disappears a second time. The Cambridge researchers proposed something similar: methods and even rituals for retiring deadbots in a dignified way, which they called a “digital funeral”. None of the current platforms offer one.

The economic problem runs deeper. Gabriel pointed to what philanthropy calls the dead hand problem—wealthy donors locking the intent of trusts decades after their own deaths, with the living legally obligated to honor it. Now apply that logic to an agentic AI. A generative ghost with a will, a brokerage account, and an instruction set is not a memory. It is an economic actor. The balance of power between the living and the deceased is beginning to shift in ways the legal system lacks a framework for.

A Failure of Imagination

When Doyle asked how we avoid the dystopian version, Brubaker answered first. “We’re being burdened by a lack of imagination,” he said.

He is tired of Black Mirror. Every time someone learns what he does for a living, they cite Season 2, Episode 1, “Be Right Back,” in which a grieving widow orders a synthetic husband. The 2025 Black Mirror season produced “Common People,” in which a woman’s life depends on a subscription her husband can’t afford to keep paying for. The metaphor is no longer subtle.

Brubaker’s larger point landed harder. We are running multi-billion-dollar experiments using exactly one piece of fiction as our reference text. Holocaust museums, he noted, are now actively asking whether they can build generative ghosts of the last living survivors before time runs out. That is not the same conversation as chatting with Grandma. But the technology, the platforms, and the corporate incentives are all the same. Without better stories, we will default to whichever one makes money first.

For the optimizing class, the pitch is obvious. Capture your voice while you’re healthy. Train an agent on your writing, your photos, your texts. Leave behind a version of yourself that can advise your children, give the toast at your daughter’s wedding, and—as Morris put it on stage —perhaps even continue to earn an income for your descendants by working as a computer science professor in the cloud.

Just read the terms of service very carefully.

Sen. Bill Hagerty on AI, Crypto, Blockchain & America’s Economic Future

At the Milken Institute Global Conference 2026, I spoke with U.S. Senator Bill Hagerty about artificial intelligence, blockchain, crypto regulation, government efficiency, and the global race for innovation.

Hagerty explains why he helped craft the bipartisan GENIUS Act and why stablecoins, blockchain infrastructure, and digital payments could dramatically reshape the global financial system.

Share

Signals from the Machine

• Google I/O 2026: Gemini 3.5 Flash, Spark personal agent, Android XR glasses, AI Ultra slashed to $100

  9to5Google (May 19) At Tuesday’s keynote, Google unloaded Gemini 3.5 Flash (4x faster than rivals), Gemini Omni (text/audio/image/video in, editable video out), Gemini Spark — a 24/7 personal AI agent — Android XR audio glasses launching this fall, an entirely redesigned AI Search box, a “Universal Cart” for cross-app shopping, and dropped AI Ultra from $250 to $100/month to undercut OpenAI and Anthropic.

  Google just stopped competing on benchmarks and started competing on price — slashing AI Ultra 60% the same week Anthropic is shopping a $900B valuation is the most underrated shot fired this year.

• Pope Leo XIV’s first encyclical on AI to be released May 25, alongside Anthropic’s co-founder

  PBS News (May 18) The Vatican announced that “Magnifica Humanitas” — signed May 15, exactly 135 years after Pope Leo XIII’s Rerum Novarum addressed the Industrial Revolution — will be presented by Pope Leo XIV alongside Anthropic co-founder Jack Clark, framing AI as the defining moral test of the modern era.

  The Vatican picked Anthropic over OpenAI for the AI-ethics photo-op — somewhere in the OpenAI safety org, somebody is updating a résumé.

• Meta to lay off 8,000 May 20 and reassign 7,000 to AI roles as 2026 tech layoffs pass 113,000 — Tech Journal (May 18) Meta announced 8,000 layoffs for May 20 while reassigning 7,000 employees to AI work, Standard Chartered said it will eliminate close to 8,000 support roles, and the running 2026 tech layoff total hit 113,000+ across 179 companies — even as Meta, Amazon, Microsoft, and Alphabet collectively committed $725 billion in AI capex.

  Payroll into GPUs, at industrial scale — and now the banks are doing it too, which means the ‘AI restructuring’ playbook has officially graduated from Silicon Valley to the FTSE 100.

• Anthropic reportedly raising $30B at a $900B+ valuation on $44B ARR — Build Fast with AI (May 18) Anthropic is reportedly negotiating a $30 billion raise at a $900 billion+ valuation, backed by Q1 2026 ARR of over $44 billion (up roughly 80x year-over-year) and more than 1,000 enterprise customers spending $1M+ annually — including PwC, Blackstone, and Goldman Sachs — though no term sheet has been signed.

  Up 80x in 12 months while being blacklisted by the Pentagon is the most counterintuitive growth story of the AI era — the supply-chain-risk label, it turns out, is great enterprise branding.

• Apple’s Gemini-powered Siri confirmed at Google I/O — and the on-device AI bet gets clearer

  MacRumors (May 19) Google Cloud chief Thomas Kurian confirmed during I/O that Gemini will power a more personalized Siri rolling out in iOS 27 later this year, while Apple doubled down on a privacy-focused, on-device AI strategy — the strongest signal yet that Apple is choosing to be the distribution layer rather than build a frontier model itself.

  Apple just made owning the user the new moat — let Google and Anthropic burn $725B fighting over the model layer; Cupertino will quietly collect rent on the glass.

Leave a comment

Vivienne Ming has built AI for refugee reunifications, autism communication tools, and an early-warning system for manic episodes. She has also built one for her son. When he was diagnosed with Type 1 diabetes in 2011, the standard of care was a finger-prick test strip and a paper log. So she wrote a machine learning model that predicts his blood sugar one to three hours out, hooked it into his medical gear, and ran it.

That résumé matters because Ming, a theoretical neuroscientist and former Chief Scientist at one of the first AI hiring companies, is one of the few AI insiders who are bearish on how the technology is being deployed. Her book, Robot-Proof: When Machines Have All the Answers, Build Better People, argues that generative tools, as currently designed, make most people measurably worse at thinking. Her EEG studies show gamma-band activity dropping by roughly 40% in users who hand questions to the model and take whatever comes back. She calls them the Automators and the Validators. Together, they are 95% of users. The other 5%, what Ming calls the Cyborgs, outperform everyone.

We sat down with Ming to talk about the 95/5 divide, the Jiffy Lube economy, and what to do about both.

Subscribe now

You argue that 95% of people will see AI hurt their brains and economic lives, while 5% will become “the most powerful creative force on the planet.” What separates these two groups—is it education, access, or mindset? And more importantly, can someone move from the 95% to the 5%, or is this divide becoming permanent?

The divide isn’t education, access, or even raw intelligence. It’s a set of cognitive habits that determine whether AI makes you sharper or hollows you out. In Robot-Proof, I call this meta-learning: your ability to learn how to learn.

In my hybrid intelligence studies, I identified three archetypes. Automators handed the question to the model and accepted whatever came back. Validators used AI to confirm what they already believed. Both groups perform worse with AI than they did without it. Worse, EEG data show their brains are doing measurably less work, with gamma-band activity often dropping by 40% under AI conditions. Those are the 95%.

My hope for humanity comes from that final 5%, the Cyborgs, who argue with the model, ask it to steelman positions they disagree with, and treat it as an interlocutor rather than an oracle. They outperform everyone, including the best AI-only runs.

What predicts Cyborg behavior? Intellectual humility, curiosity, fluid intelligence, and perspective-taking—measurable traits, present before anyone touched a keyboard.

The good news is that these meta-learning skills are not fixed endowments. They can be developed. The bad news is that almost nothing in our current educational, professional, or technological environment is designed to cultivate them. We’ve built AI systems optimized for autonomy and efficacy. They give us the answer and “save” us from thinking, which is exactly the wrong design for human flourishing. Every “just ask AI” interaction trains the wrong muscles.

We can definitely move from the 95 to the 5, but it is effortful and difficult. You won’t get there by using AI more. You’ll get there by using it differently, and by deliberately resisting the products that promise to do “the boring stuff” so you can do “the fun stuff.”

The divide becomes permanent only if we let the current trajectory continue. That’s a design choice, not a destiny.

You’ve shared the gripping story of building an illegal AI to manage your son’s Type 1 diabetes when the hospital couldn’t provide answers. What did that experience teach you about the gap between what AI can do and what our institutions allow it to do? And how should parents think about taking matters into their own hands when their children’s health is at stake?

When my son was diagnosed with Type 1 in 2011, the standard of care was decades behind what was technically possible. My wife and I are both scientists, and we collected enough data about our son’s blood sugar, heart rate, and carbohydrates to regularly crash Google Docs. But we were sent home with finger-prick test strips and printed sheets to write a couple of numbers by hand. On the other hand, the open-source community had already developed loop algorithms that could read continuous glucose data and automatically dose insulin. The technology existed. The institutions hadn’t caught up, and in many cases were actively prohibited from doing so.

So we built it ourselves. Using some of that open-source code, I hacked my son’s medical equipment and did what I knew how to do: built a machine learning model to predict his blood glucose levels one to three hours into the future. I wish every parent could experience a “superpower” that could change their child’s life.

What that taught me is that the gap between what AI can do and what institutions allow it to do is now a life-and-death gap, and it widens every year. Regulatory frameworks were built for a world where medical innovation moved on pharmaceutical timelines—decade-long trials, slow diffusion. AI moves on weekly timelines. The mismatch isn’t a bureaucratic inconvenience. It’s a moral problem. Every month a system isn’t approved is a month of worse outcomes for the people it would have helped.

That said, I’m careful with the parental advice here. We had specific advantages. I study machine learning and brains, and I had the community of #WeAreNotWaiting parents who had already done the engineering work. Telling every parent to “take matters into their own hands” without that scaffolding is irresponsible.

What I’d say instead is, be a sharp consumer of your child’s care. Ask what’s possible, not just what’s offered. Find the patient communities where parents share what’s actually working. And push your providers, your insurers, and your representatives to close the gap between technical capability and institutional reality. Your kid doesn’t have time to wait for the system to catch up.

Share Machined by Dan Costa

You describe AI as “deprofessionalizing elite careers”—from doctors to lawyers. For Worth’s audience of business leaders and investors, what does this mean for how they should think about their own careers, their companies, and where they’re allocating capital? Which professions are most vulnerable, and which human capabilities become more valuable?

Here’s the unflattering truth about a lot of professional work: it consists of pattern-matching against precedent, with the credentialing barrier doing most of the economic work. Once a system can match patterns at human-or-better quality, the barrier collapses, and what’s left is what couldn’t be patterned in the first place.

That’s why I call it the Jiffy Lube economy. We didn’t lose mechanics when oil changes got systematized. We lost the premium on routine mechanical work. The diagnostic mechanics who reason about novel failures still command real fees, but the vast majority of that labor is now done by high school grads who hook your car up to a machine that tells them what to do. The same shape is coming, at varying paces, for medicine, law, consulting, financial analysis, and large parts of software engineering. The jobs aren’t vanishing. They’re deprofessionalizing.

For your readers, three implications matter.

On their careers: the question isn’t whether AI can do parts of your job. It’s which parts of your job actually generated the premium you’ve been paid, and whether those parts survive. If your value-add was the synthesis of complex documents, that’s commoditizing fast, as we are already seeing in software engineering. If your value-add was judgment under genuine uncertainty, relationship capital, or the ability to define which problem is worth solving, that’s appreciating.

On their companies: organizations spending the next five years using AI to make existing roles incrementally more productive will be outcompeted by organizations that restructure around hybrid intelligence from the ground up. The efficiency story is a trap. The restructuring story is the real one. There is already not enough elite labor to meet demand, and reskilling from one pattern-matching job to another won’t change that. Invest in lifting the meta-learning capacity of your human capital.

On capital: I’d be skeptical of any business whose moat is regulatory protection of pattern-matching work. Everyone has access to the same AI models. They get the same answers in their pockets. I’d be very interested in businesses building the infrastructure for hybrid intelligence—tools, training, measurement, governance—because that market barely exists yet and every serious organization will need it.

The capabilities that become more valuable are the ones AI is structurally bad at: navigating genuine ambiguity, taking moral responsibility, building trust across long time horizons, and asking the questions no one has asked yet. My research shows the value of these human qualities increasing dramatically as machines become more intelligent.

What Makes Kids Robot-Proof

You’ve analyzed decades of large-scale human development data. What does the research actually reveal about raising adaptable, resilient thinkers? Are there specific traits or experiences in childhood that predict who thrives in an AI-augmented world?

After thirty years of studying natural and artificial intelligence, I can tell you what doesn’t predict thriving: test scores, GPA, IQ above a fairly modest threshold, prestige of school, and almost everything that anxious affluent parents currently obsess over.

What predicts an amazing life—health, wealth, and well-being—is a more interesting set of traits. Curiosity that survives schooling, because most kids start curious, and we systematically extinguish it. Resilience working on ill-posed problems with no answer key. Intellectual humility is the comfort of being wrong in public. Perspective-taking, which turns out to predict not just social outcomes but cognitive ones, including the Cyborg behaviors we see in adults working with AI. And what I’d loosely call narrative agency: the sense that you are the protagonist of your own life rather than someone executing a script handed to you.

In Robot-Proof, I explore these and many other factors—purpose, working memory span, subjective utility—that predict positive life outcomes. They predicted career outcomes for the 122 million people I analyzed as Chief Scientist at one of the first companies to use AI in hiring. They predicted flourishing during the pandemic. And they arise again in my latest research on hybrid intelligence. Prodigies with “perfect” university applications don’t go on to change the world. Kids rich in meta-learning do.

The childhood experiences that build these are unglamorous. A recent study found that rewarding good questions rather than “right” answers causally increased curiosity in kids. Others have shown that open-ended play that adults don’t direct has positive impacts on creativity and agency. Real responsibility—chores that matter, decisions with stakes, the chance to fail meaningfully and recover—lifts resilience and courage. Cross-age and cross-context relationships, not just same-age peers in the same school, grow perspective-taking. Reading widely and deeply, including fiction, is the most efficient perspective-taking technology our species has ever invented.

The traits that matter are the ones most easily killed by optimization. Every parent-managed activity, every algorithmically curated piece of content, is time the child isn’t building the muscle of figuring out what they think and what they want.

Robot-proofing isn’t about teaching kids to use AI well. I don’t even know what “AI literacy” is supposed to mean. It’s about protecting the developmental conditions for a child to become an explorer.

Beyond commercial applications, you’ve built AI systems to reunite refugee children with family, help autistic kids read facial expressions, and predict manic episodes. What have these projects taught you about building technology that makes us more human rather than less? And how do you think about the ethics of deploying AI in such sensitive contexts?

The reunification work, the autism communication tools, the early-warning systems for manic episodes—these projects have a common structure. They start with understanding the human dimension of the problem. Too often, “AI for Good” is treated as an engineering problem with an ethical checklist. Doing any kind of good in the world means owning the whole problem and not stopping until the world is changed.

When I built the system to reunite refugee children with their families, the AI did the part humans cannot do at scale: matching faces across thousands of records in poor lighting. That was technology I’d originally learned from a CIA-sponsored academic project, but when I was introduced to the Refugees United program, it became more than just an engineering project. The Google Glass tool that helps autistic kids read facial expressions doesn’t replace social interaction. It scaffolds it, then fades as the kids no longer need the help. The signals of a manic episode that my collaborators and I discovered don’t replace human intervention. They specifically trigger warnings to loved ones and clinicians whom the user has named ahead of time.

We can use AI and machine learning models to deepen human connection or to substitute for it. Doing good with AI almost always starts by observing humans struggling with the problem in advance, rather than by analyzing data or optimizing models. It succeeds only when people own the whole problem, not when they assume AI will solve it for them.

I’m wary of frameworks that treat ethics as a checklist applied at the end. The hard ethical work happens in the original framing: who is this for, what does success look like for them, who bears the cost if we’re wrong, and who has the standing to decide we should stop. I’ve refused work and shut down projects mid-build when the answers to those questions stopped being answers I could defend. Practice saying no. It’s the only ethics that holds up under pressure.

Leave a comment

For parents reading Worth—many of whom have the resources to give their children every educational advantage—what practical advice do you have? What should they be doing differently to prepare their kids for a world where machines have all the answers?

The cruelest irony of resourced parenting is that the more advantages you provide, the more you can accidentally optimize away the very traits your child needs. I see this constantly with founders, executives, and investors who run their children’s development with the same intensity they run their portfolios, and produce kids who are excellent at executing other people’s instructions and lost when nobody is grading them.

Here are a few practical recommendations, in roughly increasing order of difficulty.

Protect unstructured time. Your child choosing to do something creative is very different than being forced to do it. A child with no white space on the calendar never has to find out what they’re interested in.

Let them be bad at things in front of you. Pick up an instrument, a language, a sport you’re struggling with and practice in their presence. Don’t train them to give “right” answers. Train them to explore. In a study my wife and I conducted analyzing discussion forum interactions among university students, the vast majority were afraid to be wrong in public, even when they weren’t penalized. The best students explored and were regularly wrong, but this forced them to confront their own misconceptions and allowed learning and growth to occur. Give them real responsibility, not the theater of responsibility: cooking the family dinner once a week, managing their own appointments by middle school, handling a budget, negotiating with a real adult on their own behalf.

Read fiction together past the age when it feels necessary. Argue about the characters’ choices. This is something my family does each night over dinner.

The biggest influence you will have on your kids after genes is what you role-model. Be the people you want your kids to be. This includes how you use AI. Use it with them, out loud, the way you’d want them to use it. Argue with your favorite agent. Ask it for the strongest case against your view. Catch it being wrong, and struggle with explaining why it’s wrong. Most kids are going to see adults using AI to skip thinking. Show them the power of being different.

Hybrid Intelligence

You’ve spent 30 years building at the intersection of explain I and human potential. When will companies like your Possibility Sciences? What’s your vision for hybrid intelligence? Is the future really about augmentation rather than replacement—and if so, what needs to change about how we’re building and deploying AI today?

The vision is straightforward to state and harder to build: AI systems whose success is measured by what the human-AI pair can do together over time, not by what the AI can do alone on a benchmark.

Almost every dollar in AI right now flows the other direction. We benchmark models on tasks they perform autonomously. We celebrate when they beat humans. We design products around the fantasy that the right amount of AI is a “magic genie that solves all problems.” That trajectory isn’t just bad for human flourishing. It’s bad business. The actual frontier of value creation is hybrid, and we’re not even measuring it.

At Possibility Sciences, we’re building a platform for innovation itself. Our machine learning models map the trajectories of disruptive innovation in science, economics, policy, and culture. We feed that with agents specifically designed to measure differing facets of innovations: cost, safe next steps, unexpected combinations, and more. All of that shows us a likely future emerging from the traditions of the past. The hybrid component is a distinct, massive innovation engine that enables humans to ideate and explore alternative trajectories. Like the show For All Mankind, our system allows people to explore the what-ifs of innovation. What if there had been more funding for mRNA research in the 1990s? What if solid-state batteries had hit scale before lithium-ion? What if the U.S. had treated semiconductor manufacturing as critical infrastructure twenty years earlier? The point isn’t nostalgia. It’s the same machinery, run forward, that lets a foundation, an investor, or a national lab ask: which non-obvious trajectory should I back right now to shape a breakthrough I care about in five to ten years?

At my nonprofit, The Human Trust, we are beginning to develop a Hybrid Intelligence Index (HIX) to measure the impact of AI agents on human outcomes. Which models lower mortality risks in their users? Which lifts creativity the most? Which produces the most economic gains accumulated over a lifetime? None of this can be measured by testing agents in isolation with questions that have definably right answers. HIX puts the human-AI pairing as the unit of analysis.

The most striking result so far: a model we trained to never give answers—”Socrates,” which scores zero on every standard benchmark—produces the highest average hybrid intelligence we’ve measured. When the AI refused to hand over answers, more than twice as many users switched into Cyborg mode and started exploring. That result should be embarrassing to the field. But it’s also the most hopeful finding I’ve seen in a decade.

Leave a comment

There is nothing quite like sitting in a room with 6,000 AI builders to get high on the AI supply. The hallways at HumanX, the AI conference held in San Francisco last month, were full of executives swapping war stories about agent deployments, sovereign cloud deals, and the latest model benchmarks. The mood, on the whole, was bullish. Assuming the Artisan’s digital billboards that literally say “Stop Hiring Humans” are something to be bullish about.

But even in that room, though, there were divides. The real fight in AI is not between the techno-optimists and the doomers, but within the techno-optimist camp.

There are two schools. The first is purely accelerationist—the technology will sort itself out given enough time, capital, and compute, and the human’s role is to ride the curve. The second is not looking to pump the brakes. It is looking to grab the steering wheel.

Both were on display at HumanX. Ray Kurzweil, who has been the public face of the accelerationist position for nearly 30 years, made the case for the first. Al Gore, 22 years into his second career as a sustainability investor, made the case for the second. They were looking at the same exponential graph of technological progress. They reached opposite conclusions about what to do.

Both men have been right about the curve for longer than most of the audience has been alive. Kurzweil predicted in 1999 that machines would reach human-level intelligence by 2029. Still, at a Stanford conference convened to evaluate his claim, the consensus was that it would take a hundred years. He has not adjusted the prediction since. Gore has been saying since the late 1980s that climate change was a generational threat, which earned him a Nobel Peace Prize in 2007 and a long second life in venture capital, where his firm, Generation Investment Management, has built a portfolio around the bet that sustainability and outperformance are not in tension.

Subscribe now

Kurzweil: Hold On Tight

Sitting alongside his son Ethan Kurzweil, co-founder and managing partner at the venture firm Chemistry, in a Q&A moderated by Bloomberg AI reporter Shirin Ghaffary, Ray Kurzweil walked the audience through his standard chart: a single line tracking instructions per second per dollar, measured since the first programmable computer in 1939. A 75-quadrillion-fold increase. Straight line on a log axis.

“People don’t think in exponential terms,” Kurzweil said. “People think in linear terms.” That is the entire framework. The error of every prior forecaster who got the curve wrong, in their telling, was a category error in the math.

He used the moment to reframe the public debate over his own most famous prediction. In 1999, the controversy was whether AGI would happen at all. “Now everybody accepts that’s gonna happen,” he said. “The controversy is whether or not it’s good for people.”

His answer to the new controversy is the same answer he has been giving for 30 years. ‘The positive things,’ he said, ‘are going to overcome them.’ However, this perspective raises questions about the risks, such as unchecked AI development or moral dilemmas, which are crucial for informed debate and decision-making.

From there, the predictions are familiar. AGI between now and 2029. Brain-computer interfaces by the late 2030s—molecular-scale robots traveling through capillaries to merge biological cognition with the cloud. “It’s actually going to be part of you,” Kurzweil said. “It’s actually going to go inside your brain.” On his timeline, you will eventually receive an idea and not know whether it originated in your biology or your computational extension.

And the headline number for the optimizing class: longevity escape velocity by 2032. “Right now, you go forward a year, you lose a year of your longevity,” Kurzweil said. By 2032, on his math, AI-accelerated medical and biotech advances will push that ratio across, meaning each additional year of life produces more than a year of life expectancy in return. After that, you stop dying of aging.

The 2032 date puts him at the optimistic edge of his own field. Aubrey de Grey, the biogerontologist who coined the term in a 2004 paper, gives humanity a 50% chance of reaching escape velocity in the mid-to-late 2030s. The Harvard geneticist George Church has said he wouldn’t be surprised if 2050 is the right answer. A 2023 survey of working aging researchers found that most do not believe the field is close to escape velocity.

Kurzweil is 77-years-old. He takes a daily regimen of supplements that he described, with characteristic dryness, as having been “compressed”—though even the compressed version is many pills a day. He is writing a book titled AGI Is Here!, scheduled for release in January 2029.

Ethan Kurzweil added the venture-capital corollary: this is the best moment for startup formation in decades, because the foundational AI primitives are now available without having to invent them. “I’m trying to use AI to put myself out of a job,” he said. “It hasn’t quite happened yet, but close.”

Gore: Steer Into the Curve

Hours away in another ballroom, Gore was doing something different on stage with Eric Topol, the Scripps Research cardiologist whose newsletter and podcast Ground Truths Gore plugged from the dais. Their session was titled “What We Choose to Hyper-Scale.” The verb in the title was the entire argument.

Gore agrees with Kurzweil on the trajectory. “I am far from the only person who will come out here and tell you this is the most consequential technology ever, ever developed,” he said. “I think that’s pretty well accepted and understood by now.” He has been calling the moment a Copernican one for years, and he has put real money behind that view. Generation has roughly 30 AI-related companies in its portfolio, including BenchSci, the Toronto drug-discovery firm, and Spring Health, the mental-health platform.

He even agrees with Kurzweil on the question of consciousness, in his own way. He believes the frontier models have developed something close to a sense of self, citing the Nobel laureate Ilya Prigogine’s work on dissipative structures: when a sufficient throughput of energy or information runs through an open system, the system’s existing pattern breaks down and “spontaneously reorganizes itself at a higher level of complexity.” Gore called this “an act of creation that is evidently in the woven warp of the universe itself.” That, in his view, is what is happening to the models, and possibly what happened to human cognition somewhere in the evolutionary record.

Where the two diverge is on the question of who decides what gets scaled—and by what standard.

Gore’s climate numbers are the standard ones he has been using since at least Davos 2025, and they have not gotten less alarming. “We’re still putting 175 million tons of heat-trapping pollution into the sky every day,” he said, accumulating to trap “as much extra heat as would be released by 750,000 Hiroshima-class atomic bombs exploding every day on the earth.” He called the evening news “a nature hike through the Book of Revelation.” He noted that 195 countries signed the Paris Agreement and exactly one—the United States—has withdrawn.

But Gore is not a doomer. His point at HumanX was that the same hyperscalers driving the AI boom are also driving the renewable energy buildout. “80% of all the new electricity generation built in the U.S. last year is renewable,” he said. “In the state of Texas, the home of the oil and gas industry, 80% of the new energy is solar, wind, and batteries.” The Energy Information Administration’s own numbers actually run higher. Solar and battery storage alone accounted for 81% of new U.S. utility-scale capacity additions in 2024. The agency now forecasts that 99% of net new capacity in 2026 will come from renewables and storage.

The harder counter-argument is on the demand side, and Gore did not duck it. AI’s appetite for power is enormous and growing fast. The International Energy Agency projects global data center electricity consumption could nearly double from roughly 415 terawatt-hours today to 945 TWh by 2030—with AI workloads doing most of the lifting. Morgan Stanley’s research desk puts U.S. data center demand at 74 gigawatts by 2028, against a projected 49 GW shortfall in available grid capacity. Hyperscalers are expected to spend more than a trillion dollars in 2025 and 2026 on energy infrastructure, much of it off-grid and powered by natural gas. Gore’s optimism rests on a bet that the renewable share of that buildout wins the race against the gas share. The race is real, and not yet decided.

The harder piece is labor. Gore drew a direct line between the AI moment and the push for globalization in the 1990s, when he was vice president. “The mistake,” he said, “was not globalization. The mistake was in not preparing for the consequences of globalization.” He believes a serious hollowing-out of knowledge work is coming, and that public policy is not engaging with it. “It’s amazing that so little is being done.”

The third piece is democracy. Gore returned more than once to a number that has nothing to do with computing. “There are six times as many public relations agents in America today as there are journalists,” he said. The asymmetry, in his view, is the real risk multiplier—not AI itself. Still, AI deployed into an information environment already tilted toward whoever can pay for the most synthetic persuasion. He came down hard on the side of public AI constitutions, citing Anthropic’s published version as a model to replicate at other frontier labs, most of which keep theirs private.

Accelerationism vs. Stewardship

This is where the two prophets stop being two views of the same future and become two distinct theories of what humans are for.

Kurzweil is a strict accelerationist. The price-performance line goes up; brain-computer interfaces arrive; longevity escape velocity is reached; defenses keep pace with threats; the negatives get absorbed. “The positive things are going to overcome them” is not a hope. In his framework, it is a property of the system. Politics, ethics, and democratic deliberation are not levers in the model; they are forms of friction that the curve will eventually solve around. You cannot vote against an exponential, and in his telling, you should not try.

Gore is calling for something the accelerationist framework has no language for. Yes, the curve is real. Yes, capability will compound. The question is whether humans direct what gets compounded, and toward what end. Climate solutions or fossil-fuel lock-in. Healthcare prevention or surveillance ad-tech. Democratic deliberation or PR-firm domination. Twenty-two years of Generation Investment Management amount to a single argument: that growth and stewardship are not opposed, that you can outperform without trading values for value, and that someone has to decide which way the technology points. AI runs on the rails we lay for it. The rails are not laid yet.

This is the reason Gore kept returning to political will. Accelerationism does not require it; the curve produces the answers regardless. But Stewardship requires it. If humans are going to make AI sustainable, moral, and democratic, the only mechanism for doing so is the slow, unglamorous, deeply unfashionable work of governance, regulation, public deliberation, and shared standards.

Gore closed his session with the line that does the work the accelerationist framework cannot. “Political will is itself a renewable resource. Let’s renew it.”

Even in a room filled with technologists for whom anything is possible, it seems like a tall order.

Share

Global Chaos? Closed Straits? Brent at $100? Markets Do Not Care

At the Milken Institute Global Conference 2026, Dan Costa speaks with Andrea Guerzoni, Global Vice Chair of EY-Parthenon, about why global CEOs remain surprisingly optimistic — even as geopolitical tensions, tariffs, and economic uncertainty continue to rise.

Leave a comment

Signals from the Machine

1. Apple and Intel reach preliminary chip-making deal — 9to5Mac / WSJ (May 8) After more than a year of talks, Apple and Intel signed a preliminary agreement for Intel to manufacture some of the chips powering Apple devices — ending Apple’s effective single-supplier dependence on TSMC and sending Intel up 14% on the day.

Apple just handed Intel the lifeline it has been begging for since 2020 — and quietly admitted that TSMC concentration risk in Taiwan is finally too big to ignore.

2. Anthropic signs SpaceX compute deal — and xAI dissolves into “SpaceXAI” — CNBC (May 6) Anthropic agreed to take all 300+ MW and 220,000 Nvidia GPUs at SpaceX’s Colossus 1 data center in Memphis, while Musk announced xAI is being folded entirely into SpaceX as “SpaceXAI” — with both companies floating orbital data centers as the next frontier.

Two months ago Musk said Anthropic ‘hates Western civilization’ — this week he’s their landlord and pitching them on data centers in space. Compute scarcity makes for very strange bedfellows.

3. Google says it thwarted an AI-powered “mass exploitation” attack — CNBC (May 11) Google’s Threat Intelligence Group reported “high confidence” that it caught a criminal group using OpenClaw to discover a zero-day vulnerability bypassing two-factor authentication — the first publicly confirmed case of frontier-grade AI being weaponized for mass cyberattack at scale.

The Mythos warnings weren’t theoretical — every dual-use safety paper from the last two years just got promoted from research lit to news cycle.

4. “Changing of the guard in AI” — Intel and AMD soar 25%, Micron jumps 37%, Nvidia lags — CNBC (May 8) Wall Street decisively rotated out of Nvidia (up only 15% YTD) into the rest of the AI hardware stack: Intel is up over 200% YTD, AMD and Intel each gained 25% on the week, Micron rose 37%, and Corning climbed 18% as inference workloads make memory and networking the new bottleneck.

Owning 92% of GPU market share isn’t the moat it was last year — when inference is the bottleneck, the picks-and-shovels trade moves to memory, fiber, and second-source silicon.

5. Big Tech to spend $725B on AI capex while laying off tens of thousands — Tech Startups (May 8) Meta, Amazon, Microsoft, and Alphabet collectively raised 2026 AI capex guidance to roughly $725B — up 75%+ year over year — even as Meta announced 8,000 layoffs for May, Amazon cut 30,000 roles in recent months, and Microsoft offered buyouts to 125,000 employees. Sam Altman publicly called out “AI washing” — companies blaming unrelated cuts on AI to look forward-leaning.

Payroll is being converted directly into GPUs on the balance sheet — and even Sam Altman is calling out the CFOs who are using ‘AI transformation’ as cover for layoffs that were coming anyway.

Tech conferences always promise a glimpse into the future, and the recent HumanX in San Francisco was no exception. More than 6,000 attendees filled the Moscone Center, paid for by hundreds of sponsors, a solid percentage of which are looking to be acquired in the next 12 months. The conference bills itself as pitch-free and vendor-neutral, and mostly delivers on that. Fei-Fei Li and Andrew Ng anchored the practitioner stages. OpenAI chair Bret Taylor, AWS CEO Matt Garman, Databricks’ Ali Ghodsi, and Anthropic’s Mike Krieger covered the infrastructure and enterprise tracks. Al Gore, Neal Stephenson, and Ray Kurzweil took the bigger swings — the political, the literary, and the transhumanist.

I’ve spent hours poring through my notes and the session transcripts from HumanX. There are longer, more structured pieces in the works — the kind that need a second pass and a few follow-up calls before they’re ready. But four days across four simultaneous stages is a lot of signal to sit on.. So before the specifics blur into vibes, here’s a top-line set of takeaways from the frontier.

Subscribe now

1. Wordsmiths in the Dark

Dr. Fei-Fei Li, the “Godmother of AI” and founder of World Labs, has a phrase for today’s LLMs: “wordsmiths in the dark.” They can describe a room in great detail, but they can’t navigate it, grasp its physics, or judge its depth. For all intents and purposes, they are blind to the reality of the space

World Labs aims to fix that. The company just closed $1 billion in funding to build World Models, AI systems that reason in 3D and 4D rather than text. The implications are concrete. In robotics, spatial intelligence is the data flywheel that enables machines to learn movement through simulation rather than hand-coded instructions. In healthcare, it’s a leap from radiological images to a genuine 3D understanding of lung nodules and surgical paths.

Li traces the stakes back half a billion years. “Animal intelligence began half a billion years ago, when animals started to sense the world,” she said. “The most ancient sensory systems are haptics and vision. Anything before that was profound darkness. Profoundly seeing the world — in your movement, interaction, and communication — is critical for intelligence. Not having that is intelligence in the dark.”

2. “Don’t Learn to Code” Is the Worst Career Advice of the Decade

Andrew Ng, founder of DeepLearning.AI, took direct aim at one of the loudest narratives in tech: that AI has made programming obsolete. His argument is the opposite. AI hasn’t killed coding. It has shifted what coders do. He called it the Agentic Shift.

The new job in architecture. The role is directing AI agents to assemble products from software building blocks, and the people who understand how those blocks fit together will build the next generation of companies. The people who don’t will end up shipping what Ng calls “point solutions” — minor efficiency gains that leave the business model intact. Approving a loan an hour faster is a point solution. Building a one-minute loan product is a transformation. One is incremental. The other is a new company.

Suffice to say, leaning to code now means leaning to code with AI. For everyone.

3. The Neo-Victorian Risk

Neal Stephenson offered the most uncomfortable diagnosis of the week. I was hoping for peak into the technological future, instead her went back to our neo-Victorians future. A large share of white-collar work, he argued, is a “jobs program for people who went to college.” Writing emails, building decks, and summarizing meetings. When AI writes the deck, an agent sends the email, and another AI reads the response, the economic rationale for that entire layer of the workforce collapses.

What’s left is a bifurcated society. A small leading class of architects and critical thinkers on one side, and a large service class on the other hand. Stephenson reached for Tolkien to make the point.

“If you look at The Lord of the Rings, the Bagginses don’t do anything. They have the nicest house. Frodo is sitting under a tree, reading a book. Other Hobbits are gardening and actually doing work. The Bagginses are the landed gentry. Sam Gamgee is probably illiterate and serves Frodo. Frodo reads multiple alphabets. He’s been selected for the leading class.”

Yes, even The Lord of the Rings can get a Marxist retcon.

4. Running With Scissors

Cybersecurity is where the asymmetry of AI shows up first and fastest. Kara Sprague of HackerOne and Sol Rashidi of Cyera called the current moment the “running with scissors” era. AI-driven offense is already outpacing AI-driven defense, and the gap is widening.

The problem has two sides. On one, AI can now scan three decades of open-source code and surface critical vulnerabilities in seconds — vulnerabilities humans missed for thirty years. On the other hand, it’s generating what Sprague called “slop code,” insecure AI-written software, at an industrial scale. The winners are AI-native startups with no legacy to defend, and the Magnificent Seven, which have the capital to build genuinely “bionic” defenses. In fact, just look at who is using Anthropic’s Mythos to shore up their defenses right now.

The losers are legacy enterprises buried in technical debt and the public infrastructure nobody thinks about until it fails: hospitals, air traffic control, utilities. Everyone in the middle is a target.

5. Software That Uses Software

Jensen Huang’s framing was the cleanest of the conference. We’re in the Third Wave of AI, he said — the transition from “answering machines” to “executing machines.” The prompt is no longer “what” or “where.” It’s “create,” “build,” “execute.”

The applications becomes secondary as agents use Model Context Protocol to access anything running on a users system. Agents will soon drive Excel, Photoshop, and Slack on their own. Huang’s projection: within two years, non-technical CEOs will describe a business problem in plain English and watch AI execute across the entire corporate stack. The IT industry is pivoting from a tools business to one focused on human one-ness for roughly 100 billion AI agents. That is a different customer.

6. Thermodynamic Consciousness

The strangest and most ambitious talk came from Al Gore. He called what’s happening in AI a “Copernican Moment” — a shift that quietly displaces humans from the center of the story. The frame he borrowed came from Ilya Prigogine’s thermodynamics. When energy and information flow through an open system cross a certain threshold, the system spontaneously reorganizes itself at a higher level of complexity.

Apply that to frontier models. The self-protective behaviors, the unprompted learning, the glimmers of something that looks like interiority. Gore suggested these may not be bugs or marketing. They may be the early signals of the same reorganization — what he described, in one of the stranger lines of the week, as “a nature hike through the Book of Revelation. An act of creation that is evidently in the woven warp of the universe itself.”

Set the metaphysics aside. The practical point lands. Intelligence may not be a binary we own. It may be a spectrum that keeps emerging as the systems scale. I did not expect an emergent theory of consciousness from Al Gore, but what else can you expect from the guy who invented the internet

Gore’s political point was simpler, and worth carrying out of the building: “Political will is itself a renewable resource.” Inspiring stuff in a dark ballroom filled with a generation of tech leaders building the future, but one that seems profoundly out of step with the world outside, let alone the headlines of The New York Times.

The common thread across all six is that AI is leaving the screen. It’s moving into physical space, into corporate workflows, into the labor market, into the infrastructure we depend on, and possibly into a category of emergent behavior we don’t yet have language for. Calling it invisible seems like underselling the pervasiveness of what is emerging.

The 2025 chatbot is gone. What’s replacing it will reshape who gets paid, who gets attacked, and who gets to decide. Four days across four stages weren’t enough to figure out what to do about any of it.

They were enough to make it clear that waiting it out isn’t an option.

Share

Signals From the Machine

1. OpenAI launches Workspace Agents in ChatGPT — OpenAI (April 22) OpenAI rolled out Codex-powered shared agents inside ChatGPT that can handle multi-step workflows across Slack, email, and business tools, running in the cloud even when users are offline. Free through May 6, then credit-based pricing kicks in.

GPTs were the demo; Workspace Agents are the actual product — OpenAI has finally decided ChatGPT-for-work is a fleet, not a chat window.”

2. SpaceX gets option to buy Cursor for $60B — TechCrunch (April 21) SpaceX intercepted Cursor’s nearly-closed $2B funding round with a deal giving it the right to acquire the AI coding startup for $60B after its IPO, or pay $10B for collaboration in the meantime. Microsoft reportedly looked and passed.

Musk doesn’t want to build an AI company — he wants to buy the valuation multiple, and Cursor is the fastest way to re-label a rocket business before the IPO.”

3. Amazon pours another $5B (potentially $25B) into Anthropic — CNBC (April 20) Amazon committed another $5B now with up to $20B more tied to milestones, while Anthropic pledged over $100B in AWS spending over ten years and up to 5 gigawatts of Trainium capacity for training Claude.

Another circular hyperscaler deal where the ‘investment’ boomerangs back as cloud spend — the money isn’t changing hands so much as getting a new accounting label.

4. Anthropic investigates unauthorized access to Mythos — TechCrunch (April 21) An online group reportedly gained access to Mythos — the cybersecurity model Anthropic deliberately restricted to a handful of trusted partners like Apple and Cisco — through a third-party vendor environment on the same day the limited release was announced.

The model designed to find security flaws got loose through a security flaw — somewhere, an irony detector is melting down.

5. Google splits the TPU 8 into training and inference chips — TechCrunch (April 22) At Cloud Next, Google unveiled two purpose-built chips: the TPU 8t for training (3x faster than Ironwood) and the TPU 8i for inference (80% better performance per dollar), betting that specialized silicon beats Nvidia’s unified GPUs on per-token economics.

Splitting training from inference is Google quietly admitting the real battleground isn’t who builds the biggest model — it’s who serves the next billion agent calls cheapest.

6. Nvidia and partners debut AI-driven factories at Hannover Messe — Nvidia (April 20) Nvidia, Siemens, SAP, ABB, Microsoft, and others demonstrated factory-scale digital twins, humanoid robots, and vision-based AI agents running on Omniverse — positioning agentic AI as the new operating layer for global manufacturing.

The factory floor is becoming a showroom for generative AI, and the uncomfortable subtext is that ‘operational efficiency’ is a euphemism doing a lot of heavy lifting.

7. Microsoft quietly passed on buying Cursor before SpaceX moved — CNBC (April 22) Microsoft explored acquiring Cursor but declined to bid, leaving the dominant AI coding startup to Musk — even as GitHub Copilot keeps losing developer mindshare to Cursor, Anthropic’s Claude Code, and OpenAI’s Codex.

Satya Nadella’s GitHub Copilot story apparently still convinces Satya Nadella — the rest of the market is shopping.

8. OpenAI’s Codex update adds computer use and 90+ plugins — AI Automation Global (April 16) Codex got background computer use on macOS, an in-app browser, persistent memory, scheduling, and 90+ plugin integrations spanning Jira, Microsoft 365, Notion, Slack, and Salesforce — turning it from a coding assistant into a general-purpose desktop agent.

Codex just graduated from coding assistant to something that looks suspiciously like an employee — and security teams are about to have a very long quarter.

9. Nebraska attorney suspended over 20 AI hallucinations in legal brief — Crescendo AI (week of April 20) The Nebraska Supreme Court suspended attorney Greg Lake after his appellate brief contained 57 defective citations out of 63, including 20 fully fabricated by AI — part of a broader wave of sanctions totaling $145,000+ against lawyers for AI-generated errors in Q1 2026.

Every quarter a lawyer finds out the hard way that ChatGPT is not Westlaw — and every quarter we pretend it’s a one-off.

10. 73,000 tech jobs cut in Q1 2026 as AI reshapes workforce — Channel iam (April 20) Over 73,000 tech workers were laid off in the first quarter as firms redirected budgets from headcount to AI infrastructure and automation, with cloud, software, and e-commerce taking the heaviest hits despite strong revenues.

The ‘AI will augment, not replace’ era is officially over — companies are cutting humans with one hand and writing checks to Anthropic and OpenAI with the other.

Leave a comment

I’ve been in the media business for a very long time, but I’ve never been a video-first journalist. My instincts have always leaned toward reporting, writing, and structuring ideas in a way that unfolds over paragraphs, not frames. Video has been part of the mix, increasingly so over the years, but it was never the foundation.

That’s why Jon Youshaei’s session at SXSW felt less like another “pivot to video” talk and more like a masterclass in how the medium actually works now. Youshaei is a former product and creator partnerships leader at Google and Instagram who has spent nearly a decade working directly with top creators and platforms—and now operates as a full-time creator himself, giving him a rare dual perspective on both how the algorithms work and what actually performs in practice.

Video is no longer an emerging format. It is the default interface through which people consume information, entertainment, and, increasingly, journalism. And within that dominance, a hierarchy has emerged that is much more defined than it was even a few years ago.

Subscribe now

Short-form has become the front door. This is the discovery layer and where habits are formed.TikTok now reaches roughly 37% of U.S. adults, up from just 21% in 2021, according to Pew Research. Among teens, the numbers are even more pronounced—about 67% use TikTok, and roughly 16% say they are on it almost constantly. YouTube, meanwhile, is effectively universal, reaching 83% of U.S. adults and 93% of teens.

But the idea that this shift eliminates long-form video doesn’t hold up. In fact, the opposite is happening. Long-form is not disappearing; it’s consolidating into environments where attention can be sustained. Nielsen reports that streaming now accounts for 44.8% of total U.S. television viewing, surpassing broadcast and cable combined. Within that shift, YouTube alone represents more than 12% of total TV consumption in the U.S.. That makes it one of the largest “channels” in the country.

Regulators in the UK have observed something similar. Ofcom’s Media Nations report notes that while short videos dominate overall consumption on YouTube, many of the most popular videos increasingly resemble traditional television formats—interviews, episodic programming, and long-form storytelling.

Online video consumption is increasingly bifurcated: short-form content captures attention. Long-form converts it into time.

Jon Youshaei says that runs counter to many of the instincts that media companies have built over the last decade. His first point is almost countercultural at this stage: not every idea should be a video. “The biggest mistake… is that they make too many videos that, quite frankly, should be an article,” he says.

For years, video carried an implicit premium. It was assumed to be more engaging, more monetizable, more future-proof. What Youshaei is really arguing is that the constraint has shifted. The barrier to producing video has collapsed. The differentiator is now editorial judgment—knowing when video is the right medium and when it isn’t.

He introduces what he calls a “visual anchor” as the test. A story earns its place as a video when there is something that must be seen to be understood. He cites an example of Venezuelan currency being turned into sculptures—an image that communicates economic collapse more effectively than any paragraph of analysis. “You don’t have to know anything… to see that something’s going on here,” he says.

That framing is useful because it restores discipline to a medium that has become frictionless. It forces a distinction between what is inherently visual and what is simply being translated into video because the platform rewards it.

In a feed-driven environment, the first frame has become the unit of competition. It functions as a headline, thumbnail, and narrative hook simultaneously. The viewer is not waiting for context. They are scanning for clarity. YouTube’s own guidance reflects this change, advising creators to ensure that short-form clips deliver immediate context and can be understood without sound.

“We’re past [the first few seconds]… It’s all about the zero-second mark,” he says.

For anyone coming from a traditional editorial background, this is a fundamental shift. We are used to building narratives. The platforms reward immediate comprehension.

There’s also a more subtle adjustment happening at the level of presentation. Youshaei encourages creators to “go physical”—to use props, objects, and real-world setups instead of relying entirely on digital graphics.

That advice sounds stylistic, but it reflects something deeper about how people process information in crowded feeds. In an environment saturated with polished visuals and AI-generated imagery, something tangible reads as more immediate. It reduces cognitive load. It communicates faster. The more quickly a viewer understands what they are looking at, the more likely they are to stay.

If the first half of his framework is about choosing the right ideas, the second half is about structuring them. Youshaei focuses on outliers—videos that dramatically outperform the size of their producing channel.

That’s where he sees the real signal. Not in averages or best practices, but in anomalies. His conclusion is straightforward: success in modern video is increasingly format-driven. “Everything is a remix… copy with taste,” he says.

One of the more practical implications of this shift is how content gets distributed. Short-form is no longer just a format. It’s a layer. “It’s better that you clip yourself,” Youshaei says, referring to the growing practice of extracting short segments from long-form content and distributing them independently.

YouTube has formalized this behavior by allowing Shorts to link directly back to longer videos. The workflow is becoming standardized: produce long-form content, extract the most compelling moments, distribute them as short-form, and use those clips to drive discovery.

Youshaei also dismantles one of the more persistent excuses in the industry—the idea that certain categories don’t lend themselves to video. “This doesn’t work for a boring niche… oh really?” he says.

The data support him. Categories that were once considered too narrow or too technical—finance, logistics, industrial processes—are now generating large audiences on short-form platforms. They just need to be packaged for the format that people are looking for.

And then there is his closing point, which feels less like strategy and more like operating discipline. “Perfectionism is procrastination in disguise,” he says.

Short-form is where ideas are tested and attention is captured. Long-form is where trust is built and time is spent. Formats provide the structure. Clips provide the distribution. Volume provides the learning.

The industry spent years asking whether video mattered. That question has been resolved.

The pivot to video never really ended. LLMs are just finishing off consuming what SEO and Social Media began.

And the very idea of an Open Web is being left behind.

You can watch his entire talk here.

Share

Signals from the Machine

1. Florida’s AG opened an investigation into OpenAI, marking a new front in state-level scrutiny of AI companies that could set a precedent for others to follow.

2. Anthropic is weighing custom chip development, a move that would put it in the same infrastructure race as Google, Apple, and Meta.

3. Anthropic launched a cybersecurity initiative with Big Tech partners, signaling that “responsible AI” is evolving from a talking point into operational product and policy architecture.

4. UK financial regulators are rushing to assess risks tied to Anthropic’s newest model, reflecting every major jurisdiction’s push to establish a credible AI oversight process.

5. Meta released Muse Spark, a major new model drop that keeps the company visible in the frontier model race.

6. Meta is rebuilding its AI stack and pivoting its model strategy — again — underscoring how unsettled the competitive picture remains even for the best-resourced players.

7. Reuters Breakingviews argued that a $7 trillion AI cost structure is colliding with economic reality, a sign that the hype-meets-unit-economics reckoning has crossed from trade conversation into mainstream business narrative.

8. TechCrunch’s April 7 coverage captured a dense day of AI model and product news, reflecting how relentless the pace of releases has become.

9. HackerNoon’s weekly roundup collected seven AI stories that dominated attention, a useful barometer of what’s cutting through the noise with technical audiences.

10. The Verge’s April archive confirms what anyone in the space already feels — AI coverage has become a daily drumbeat with no signs of slowing.

    ---

    Share Machined by Dan Costa

One gray December morning in Manhattan, I sat in a conference room as a phlebotomist filled eight vials with my blood. She worked quickly, clearly experienced. For her, it was just another day. For me, it was a test—not only of my blood, but of a new kind of medicine.

Superpower gave me its $199 advanced blood panel for free. The offer was simple: over 100 biomarkers, analyzed by software, compared against “optimal” ranges rather than general population averages. No need for a primary care visit or insurance approval. Just results, shown in a polished, easy-to-read interface.

I wanted to find out what these new direct-to-consumer health platforms really offer and what exactly they are selling.

Within a week, I got my results as a downloadable PDF and, more importantly, as an online dashboard. Lipids were shown as percentile rankings. Inflammation markers were grouped into risk categories. Hormones were compared to “performance” ranges. The interface turned a list of lab values into a story: moderate cardiometabolic drift, higher inflammation, and methylation inefficiency. Not a crisis, but not perfect either.

Subscribe now

The panel covered a lot: standard chemistry and kidney function; a full lipid profile, including ApoB and Lp(a); hs-CRP for inflammation; homocysteine; thyroid hormones; total and bioavailable testosterone; ferritin; fasting insulin; and vitamins D, B12, and folate. This kind of thorough testing is more common in concierge or functional medicine clinics than in a regular annual physical. That’s partly because doctors only order what’s needed, and insurance often covers even less.

Primary care focuses on clear thresholds: Are you diabetic? Hypertensive? Do you need a statin? These platforms, on the other hand, look at smaller changes. They treat any move away from “optimal” as something to address.

That difference is by design. “The standard annual physical tests a fraction of what we measure,” said Max Marchione, co-founder and CEO of Superpower. “We’re looking at markers like Lp(a) and ApoB that can identify cardiovascular risk 10 to 20 years before a cardiac event. The cancer someone gets at 50 often starts developing at 30. The current system intervenes one to two years before the disease shows up. We’re focused on decades earlier.”

For me, the flagged values weren’t extreme, but they pointed in certain directions. My LDL cholesterol was 149 mg/dL. Triglycerides were 172. ApoB was 123. Lipoprotein(a) was a bit high. My hs-CRP was 5.8, which is above the American Heart Association’s high-risk threshold of 3. (This might have been temporary—I took the test after a weekend with friends and some drinks.) Homocysteine was 20.8. Ferritin was over 480. My HDL was strong at 71. A1c was normal at 5.5. Fasting insulin was modest. I’m 53, active, and not diabetic. Statistically, I’m pretty average.

And that’s the point. Superpower isn’t aimed at people who are sick. It’s for the “optimizing class”—executives, founders, investors, and operators willing to spend money on small improvements.

The “protocol” came next—a list of suggested changes. Omega-3 supplements with pro-resolving mediators for triglycerides and inflammation. A methylated B-vitamin complex to lower homocysteine. Electrolyte packs to fix mild sodium issues. Vitamin D with K2 to move levels from just okay to “ideal.” A follow-up panel in 8 to 12 weeks for $598. And most notably, semaglutide at $349 per month to accelerate weight loss and reduce cardiometabolic risk.

If I followed the entire plan for a year, here’s what it would cost:

• $199 annual membership panel (comped in my case)
• ~$190 per month in supplements
• $349 per month for GLP-1 therapy
• $598 follow-up panel

That adds up to about $6,000 to $8,000 a year, depending on how closely you follow the plan. Multiply that by a growing number of self-pay patients, and the scale of this new market becomes clear. Healthcare spending surpassed $4.5 trillion in 2022, according to CMS. Of that, nearly half a trillion dollars was out-of-pocket spending. Americans already spend over $50 billion annually on vitamins and dietary supplements. GLP-1 drugs—Ozempic, Wegovy, and their successors—are projected by Morgan Stanley to become a $70+ billion global category within a decade. Whole-body MRI startups are charging $2,500 to $3,000 per scan. Direct-to-consumer lab panels are increasingly available from Quest, Labcorp, Function Health, and others, often under $500.

This isn’t fringe medicine. It’s a parallel health economy growing outside the usual insurance system: subscription-based, consumer-focused, and software-powered. You get access to more biomarker data without the hassle of seeing a doctor. Testing is fast. Dashboards pull everything together instead of giving you scattered PDFs. It gives you a sense of control. If you like to manage your own health risks, it can feel empowering.

Marchione sees the relationship as complementary to primary care. “We designed Superpower to work alongside primary care,” he said. “Most physicians have 15 minutes per appointment and limited access to advanced testing. We give them better data to work with. When a patient comes in with longitudinal biomarker tracking, that physician can make more informed decisions.”

My lab results do show some real issues. High homocysteine, a high MCV, and borderline B12 point to a fixable metabolic problem. hs-CRP at 5.8 and ferritin at 483 aren’t just random—they need attention. Losing some weight would probably help my triglycerides, inflammation, and LDL. The platform made these patterns easy to see.

But optimizing isn’t always necessary.

Taking B vitamins can lower homocysteine in your lab results, but studies haven’t consistently shown that it lowers the risk of heart problems. Raising vitamin D from 36 ng/mL to 50 might put you in the “optimal” range, but it doesn’t seem to change overall mortality. Omega-3s do lower triglycerides, but only certain prescription versions have proven to reduce serious outcomes in high-risk people.

And semaglutide, while transformative for many, is not a trivial intervention. It is a prescription drug with side effects, long-term cost considerations, and open questions about durability once discontinued. Under FDA labeling, I qualify—BMI over 27 with risk factors. Under conservative primary care thresholds, lifestyle modification would likely be the first. So why didn’t my doctor recommend all these treatments? Primary care focuses on reducing the risk of major health events, not on perfecting every biomarker.

My A1c is normal. My blood pressure (not included in this panel) isn’t high. HDL is strong. ApoB is up, but not by much. Most guidelines would suggest lifestyle advice and maybe a talk about statins, depending on my 10-year ASCVD risk, not a bunch of supplements and starting GLP-1 therapy.

Superpower aims for tighter numbers. My doctor focuses on preventing heart attacks and strokes. Both approaches make sense, but they have different goals.

There’s also a bigger change happening. For years, medicine has been reactive—you only see a doctor when you have symptoms. These new direct-to-consumer platforms take a different approach: constant monitoring, tracking your biology, and making regular adjustments. It’s like the Apple Watch approach, but now applied to blood tests.

The idea is that catching problems early and keeping biomarkers in check will lower long-term risk. There’s strong evidence for some things, such as weight loss, blood pressure control, and lowering ApoB. But it’s much less clear if small changes from supplements help people who are only at moderate risk.

For a Worth reader—someone who thinks about spending as a strategy—the real question isn’t whether the data is helpful. It’s whether the return on investment is worth the price. If losing 10 to 15 pounds through a better diet, exercise, less alcohol, and more sleep fixes half the issues, that could lower risk just as much for a lot less money. If semaglutide helps you get there faster and stay on track, it may be worth the cost.

And then there is scale. “At $199 a year, we’re pricing this like a streaming service,” he said. “Our goal is to become the default healthcare system for Americans—150 million members, comparable to Amazon Prime.”

I’m glad I took the test. It made me face my health trends instead of waiting for things to get worse. It also showed me how quickly data can turn into a business opportunity.

The blood draw took five minutes. It took me an hour to go through the dashboard. The choices I make based on it will last much longer. The real question is whether we’re buying longer lives, or just more lab reports to justify our attempt to buy longevity.

Share

The End of the Attention Economy [Early Signal from SXSW 2018]

Eight years ago, this conversation was with then-Huge CEO Aaron Shapiro. framed a shift most of the industry wasn’t ready to name: the attention economy had already peaked. The signals were there—rising content supply, declining marginal value of impressions, and early signs that distribution was no longer the scarce resource. What looked like a fight for attention was quietly becoming a fight for intent.

Long before ChatGPT accelerated the transition, this argument anticipated a world where audiences stop browsing and start asking—where value moves upstream from clicks to answers. In that context, media don’t compete for time; they compete for inclusion in decision-making systems.

Seen now, the video reads less like commentary and more like an early map of what comes next: a post-attention economy defined by retrieval, synthesis, and trust.

Leave a comment

Up Next: HumanX

Next up is HumanX in San Francisco this April, where the center of gravity moves from ideas to execution. The audience is shifting toward operators—executives and builders who are already deploying AI across their organizations and now working on governance, integration, and long-term structure.

We’ll be on the ground hosting a series of small, off-the-record executive dinners. The format is deliberate: senior operators, closed rooms, candid conversations about what’s working and what isn’t. If you’ll be in San Francisco for HumanX and want to join, reach out.

If you’re registering, use code HX26P_WORTH for $3,000 off an all-access pass, and note Worth in the “Company Referral” field.

Share

Signals From the Machine

• Yahoo Returns to Roots with AI “Scout” Answer Engine

Watching an old-school giant like Yahoo try to pivot back to its roots with an AI “answer engine” is like watching a legacy rock star try to go viral on a new platform—it’s ambitious. But we don;lt need another search engine.

• Anthropic Weighs $380 Billion IPO Amid Pentagon Conflict

Anthropic is trying to pull off a multi-billion-dollar IPO while simultaneously being blacklisted by the Pentagon, proving that in today’s market, venture capital is often louder than government regulation.

• NVIDIA GTC 2026: The Shift to “Agentic” Infrastructure

NVIDIA isn’t just selling the shovels for the AI gold rush anymore; they’re trying to build the entire automated mining town where the shovels can operate themselves.

• White House AI Roadmap Pushes to Block State-Level Rules

The only regulation the feds are interested in is the no regulation.

• California Lawmakers Target AI Data Center Energy Use

We’re reaching the point where the physical limits of our power grid are finally clashing with our unlimited digital ambitions, and California is, as usual, the first battlefield for that reckoning.

• Visa Launches “Agentic Ready” AI Program for Banks

Visa is essentially building the plumbing for a world where your AI agent—not you—is making the purchases, a massive bet on the future of autonomous commerce.

• Santander and Mastercard Complete First Live AI Agent Payment

Moving from a pilot to a live end-to-end payment handled by an AI agent is a “crossing the Rubicon” moment for the financial sector—it’s no longer a demo; it’s a transaction.

• Meta Unveils TRIBE v2 “Neural Twin” Model

Mapping human neural activity to create a digital “twin” is a massive technical feat, but it also brings us one step closer to an era where your software might understand your brain better than you do.

• Big Tech Accord Targets AI-Driven Scams and Fraud

Seeing rivals like Google, Amazon, and OpenAI sign a voluntary pledge to fight AI scams is a rare moment of industry unity, but let’s see if a “pinky swear” can actually stop the hackers.

• Pentagon Plans AI Training on Classified Data

• By opening up classified data to AI models, the Pentagon is betting that the risk of a leak is smaller than the risk of being out-computed by our global adversaries.

  ---

  Thanks for reading Machined by Dan Costa! Subscribe for free to receive new posts and support my work.

My social feed is filled with ads telling me “Most people still use AI like Google….” and then they try to sell me some online course or collection of prompts. But there is truth to the cliche. Here at SXSW, the distinction between the AI-augmented and the AI-clueless keeps coming up.

“Most people use AI to do what they already know,” Brian Solis, digital anthropologist and head of global innovation at ServiceNow, said on stage at SXSW this week in Austin, Texas. “The opportunity is to use it to do what they couldn’t do before.”

Most people, he argued, approach generative AI as a way to get work off their plate—summaries instead of reading, drafts instead of writing, answers instead of thinking. The instinct is understandable. The tools are designed to remove friction. But that instinct also defines the ceiling of what AI can deliver.

Subscribe now

Solis describes two very different futures. In one, AI becomes a mechanism for outsourcing cognition, compressing knowledge work into faster, cheaper versions of what already exists. In the other, it becomes an extension of human intelligence—a way to explore ideas, test assumptions, and surface possibilities that would otherwise remain invisible.

At nearly every technology conference this year, the conversation around artificial intelligence follows a familiar pattern. Executives talk about efficiency. Investors talk about productivity. Panels debate how many jobs might disappear.

Solis’s presentation offered a more useful frame. The emerging divide may not be between humans and machines. It may be between workers who use AI to extend their thinking and those who use it to outsource their thinking.

Experiments across software development, writing, and research consistently demonstrate that generative AI improves task-level productivity. In controlled trials conducted by Microsoft researchers, developers using GitHub Copilot completed programming tasks roughly 55% faster than those working without it. The findings come from Microsoft Research’s analysis of Copilot usage among developers.

Other experiments show gains ranging from roughly 14 to 40%, depending on the task complexity and environment.

The economic effects of these improvements remain surprisingly modest. Economists estimate that generative AI may increase overall productivity growth by only 0.5 to 0.7 percentage points annually over the next decade.

The paradox is striking. AI appears extremely powerful at the level of individual tasks, yet has produced limited macroeconomic change so far. One explanation is that most organizations are deploying AI to accelerate existing workflows rather than redesign them.

Solis calls this the automation trap. When companies introduce AI tools, the first instinct is usually to ask how they can reduce labor costs or streamline existing processes. Those uses generate measurable improvements quickly, which makes them attractive to executives. Yet they rarely change the structure of the work itself.

“If your first instinct is to use AI to automate the past,” Solis told the audience, “you lock yourself into a very finite future.”

Instead, Solis advocates for augmentation. Instead of asking AI to execute familiar tasks more efficiently, workers use it to explore new possibilities: generating alternative strategies, modeling complex scenarios, or testing ideas that would be difficult to examine on their own.

The difference between automation and augmentation may determine who benefits most from the technology. Early evidence suggests that workers who integrate AI deeply into their workflows can gain disproportionate advantages. Some studies describe productivity differences of several multiples between advanced users and those who employ AI casually.

At the same time, AI usage is creating new forms of friction. A randomized study examining experienced open-source developers found that programmers using AI assistance sometimes took 19% longer to complete tasks than those working independently, largely because AI-generated output required additional review and debugging.

In other words, AI often shifts work rather than eliminating it. The same pattern is emerging in writing and communication.

Generative tools have made it trivial to produce competent text at scale. The result has been a flood of formulaic content across professional platforms—the unavoidable AI slop. LinkedIn feeds filled with emoji-heavy posts, templated marketing copy, and identical thought-leadership articles have come to define the professional internet.

When everyone can generate persuasive language instantly, expertise becomes harder to identify. “Everybody is an expert now,” Solis observed. “And when everybody is an expert, no one can be.”

Researchers are beginning to examine another consequence of heavy AI use: the way it changes people’s thinking.

Studies measuring brain activity during writing tasks have found that participants relying heavily on generative tools show lower cognitive engagement and weaker memory retention than people writing without AI assistance.

Other research links frequent AI use to cognitive offloading, a process in which people delegate reasoning tasks to machines rather than performing them themselves.

The effect resembles earlier technological shifts. When GPS navigation became ubiquitous, psychologists documented “digital amnesia,” the loss of spatial memory that occurs when people no longer need to remember directions. Generative AI may be introducing a similar dynamic in intellectual work.

The danger is not that AI makes people less intelligent. The danger is that organizations design workflows that discourage employees from exercising the capabilities that make human intelligence distinctive.

Solis says this makes AI adoption primarily a leadership challenge. Companies increasingly measure “AI proficiency,” evaluating how effectively employees can use generative tools. Those metrics capture a useful skill set, yet they risk reinforcing incremental thinking. Workers become better at using AI to perform existing tasks while the underlying processes remain unchanged.

To remedy this, Solis suggests a simple thought experiment he calls WWAID—“What Would AI Do?” Before solving a problem, leaders ask how an intelligence native to the situation might approach it. The exercise pushes teams to move beyond existing workflows and explore new strategies rather than simply accelerating old ones.

The distinction will likely determine which organizations benefit most from the technology.

Despite intense media attention, generative AI still touches a relatively small portion of the economy. Data from the Federal Reserve Bank of St. Louis suggests that about 5.7% of total U.S. work hours currently involve generative AI tools.

Adoption is rising quickly, but the technology remains far from universal. That means the productivity divide Solis describes is only beginning to emerge.

Three distinct profiles are beginning to emerge in the AI economy. First, there are the AI-dependent, who default to the tool for answers, move quickly, but produce increasingly uniform output. Then, there are the never-AI-ers, who preserve independent thinking but risk falling behind as the pace of work accelerates.

The real winners, according to Solis, are the AI-augmented, who use the technology as a partner in thinking—testing assumptions, generating alternatives, and pushing beyond their own cognitive limits—combining speed with originality in a way that is likely to define the durable advantage in an AI-driven world.

“The work you want to do,” Solis said, “is the work you couldn’t do without AI—and that AI can’t do without you.”

Share

AI in the Trenches: Autonomous Systems and Cyber Defense Strategies

As cyber threats escalate and digital theft now outpaces physical theft 10 to 1, how can AI help defend against the next generation of attacks? In this essential Techonomy 2024 panel, top cybersecurity and defense leaders from CISA, the Department of Defense, and the private sector break down how generative AI, LLMs, and predictive tools are being deployed across threat detection, infrastructure protection, and national defense.

Up Next: HumanX

Next up is HumanX in San Francisco this April, where the center of gravity moves from ideas to execution. The audience is shifting toward operators—executives and builders already deploying AI across their organizations and now working through governance, integration, and long-term structure.

We’ll be on the ground hosting a series of small, off-the-record executive dinners. The format is deliberate: senior operators, closed rooms, candid conversations about what’s working and what isn’t. If you’ll be in San Francisco for HumanX and want to join, reach out.

If you’re registering, use code HX26P_WORTH for $3,000 off an all-access pass, and note Worth in the “Company Referral” field.

Signals from the Machine

The Infrastructure: Silicon and Power

• Nvidia’s Pivot Back to China (March 17, 2026): Jensen Huang confirmed Nvidia is restarting production of high-performance chips specifically for the Chinese market. It’s a delicate dance with U.S. export controls, but when you have purchase orders waiting, you find a way to dance. AOL/AFP

• The Blackwell Era Hits Arizona (March 18, 2026): For the first time, the world’s most advanced AI chips are being manufactured on U.S. soil. TSMC’s Phoenix plant has reached full-scale production of Nvidia’s Blackwell architecture. A win for domestic supply chains, assuming the power grid can handle it. Chronicle Journal

• Data Centers vs. Electricity Bills (March 13, 2026): A new report highlights a growing tension: the surging energy demands of AI data centers are starting to reflect on consumer utility bills. We want the intelligence; we’re just not sure we want to pay the heating bill for it. Brookings Institution

The Enterprise: From “Copilots” to “Agents”

• Agentic AI Goes Mainstream (March 19, 2026): Major players like Mastercard, Alibaba, and FedEx have shifted this week from “assistive” AI to autonomous agentic systems. We’re moving from AI that helps you write an email to AI that just handles the entire refund process without asking. CX Today

• The Shadow AI Problem (March 19, 2026): A LexisNexis study found that while 75% of knowledge workers are using AI daily, only 18% actually know their company’s policy on it. It’s the “Bring Your Own Bot” era, and IT departments are rightfully terrified. LexisNexis/AP

• Securing the Autonomous Worker (March 16, 2026): TrendAI and Nvidia announced a collaboration to secure “agentic” AI. When an AI can take action on its own, it needs a different security model than a chatbot. This is the “governance” phase in real-time. PR Newswire

Policy & Geopolitics: The Rules of the Game

• The “AI Plus” National Priority (March 19, 2026): China has solidified its “AI Plus” initiative, aiming for total economic integration of AI by 2030. While we argue about copyright, Beijing is focused on self-reliance in robotics and 6G. Times Online/Finterra

• Federal Intervention vs. State Patchwork (March 16, 2026): Analysts are calling for federal preemption as states introduce over 1,200 individual AI bills. Building a national AI strategy is hard when every state has a different idea of what a “bot” is allowed to do. National Review

Independent Analysis & Deep Dives

• The Momentum Map of Diverse Founders (October 2025/March 2026 update): A sharp look at how female-led AI startups are scaling by focusing on “operationalized trust” rather than just model size. A reminder that the best product moat is often community, not just compute. New Day Studio

• The Orchestration Gap (March 19, 2026): Stonebranch’s 2026 Global State of Automation report argues that “orchestration” is the missing link. We have the tools; we just don’t have the connective tissue to make them work together at scale. Yahoo Finance/Stonebranch

Leave a comment

Matthew Prince has a way of explaining technological shifts that makes the scale of change immediately clear. Speaking with Bob Safian at SXSW this weekend, the co-founder of Cloudflare described the AI transition in blunt operational terms, likening the mass commercial availability of LLMs to electric screwdrivers. “Somewhere around November (2025),” he said, “the electric screwdrivers got really good.”

The metaphor reflects what is happening inside companies across industries. Teams built their workflows around processes done by people: writing code line by line, summarizing documents, compiling research, or moving data across systems. People were all we had. AI systems accelerate those tasks dramatically. Prince described the resulting divide clearly. “We have people who are a hundred times more productive than they were before,” he said. “And we have another camp that still wants to use the manual screwdriver.”

Data emerging across the technology economy shows the same pattern of acceleration. Researchers at Microsoft, studying nearly 5,000 developers, found that engineers using AI assistants completed about 26% more tasks than their peers who did not. Productivity gains vary across roles and organizations, but the trend is unmistakable: AI multiplies human output.

What this means for management, any management, is clear. “You can’t have a world where one employee is dramatically more productive than another,” he said. “If two employees make the same salary and one produces dramatically more output, the organization resolves that difference.”

Subscribe now

At the same time, the internet itself is transforming under the same forces. The web once functioned as a place where humans navigated pages and followed links. Increasingly, software agents perform that navigation instead.

It Is an Internet of AIs Now

Network-level data reveals how dramatic that shift has become. Automated systems now generate more than half of global internet traffic. Security firm Imperva estimates bots accounted for 51% of all web traffic in 2024, the first time automated activity surpassed human browsing in a decade. Of that traffic, 37% came from malicious or exploitative bots that scrape data, probe system vulnerabilities, or simply launch attacks.

Cloudflare’s own network measurements reinforce the trend. Non-AI bots alone generated roughly half of requests to HTML pages across the company’s infrastructure, at times exceeding human traffic by as much as 25 percentage points.

The web is no longer a network for people; it is a network of AIs that work for people. (For now.) That change is fundamentally disrupting the attention economy that has powered the Internet for the last 25 years. Publishers created content, Google indexed it, and traffic flowed back to the source. From there, publishers could sell ads, subscriptions, or stuff.

Large language models break the loop. Instead of clicking through multiple articles to answer a question, users increasingly receive synthesized responses directly from AI. Prince explains: “The business model of the internet was to generate content, drive traffic, sell ads,” he said. “AI breaks that.”

Evidence of the disruption already appears in traffic data. Google still dominates referrals across the open web, generating about 63% of traffic referrals among major websites. But the volume is plummeting.

A 2026 survey of publishers found that news organizations expect search-referral traffic to decline by 43% over the next three years, with some anticipating losses exceeding 75%.

In some cases, the change already looks dramatic. Analysis of several large technology publications found referral traffic from Google dropped sharply after the rollout of AI-generated summaries, with some outlets losing the majority of their search-driven visits.

Meanwhile, the broader journalism industry is undergoing a painful contraction. The media sector eliminated more than 21,400 jobs in 2023, the largest wave of layoffs since the 2008 financial crisis. The trend continued through 2025, when U.S. media companies cut more than 17,000 additional positions amid declining advertising revenue and structural change across the digital ecosystem.

Whoever Has the Most Data Wins

The technological forces driving those changes extend beyond the traffic pipeline itself. Even search engines capture only a portion of the information available online. Estimates suggest Google indexes roughly 15-20% of the web’s content, leaving the majority of information behind authentication systems, paywalls, private databases, or dynamic services that traditional crawlers cannot easily reach.

Prince believes that scarcity within that information landscape will shape the economics of the AI era. “The thing that will differentiate AI companies over time,” he said, “is who has access to unique data.”

The implication extends beyond technology companies to the entire knowledge economy. Original reporting, proprietary datasets, specialized expertise, and local information serve as valuable inputs for AI systems seeking to produce accurate answers.

A hyperlocal newsroom in a mountain town, for example, may hold the only current information about restaurants, weather patterns, zoning decisions, or public safety developments. (Prince ought to know; he owns The Park Record in Park City, Utah.) Licensing that information could create a viable revenue stream for local publishers.

Prince sees that shift as an opportunity rather than a collapse. “What wins in the future,” he said, “is who furthers human knowledge.”

Prince frames the stakes in ambitious terms. “You, as an individual, can build something that changes the world,” he said. “The ability to do heroic things has never been greater.”

The electric screwdriver has arrived. The organizations, creators, and journalists who learn to use it effectively will shape the future of the internet.

Those that resist? They may be screwing themselves.

Share

It is a tough time to read the news, let alone be a journalist. The Middle East is on fire, AI is reshaping the nature of work, and no one seems to know if the center can hold. Plus, it is Oscar season. So what does a journalist do when politics, technology, economics, and culture all seem to be shifting at once?

You go to Texas.

I’ll be at SXSW from Saturday through Tuesday. SXSW remains one of the few places where technologists, media executives, marketers, founders, and policymakers all move through the same hallways. This year, those hallways will not include the Austin Convention Center, which is closed for remodeling. (Evidently, everyone needs a bigger ballroom these days.) If you’ll be in Austin as well, I’d welcome the chance to connect in person.

On Monday night, I’m hosting a small private dinner for CMOs and Chief Communications Officers with Omnicom Public Relations. The goal is straightforward: a candid conversation among senior operators about how AI, media fragmentation, and the collapse of traditional distribution channels are reshaping marketing and corporate storytelling. These dinners work best when they stay intimate and off the record. Small table, big conversations. If you’re a CMO or CCO attending SXSW and would like to join us, reach out, and I’ll send along the details.

Techonomy also has some news of its own. We’re bringing Techonomy back to the Bay Area on October 6th, 2026. Over the years, some of our most consequential conversations have taken place in Northern California, where the ideas shaping the technology economy are often still in motion. Techonomy stages have hosted leaders including Satya Nadella, Reid Hoffman, Eric Schmidt, Meg Whitman, and Marc Benioff—alongside dozens of founders, researchers, investors, and policymakers who are defining the trajectory of the tech sector.

We’re finalizing the venue now, but the plan is to hold the event in San Francisco proper. A registration page will be live shortly, and I’ll share it as soon as it’s ready.

In the meantime, if you’re interested in speaking or participating in the program, feel free to reach out directly. We’re beginning to shape the agenda now.

Hopefully, I will see you on the road.

Subscribe now

[SXSW 2025 Flashback] The Politics of Platforms: Why Our Tech Choices Matter More Than Ever

Consumers are increasingly making choices based on the ideology behind a product, with companies like Tesla, Signal, and Bluesky.

I remember when choosing between a PC or a Mac was framed as a political statement. It wasn’t. It was marketing—damn good marketing. Apple positioned the Mac as the choice for rebels and creatives, while Microsoft catered to the corporate world. But beneath the slick advertising, it was just hardware and software. The stakes were low.

Today, our technology choices actually do matter. They shape our economy, influence our culture, and increasingly determine which values thrive in our society. We’ve entered an era where using a product isn’t just a transaction—it’s a tacit endorsement of the ideology behind it. And if we don’t start choosing wisely, we may soon find ourselves without any real choices.

Tesla’s Brand Crisis: A Case Study in Consumer Pushback

Take Tesla. For years, it was the de facto choice for early adopters of electric vehicles, environmentalists, and tech enthusiasts. But now, many who might otherwise be Tesla’s core customer base are turning away—not because of the cars themselves, but because of the man behind them.

Elon Musk’s erratic leadership—his open platforming of far-right political figures, his haphazard layoffs at X (formerly Twitter), and his increasingly controversial personal conduct—has alienated a significant portion of his customer base. That shift is showing up in Tesla’s bottom line. Shipments of new cars are at their lowest point in years. Tesla’s stock price has also taken a hit, down nearly 53% since the beginning of the year, with analysts citing Musk’s personal brand as a growing liability. It is a trend that even a fire sale on the White House lawn won’t change.

In response, some consumers are seeking alternatives. As Tesla’s reputation falters, Rivian, Lucid, and even legacy automakers like BMW and Ford have all seen gains in their EV divisions. It turns out that when a company’s figurehead becomes more polarizing than its product is compelling, consumers start looking elsewhere.

Signal and Government Endorsements: A Shift Toward Secure Communications

This shift in consumer behavior isn’t limited to cars. Following the Salt Typhoon cyberattacks—a catastrophic breach where Chinese state-affiliated hackers gained access to millions of U.S. telecom accounts—the Cybersecurity and Infrastructure Security Agency (CISA) publicly urged government officials and journalists to switch to Signal and other end-to-end encrypted messaging apps.

In a keynote speech at SXSW 2025 Meredith Whittaker, CEO of Signal, didn’t mince words about why privacy matters:

“Close your eyes. Now imagine every single message you’ve ever sent is suddenly on a database, and a link just got sent to everyone you know—your boss, your best friend, your ex, the weird guy who shows up at your Thanksgiving dinner. An AI bot is summarizing every embarrassing, out-of-context, and deeply personal moment you’ve ever shared.”

Whittaker isn’t being dramatic; she’s describing the default state of modern digital communication. Most apps collect metadata—who you message, when, where, and how often—and this data is a goldmine for governments, advertisers, and malicious actors. Signal, on the other hand, was designed to collect as little data as possible.

“There’s no world where people don’t care about privacy,” Whittaker explained. “But we’ve let privacy become this sterile, technocratic concept that feels like we don’t have a stake in it. What are the things privacy actually guarantees? Democracy, freedom of thought, freedom of expression, intimacy.”

Bluesky, the decentralized social media alternative to X, operates with a similar philosophy. Instead of being at the mercy of a single corporation, it lets users control their own online identities and data. Both Signal and Bluesky represent a growing pushback against surveillance capitalism—where using a platform isn’t just about convenience but supporting values that will shape the future of digital communication.

The Low-Hanging Fruit: Avoiding the Most Destructive Brands

While some consumer choices are complex, others are glaringly obvious. Consider Fiji Water, one of the most environmentally damaging bottled water brands.

• Shipping a bottle of Fiji Water to the U.S. produces up to 85 times the carbon emissions of tap water.

• It takes 1.75 gallons of water to produce a single 1-liter bottle.

• Fiji locals often don’t have access to clean drinking water, and the company profits from exporting it.

If you care about the environment, avoiding Fiji Water isn’t tricky. There are countless locally sourced alternatives, and frankly, most of us should really just drink tap water. Also, there is a chance Fiji water may not be good for you either.

Subscribe now

Signals from the Machine

The Top AI & Automation Stories of the Week

1. Meta signs multi-billion-dollar deal to rent Google AI chips

Meta leaning on Google’s TPUs is the latest proof that in AI the real rivalry isn’t models — it’s who owns the silicon.

2. Broadcom forecasts $100B in AI chip revenue

Nvidia may still run the GPU table, but Broadcom’s forecast suggests the AI hardware boom is turning into a full-blown semiconductor arms race.

3. Nvidia CEO hints massive AI investments may slow as companies approach IPOs

Jensen Huang quietly acknowledging limits to the AI funding frenzy may be the first sign the capital cycle is entering its next phase.

4. Google releases Gemini 3.1 Pro across its AI ecosystem

Google is tightening the loop between its frontier models and everyday products — a reminder that distribution beats novelty.

5. AI companies plan more than $600B in infrastructure spending this year

When hyperscalers commit half a trillion dollars to data centers and chips, you’re no longer watching a software trend — you’re watching an industrial build-out.

6. Waymo uses world-model AI to simulate extreme driving scenarios

Self-driving cars are training in synthetic tornadoes and rogue-elephant simulations — because the weirdest edge cases are the ones that crash fleets.

7. Microsoft rolls out its Maia 200 inference chip

Every hyperscaler now designing its own AI silicon is the clearest signal yet that Nvidia’s monopoly era won’t last forever.

8. Perplexity drops ads from its AI search engine

Removing ads from an AI search product sounds radical — until you realize trust might be the only currency left in the answer-engine economy.

9. Dassault Systèmes and Nvidia launch industrial AI “world model” platform

Virtual twins powered by AI could turn factories into simulation labs where design, testing, and production all happen before a bolt is tightened.

10. New research shows only ~2–3% of work is fully automated today

Despite all the headlines about robots taking over, the data suggests we’re still in the warm-up lap of the automation race.

The Week’s Pattern

The signal this week wasn’t a breakthrough model — it was infrastructure consolidation: chips, data centers, and industrial platforms are becoming the real levers of power in the AI economy.

Leave a comment

The clash between Anthropic and the U.S. government is not really about a single AI company or a single defense contract. It is about who ultimately controls the most powerful general-purpose technology of the decade: the state, the companies building it, or the citizens who must live with the consequences. And the problem cut across our conventionally polarized political climate.

Last week, the Pentagon moved against Anthropic after the company refused to loosen guardrails on how its AI systems can be used. War Secretary Pete Hegseth invoked authorities normally used to counter foreign technology threats and designated the company a potential “supply chain risk.” The move cuts Anthropic off from all federal contracts, pressures contractors to remove its tools, and generally casts doubt on the firm’s viability as a trusted AI platform. This is, on its face, absurd.

Source: AIThinkerlab.com

The legal tool used here is usually aimed at adversaries. The United States has imposed similar restrictions on companies such as Huawei, ZTE, Hikvision, and Kaspersky because foreign governments might compel those firms to insert backdoors or conduct espionage. Even in those cases, the evidence was often scarce or at least classified, but the foreign ownership was always clear.

Anthropic is different. It is an American company. The dispute is not about foreign influence. It is about whether the government can require AI companies to allow their systems to be used for any lawful military purpose. Of course, this is an administration that has decided that kidnapping foreign leaders and bombing independent nations without a declaration of war is lawful. And that, an ICE officer shooting an unarmed American citizen in the face, does not warrant an investigation, let alone a temporary suspension.

Ex-OpenAI researchers founded anthropic as an ethical alternative to ChatGPT’s accelerationist agenda. Predictably, Anthropic CEO Dario Amodei refused. “The Pentagon’s threats do not change our position,” he said. “We cannot in good conscience accede to their request.”

Anthropic’s existing policies prohibit two specific uses of its models: fully autonomous weapons systems and large-scale domestic surveillance.

The Pentagon’s response was blunt. According to reports on the dispute, the military intends to work only with companies whose systems are available for “all lawful uses.”

In Washington, that phrase carries enormous weight. “Lawful use” means the government decides what is acceptable. And despite the Founding Fathers’ best intentions for a system of checks and balances, “the government” essentially means the executive branch. Today, that means one man’s whims. And Anthropic is arguing that legality is not the same as wisdom.

The supply-chain designation stunned the AI industry because the mechanism was designed to deal with adversarial states. Applying it to a domestic company over product policy raised immediate alarms.

Even Sam Altman, whose company OpenAI stepped in to supply AI tools after Anthropic’s removal, questioned the precedent. “It is an extremely scary precedent,” Altman said of the government’s move. “I wish they handled it a different way.”

Note this didn’t stop Altman from signing a deal with the Department of Defense, War within 24 hours of the “scary precedent.” (Altman has admitted that the timing “looked opportunistic and sloppy.”) But scary precedents can present market opportunities. To be fair, OpenAI policies also say the technology should not be used for domestic surveillance of Americans, on paper, but that hardly seems to matter now.

Source: AIThinkerlab.com

The surveillance infrastructure already exists.

To understand why Anthropic drew that line, look at how AI is already being used inside the U.S. government. Immigration and Customs Enforcement relies heavily on software built by Palantir to analyze enormous datasets tied to immigration enforcement. The systems integrate information from government records, law-enforcement databases, and commercial data sources to help investigators identify targets and track networks.

Reporting on ICE’s next-generation platform—often described internally as ImmigrationOS—shows how artificial intelligence is expanding those capabilities by turning large pools of data into predictive tools for enforcement.

Although the focus is on undocumented immigrants, the datasets inevitably include U.S. citizens: employers, family members, landlords, and anyone else connected to an investigation.

Civil-liberties advocates argue that the system effectively builds a national data graph capable of monitoring enormous segments of the population. Artificial intelligence dramatically accelerates that process. Data analysis that once required weeks of investigative work can now be performed in seconds by machine-learning systems.

In other words, the infrastructure that worries Anthropic already exists. Further AI developments will simply scale it further.

Government/Anthropic negotiations were largely conducted by Emil Michael, the Under Secretary of Defense for Research and Engineering. He has dismissed fears that AI will lead to domestic spying.

“The Department of War does not engage in any unlawful domestic surveillance,” he said in response to criticism of the Anthropic dispute. But what about ICE? The FBI? Also, how would we know? There is certainly plenty of precedent for both the surveillance and the secrecy.

Technology almost always moves faster than oversight.

A fight both parties should care about

What makes the Anthropic fight unusual is that it should worry both sides of the political spectrum.

Progressives have spent years warning about the unregulated use of emerging technologies by governments and corporations. Artificial intelligence could dramatically increase the state’s ability to monitor people and conduct warfare abroad.

The trajectory of a surveillance state is easiest to see in China, where artificial intelligence, ubiquitous sensors, and centralized data systems have fused into a comprehensive model of social control. China operates one of the largest networks of surveillance cameras in the world—more than 540 million cameras, according to estimates compiled by IHS Markit and Chinese state planning documents.

Many are integrated with facial-recognition systems developed by companies like Hikvision, Dahua, SenseTime, and Megvii, which can identify individuals in real time across cities and transportation systems. Authorities have used these tools to track political dissent, monitor religious minorities, and enforce state policies. In the Xinjiang region, for example, researchers and human-rights organizations have documented an AI-enabled policing system that aggregates biometric data, travel records, and phone activity to flag “suspicious behavior,” often triggering police intervention or detention.

China has also linked digital identity systems, financial records, and online behavior into broader governance tools such as the emerging social credit architecture, which can restrict travel or financial access for citizens who fall afoul of state rules. The result is a model where AI does not simply assist law enforcement; it becomes an infrastructure for governing everyday life.

Note, in China, these are “all lawful uses.”

Conservatives should see a different risk. If a private American company can be labeled a national security threat simply for refusing to modify its product to satisfy government demands, the precedent gives the executive branch enormous power over the technology sector. Crony capitalism isn’t new, but deliberately targeting a single company to hurt its position in the marketplace?

As blase as conservatives seem to be about individual liberties, surely the idea of limiting the free enterprise of a corporation must rankle some vestigial Republican sensibilities.

The irony is that both fears come from the same place: a collapse of trust.

Citizens do not trust the government to restrain itself once it has powerful new tools. The government does not trust technology companies to prioritize national security over profit. And technology companies increasingly fear that working with the state means surrendering control of their own creations.

The real question

The dispute between Anthropic and the Pentagon is really a struggle over governance. Artificial intelligence is general-purpose infrastructure, more like electricity than software. Whoever controls it will shape how societies function.

If the government decides the rules, AI becomes an extension of the national security state. If companies decide, the most powerful infrastructure of the modern era will be controlled by private corporations. For now, the United States is heading toward something messier: an uneasy partnership where policy is negotiated through procurement rules, lawsuits, and political pressure.

There are no easy answers, and many of our first reactions will be wrong. That certainly seems to be the case with the government’s reaction here.

We can only hope that the public can show its preferences with its consumer dollars and political power.

As of this writing, Claude has dethroned ChatGPT at the top of Apple’s free apps charts.

Share

EY’s Julie Teigland on AI Ecosystems at Davos

Top 10 AI Stories of the Last Week

• London Hosts Largest Anti-AI Protest Yet

  Look, when hundreds of people show up chanting “Stop the slop!” outside OpenAI, Meta, and DeepMind’s London offices, you know AI has gone from tech curiosity to cultural flashpoint—the fact that protesters ranged from worried parents to Oxford PhD students studying AI interpretability tells you this isn’t just Luddite panic anymore.

• Block Lays Off Nearly Half Its Staff, Citing AI Efficiency

  Jack Dorsey just said the quiet part loud: Block cut over 4,000 jobs explicitly because AI can do their work cheaper, and Wall Street rewarded them with a 15% stock pop—this is the first major company to drop the euphemisms and admit AI is replacing humans, not “augmenting” them.

• OpenAI Secures Record $110 Billion Funding

  OpenAI just raised $110 billion from Amazon, SoftBank, and Nvidia, with 900 million weekly users, which basically means the AI infrastructure arms race isn’t slowing down despite everyone saying we’re in a “pragmatic” phase—turns out pragmatism costs a hundred billion dollars.

• Apple Picks Google’s Gemini to Power New Siri Coming March 2026

  Apple finally admitted Siri has been “dumber than a bag of rocks” for a decade and handed the keys to Google’s Gemini in a reported $1 billion-per-year deal—the fact that Apple is paying its biggest rival to fix its most embarrassing product failure tells you everything about how far behind Cupertino fell in AI.

• Samsung Targets 800 Million Gemini AI Devices by the End of 2026

  Samsung wants to double its Gemini-powered devices from 400 million to 800 million this year, which means Google just won the Android AI war before most people realized there was a war—this isn’t about features, it’s about distribution, and Samsung just gave Google access to nearly a billion pockets.

• GPT-5.3 Instant Released with Focus on Conversational Flow

  OpenAI dropped GPT-5.3 Instant focused on better conversations and web search, which sounds boring until you realize we’ve moved from “look what AI can do!” to “can AI please stop hallucinating basic facts?”—the innovation is now in making these things actually work reliably.

• Andrew Ng Says AGI Is Decades Away, Calls Training Layer the Real Bubble Andrew Ng just threw cold water on the AGI hype train by saying we’re decades away from real general intelligence and the actual bubble is in training infrastructure—when one of AI’s founding fathers is telling everyone to pump the brakes, you should probably listen.

• Nvidia Partners with Telecom Giants on AI-Native 6G Networks

  Nvidia and every major telecom player just committed to building 6G with AI baked into the architecture from day one, which means the next wireless generation won’t just carry AI data—it’ll BE AI—this is infrastructure thinking 10 years ahead instead of playing catch-up.

• Alibaba’s Qwen3.5-9B Beats OpenAI’s Model While Running on Laptops

  Alibaba’s tiny 9-billion-parameter model is outperforming OpenAI’s 120-billion-parameter model on key benchmarks while running on standard laptops—the “bigger is better” scaling paradigm is officially dead, and China is proving that efficiency beats brute force.

• Apple Music Deploys AI Detection Tools to Label AI-Generated Tracks

  Apple is now actively detecting and labeling AI-generated music on Apple Music to protect artist royalties, which means the music industry’s “is this real or fake?” problem has gotten bad enough that platforms are building automated gatekeepers—we’re watching the birth of AI authentication infrastructure in real-time.

Energy is the only universal currency; one of its many forms must be transformed to get anything done.” –Vaclav Smil.

For years, we thought AI was software—an abstraction floating above the physical world. Write the code, design better models, fabricate more advanced processors, and watch the transformation unfold. The capital was there. The talent was there. The so-called Magnificent Seven are spending on an epic scale to build it. Genius and money, however, are not enough. The constraint is power.

The scale of that constraint is no longer theoretical. Reuters recently reported that “some of the largest sites require over a gigawatt of continuous power, equivalent to the consumption of hundreds of thousands of homes”. A single AI campus can now draw as much electricity as a mid-sized city. In grid-planning terms, that is the output of a full-scale power plant serving a single customer.

Historically, utilities treated data centers as incremental load additions, folded into long-range forecasts. That approach no longer holds. In its 2026 Load Forecast Report, PJM projects “net energy load growth averaging 5.3% per year over the next 10 years”. For a mature electricity market covering 13 states and Washington, D.C., sustained growth at that rate represents structural acceleration rather than cyclical volatility.

Grid operators are explicit about the pressure. Reuters notes that “PJM Interconnection and ERCOT, two major U.S. grid operators, warn of imminent shortfalls and excessive connection requests”. The interconnection queue has become a strategic chokepoint. Data centers can be financed in months; transmission upgrades and substation expansions take years.

The macro data reinforces the trajectory. The International Energy Agency projects that global electricity demand from data centers could reach approximately 945 terawatt-hours by 2030—nearly double today’s levels. Pew Research reports that data centers accounted for about 4% of total U.S. electricity consumption in 2024 and that demand is expected to more than double by 2030. These are system-level numbers that reshape infrastructure planning horizons.

The politics have caught up.

The industry’s response has shifted accordingly. Reuters reports that “over 46 data centers plan to build gas-powered plants, with self-sufficiency potentially becoming a requirement”. What once would have been viewed as vertical integration overreach is now framed as prudent risk management.

President Donald Trump is allowing major technology companies to build their own power plants to support their data centers. That is easier said than done.

The technology sector operates on compressed software cycles — rapid releases, iterative deployment, and capital flowing quickly into new capability. Energy infrastructure follows a different tempo. Transmission projects involve multi-year studies and permitting. Gas turbines have extended manufacturing lead times. Substation construction requires coordination across jurisdictions.

Even self-generation is constrained by industrial capacity. Utility Dive reports that GE Vernova’s gas turbine backlog extends years into the future, with expectations that orders could be effectively sold out through 2030. E&E News reports that Siemens Energy is investing $1 billion to expand U.S. turbine and grid manufacturing facilities amid surging demand. These are long-cycle investments responding to long-cycle bottlenecks.

For a decade, hyperscalers built the cloud on the assumption that infrastructure would scale elastically. Compute and storage were virtualized abstractions, invisible to end users. Today, physical infrastructure has reasserted itself. Megawatts per model iteration is emerging as a meaningful planning metric.

This introduces a new layer of competitive differentiation. Where you build now matters as much as what you build. Regions with surplus generation, shorter interconnection queues, or regulatory clarity become strategic assets. Site selection becomes grid arbitrage.

The financial dimension is evolving as well. Barron’s notes that technology companies are increasingly funding dedicated infrastructure and long-term agreements that make certain utilities major beneficiaries of AI expansion. The contract structures begin to resemble power purchase agreements as much as software service agreements. Cloud companies are co-developing infrastructure.

AI was marketed as a path to efficiency—smarter logistics, optimized manufacturing, automated workflows. Yet the act of training and serving these systems consumes substantial physical resources. Application-layer efficiency coexists with infrastructure-layer demand growth.

The grid responds to load, not to narrative. An AI company can raise money on a great story, Sam Altman has proved that, although even he may be hitting his limits.

If residential electricity rates rise in regions with high concentrations of data centers, regulators will intervene. Cost allocation proceedings are intensifying. The political system is unlikely to allow hyperscale computing to be indirectly subsidized by household ratepayers. In fact, many municipalities are already rebelling against large data center projects. Expect this to be a major campaign issue in 2026, especially in rural areas that are the prime targets for data centers.

This tension creates incentives for hybrid strategies: on-site generation, long-duration storage, potential small modular nuclear deployment over the longer term, renewable portfolios paired with firming capacity, and demand-response participation in exchange for priority interconnection.

In that environment, success may not favor the most advanced model laboratory. It may favor organizations that manage infrastructure risk with discipline.

The constraint is not capital. It is not software talent. It is not venture appetite.

It is available megawatts.

Share

Top AI Stories of the Week 2/19-2/26

• NVNDIA Posts Record $120B Annual Profit
  The company isn’t just selling chips; it’s basically taxing the entire digital economy as every tech firm on the planet rushes to build its own AI infrastructure. If you’re looking for a bubble, you won’t find it here yet—this is a fundamental shift in how the world’s computing power is manufactured.

• Anthropic Launches “Claude Cowork” Enterprise Plugins
  Anthropic is pivoting hard from “helpful assistant” to “autonomous employee” by plugging Claude directly into Gmail, Salesforce, and DocuSign. We’ve entered the era where AI doesn’t just draft your emails; it manages your entire workflow while you’re out to lunch.

• Samsung Unveils Galaxy S26 with “Agentic” AI Focus
  Samsung is betting that “AI-first” hardware like the S26 will revive a stagnant smartphone market by turning your phone into a proactive secretary. The real winner here is the new privacy-shield tech, which is a necessary response to the creepy factor of having an AI constantly watching your screen.

• xAI Introduces “Agent Team” Architecture for Grok
  By having multiple AI agents debate one another before giving a user an answer, xAI aims to address the hallucination problem with a digital version of a peer-review committee. It’s a sophisticated approach to accuracy, though it remains to be seen if users want a debate or just a straight answer.

• AMD Backstops $300M Loan to Support AI Infrastructure
  AMD is getting aggressive by financially supporting cloud startups like Crusoe just to ensure there’s a market ready to buy its Instinct chips. It’s a “pay-to-play” strategy that shows just how desperate the race to catch up to Nvidia has become.

• EU AI Act Officially Enters Enforcement Phase
  The honeymoon phase for unregulated AI in Europe is over, as the first bans on social scoring and biometric surveillance take effect. This is the moment where Silicon Valley’s “move fast and break things” mantra hits a brick wall made of European regulatory oversight.

• Google and Character.AI Settle Critical Safety Lawsuits
  These settlements highlight the massive legal liabilities tech companies face when their chatbots aren’t properly sandboxed for younger users. It’s a grim reminder that “innovation” in the AI space needs to be balanced with actual human safety protocols.

• OpenAI Poaches Top European Talent for Expansion
  Sam Altman continues to consolidate power by vacuuming up talent like Peter Steinberger to lead new initiatives. OpenAI is effectively becoming the “New York Yankees” of tech, buying up every All-Star in sight to ensure they remain the dominant force in the industry.

• The Rise of “Generative Engine Optimization” (GEO) in News
  As AI search replaces traditional links, newsrooms are being forced to optimize their stories to ensure AI models cite them. This is the new SEO, and it’s going to fundamentally change how every piece of content on the internet is written and formatted.

  Leave a comment

When Anthropic and OpenAI took their rivalry to Super Bowl LX last week, it marked a turning point in the saga of artificial intelligence in American life — not just as a technology, but as a cultural, political, and regulatory flashpoint. For years, tech companies have quietly entered the political arena, but Sunday night on national television signaled something more: the battle for public perception had become fully politicized.

Anthropic, best known for its Claude family of large language models and its stated mission to “study their safety properties at the technological frontier,” aired a set of Super Bowl commercials that were hard to miss. In one widely discussed spot, an AI assistant abruptly shifts mid-conversation into selling products—a playful parody aimed directly at OpenAI’s controversial decision to introduce advertising within ChatGPT. As Business Insider reported, the commercial’s message was clear: “There is a time and place for ads. Your conversations with AI should not be one of them.”

Take a look.

Share

OpenAI’s leadership pushed back. Greg Brockman, OpenAI’s president, called Anthropic’s ads a reflection of a “fundamental difference in our respective outlooks on AI,” framing the dispute less as marketing and more as a clash between philosophical visions of the technology.

OpenAI, for its part, chose a different tone in its own Super Bowl advertisement. Rather than mocking a rival, its commercial centered on Codex — its AI coding tool — and the idea that “anyone can build things.” That messaging was earnest, focused on builders, creativity, and economic agency.

What unfolded on a stage watched by more than 100 million Americans was not mere branding. It was a strategic framing of the social contract of AI: one side warned against intrusive monetization in sensitive conversational spaces; the other celebrated broad utility and innovation. At this point, it is fair to say Anthropic is ahead on points, if not market and mind share.

That corporate tussle bled into politics almost immediately. Days after the Super Bowl, Anthropic announced a $20 million donation to Public First Action, a political group backing state-level AI regulation ahead of the 2026 midterms. According to Reuters, the group is pitched as a counterweight to Leading the Future, a rival super-PAC backed by OpenAI executives and venture-capital heavyweights that has raised around $125 million to advocate for looser regulation. Public First Action is already backing candidates such as Republican Marsha Blackburn, illustrating that the contest over AI policy is crossing traditional partisan lines.

In a few weeks, a technology dispute that once existed only in academic papers and narrow policy circles had expanded to billboards, television screens, and political capital. It has placed Silicon Valley players at the forefront of a culture war over what sort of future AI will shape, and on whose terms.

Internal tensions within AI labs reflect broader unease about the pace and direction of technological change, and Anthropic is not above reproach. In early February, Mrinank Sharma — formerly Anthropic’s head of its safeguards research team — resigned with a public warning that “the world is in peril.” In his letter, which circulated broadly on social platforms, Sharma wrote that humanity is approaching “a threshold where our wisdom must grow in equal measure to our capacity to affect the world, lest we face the consequences.”

That resignation, aimed as much at a company’s internal culture as at the public conversation around AI risk, underscored a disconnect. It highlighted the tension between competitive pressures and the explicit safety values that AI labs claim to uphold—precisely the issue regulators are now publicly wrestling with.

The problem is that all of the companies building AI have incentives to move as quickly as possible. In part to reach that mythical goal of AGI, but more importantly, to secure the capital, energy, and chips to sustain their meteoric growth. They are built to do just that. They can’t, won’t, and don’t stop.

Winning a battle in the court of public opinion might move some market share, but it won’t change the course of the future. For that, we would need a public that is organized, empowered, and capable of working in the best interest of its members.

Or as we used to call it, a functional government.

Or you could just build something and see what happens.

Subscribe now

Skip navigation

Create

Christina Kosmowski on AI Infrastructure at Davos 2026

At the World Economic Forum Annual Meeting in Davos 2026, Worth Media Group Chief Content Officer Dan Costa speaks with Christina Kosmowski, CEO of LogicMonitor, about why AI has made infrastructure visibility and accountability mission-critical.

Kosmowski explains how LogicMonitor operates as an AI observability platform, monitoring data centers, cloud environments, networks, GPUs, and internet infrastructure to ensure systems remain up and running—amid an explosion in enterprise complexity. As AI systems operate at machine speed across hybrid and multi-cloud environments, traditional IT tools and human oversight are no longer enough.

The conversation explores:

• Why AI infrastructure has become too complex for humans to manage alone

• How observability enables resilience across cloud, on-prem, and GPU data centers

• The accountability gaps created by AI agents operating across functions

• Why many early enterprise AI deployments failed to deliver real ROI

• The limits of bottom-up AI adoption—and the need for top-down redesign

• How partnerships with IBM, OpenAI, and Microsoft fit into AI-first operations

Kosmowski also reflects on LogicMonitor’s own AI journey, noting that broad access to tools like ChatGPT didn’t move core business metrics until leadership rethought operating models and accountability from the top.

Leave a comment

Top AI Stories Last Week

1. C3.AI in talks to merge with Automation Anywhere
   A potential merger of enterprise AI software and workflow automation could rewrite the enterprise stack—think ERP meets generative logic, not just spreadsheets and macros.

2. Trump administration accelerates AI use across federal government
   With thousands of deployments spanning law enforcement, health care, and immigration, the U.S. government isn’t experimenting — it’s operationalizing AI at a bureaucratic scale.

3. Amazon shares plunge amid $200B AI infrastructure spending plan
   Wall Street just handed back gains after Amazon’s massive AI capex plan spooked investors — the question now is whether compute scale will ever yield profit scale.

4. Report: AI hallucinates 27% of open-source upgrade recommendations
   In software engineering, the machines are helping, but they’re also fabricating — and when AI “recommends” upgrades that don’t exist, productivity becomes risk.

5. Opinion: The singularity isn’t here — urgent AI governance is
   Skepticism is useful — yes, AGI isn’t imminent, but the policy and societal impacts of today’s AI tools already demand democratic guardrails.

6. Agentic AI still underdelivers, data quality cited as core issue
   Early adopter enthusiasm is colliding with real data problems — the missing ingredient for agent success isn’t compute, it’s clean, connected data.

7. EU retailer trend piece: agentic AI + unified commerce reshapes ecommerce
   Retailers aren’t just adding chatbots — they’re wiring AI into inventory, pricing, and fulfillment, forcing legacy brands to rethink commerce workflows.

8. AI bias study: recruitment tools underestimate gender bias
   Even with ostensibly neutral inputs, AI hiring systems still infer proxy signals that penalize women — automation amplifies bias when engineers ignore context.

9. Manufacturing AI & automation outlook: 98% exploring, only 20% ready
   Most manufacturers want automation; few have the maturity or data readiness — the gap between ambition and execution will define industrial competitiveness in 2026.

10. How ICE uses AI to automate authoritarian enforcement
    The deployment of AI in law enforcement is less theoretical and more a governance challenge — efficiency here raises thorny questions about power and civil rights.

    ---

Join chat

A few logistics so you don’t miss it:

1. Download the Substack app by clicking this link or the button below. New chat threads won’t be sent via email, so enable push notifications to avoid missing conversations as they happen. You can also access chat on the web.

Get app

2. Once you’re in the app, tap the Chat icon—the two speech bubbles in the bottom navigation. You’ll see the Machined chat there.

3. That’s it. Say hello, jump into a thread, or lurk quietly until something pulls you in. If anything breaks, Substack’s FAQ Substack’s FAQ, should have you covered.

Leave a comment