Leland Char

New paper from Meta's content moderation team introducing industrial-scale RL for policy-violation classification. Scaled content moderation is a thorny problem: policies are nuanced and generally fluid, data is sparse and expert labels are expensive to produce, and the threat vector is constantly evolving. Meta's content moderation team explored the use of RL to fine-tune policy classification models and found that it outperformed SFT "on tasks requiring complex, policy-grounded reasoning" while achieving up to two orders of magnitude higher data efficiency under constrained labeling regimes. The team also found that the RL approach exhibited sigmoid-like scaling efficiencies as more compute and data were utilized. The paper establishes content moderation as a classification problem: given content C and policies p_1...p_N, the content is prohibited if y_i=1 for any p_i. The team built an RL algorithm using Group Relative Policy Optimization (GRPO), an alternative to Proximal Policy Optimization (PPO) that eschews value functions and uses relative advantages across sampled responses and updates the policy using groupwise comparison. This achieves higher stability in light of data sparsity, which is an acute problem in content moderation. The team uses Monte Carlo sampling to produce a smoother probability distribution of violation scores, since the initial approach produced a bimodal distribution. The team also introduces prompt adjustments that force the model to revisit previous decisions in the reasoning stage. What the team found was that RL underperforms when classification is merely a function of verifiable accuracy, as it is with eg., math problems or code production. In soft-policy regimes like content moderation, reasoning plays an outsized role in determining rewards, and so reasoning completeness becomes the principal source of model viability. The expanded reasoning process increases the probability of selecting a correct response, improves consistency across rollouts, and reduces variance in outcomes. This was verified using best-case correctness (pass@N) and consistency across samples (maj@N). The result is that the RL-only post-trained model outperformed the SFT post-trained model in low-data environments, using 1/10th as many samples. And the team found that increased token usage, rollouts, and compute led to consistent performance gains, although it tapered off at some saturation point. Paper linked below.

13

1 Comment

Caitlin Kalinowski helped build OpenAI's robotics and hardware team (after building Meta's AR/VR hardware teams). Her take on why AI companies are investing so heavily in hardware right now: "There's a dawning realization, especially in the AI labs, that acceleration is going so vertical that what you can do behind a keyboard with AI is going to saturate. When that happens, the next frontier is the physical world. Hardware, robotics, manufacturing, industrialization. The sensing layer in the real world. The ability to move objects in the real world. And eventually, we hope, space." Full conversation: https://lnkd.in/gvni494d

226

37 Comments

Meta's latest acquisition of PlayAI is a game-changer for the future of AI-generated voice tech, bringing us one step closer to seamless communication. #MetaAI #VoiceTech #AIInnovation #TechNews #ArtificialIntelligence #DigitalTransformation

2

Recently, a friend told me he was "long on compute." My first reaction — obviously, isn't everyone long on compute? Then I started actually using Claude more seriously. I don't think the market has priced in what personal agents do to compute demand. People are already buying dedicated Mac Minis just to keep their agents running 24/7. That's not a niche hobbyist thing. That's a signal. We're used to thinking about the compute build-out in terms of hyperscalers and GPU clusters. But the real story might be at the individual level — billions of people running personal agents in the background, continuously. Every individual consuming data center-scale compute. Not shared, not enterprise. Personal. We may be a lot earlier than we think.

24

10 Comments

Founders often ask us: what does it really take to build a durable AI company? In enterprise AI, we’re seeing two models emerge again and again: Oil wells (systems of record) & Pipelines (automation layers). 1️⃣ Oil wells become the source of truth for data & associated workflows. We see two entry points: Rip + Replace: When legacy systems are weighed down by tech debt, buyers are motivated to take the risk to replace with new AI-native systems. hyperexponential rebuilt insurance pricing & underwriting from ingestion to submission; Valon collapsed 25+ mortgage servicing systems into one; Vesta redesigned loan origination so tasks could run in parallel. Greenfield: When no system exists, startups capture customers early and grow with them. Rillet started as the first ERP for SMB finance teams, automating manual workflows, and has since expanded into replacing incumbents like NetSuite. Oil wells take longer to drill, but once established, they create deep, durable moats. Owning the system of record unlocks workflows no one else can build and builds switching costs. 2️⃣ Pipelines (automation/orchestration layers) -- These sit on top of existing systems and automate the “glue work” humans do between them. We’ve seen two main patterns: Fragmented Systems: When many entrenched systems coexist, pipelines unify workflows without requiring rip-and-replace. FurtherAI provides agentic workflows for insurance, automating cumbersome processes (submissions, loss runs, compliance) across multiple systems. Human Middleware: When humans are the bridge between systems, pipelines digitize that work. Concourse builds AI agents for finance teams, connecting into several financial systems so teams can query and analyze without manual effort. Sola lets customers record a workflow once and turns it into a live AI agent for tasks like invoice reconciliation. Customers don’t have to choose. Enterprises will often buy both: a new system of record in one area, lightweight automations in another. But for founders, the strategies are distinct but both can be massive. What matters is not trying to do both at once, but knowing which game you’re playing to win. New post by me & Joe Schmidt IV in comments!

172

20 Comments

The LLM era is peaking. The SLM era is just getting started. Why care? 🐧 Tencent just dropped their 0.5B-parameter Hunyuan Small Language Model that can run offline on a smartphone or edge/constrained device — no cloud, no Wi-Fi, no “please hold while your AI is thinking.” 📈 It’s tiny compared to models like o4 (~200B) or o4-mini (~20B), yet it supports a 256K context window and both “fast” and “slow” thinking modes. Translation: it can produce high-quality outputs while living entirely on-device. Our AIOT portfolio companies like Atym, Mimiq and Tripolar Industries (stealth) are primed to take advantage of this wave. 🏁 Why SLMs > LLMs in certain domains: - Speed: Millisecond responses. - Offline operation: Works in connectivity deserts. - Privacy: No data leaves the device. - Focus: Perfect for specialized tasks. 🤖 Industry will be a killer app for SLMs: - Edge AI in factories: Local analysis of production data without risking IP leaks. - Aerospace & automotive: On-device AI guidance for additive manufacturing. - Frontline productivity: Real-time troubleshooting without a network tether. 🔥 Hot take: Within 3 years, most “AI in manufacturing” will not be powered by giant LLMs in the cloud — it’ll be nimble SLMs at the edge. The next AI arms race isn’t about who has the biggest model. It’s about who can make the smartest model that fits in your pocket. What’s your bet?

192

29 Comments

Most people look at the map and see logos. The real signal is stack depth. Different companies now control different parts of the flow. ⸻ 1. On/off-ramps and orchestration This is the choke point. Firms like Manteca, Mesh, Bridge, Rampa, Lumx, Trace Finance, OpenFX, Antom, PhotonPay, and Payoneer sit between: * Bank rails * Stablecoin rails If access is blocked, nothing else matters. ⸻ 2. Card issuance Spending is the unlock. Kulipa, Offramp, Pomelo, Rain, DPT, Reap, RedotPay, and Stables are solving one thing: Turn stablecoins into everyday payments. Users don’t want to “hold crypto.” They want to use money. ⸻ 3. B2B payments This is where volume lives. Bitso Business, Caliza, Dock, Eluter, Kill B, Kravata, Interlace, XTransfer, and Obita focus on: * Treasury * Supplier payments * Cross-border settlement Retail is noise compared to this layer. ⸻ 4. Remittances and P2P Simple use case. Massive demand. Anclap, Cinko, El Dorado, Buenbit, Lemon, Félix, Vita Wallet: * Faster transfers * Lower cost * Fewer intermediaries This is where adoption starts. ⸻ 5. Wallets and custody No custody → no system. belo, Cenoa, Chainless, Utila, Bitget Wallet, UniPass: * Storage * Access * Control Trust sits here. ⸻ 6. Compliance layer Quiet but critical. BlockSec, Crypto Flow, Bitrace, Fireblocks, Trustin: * KYT * KYB * AML Without this, institutions don’t enter. ⸻ What this tells you: * Stablecoins are moving from trading → payments * Value is fragmenting across the stack * Issuers are only one piece ⸻ Where it’s happening fastest: * Asia * LatAm Not by chance. By necessity. ⸻ The shift: The race isn’t “who issues the coin.” It’s: Who controls access Who owns distribution Who manages trust ⸻ That’s where the real leverage sits.

23

8 Comments

Meta just made AI usage a core line on every performance review. The line that decides comp, promotion, and exit. Most CEOs will copy the words. None will copy the move. Only 3 Fortune 100 companies have actually tied AI usage to compensation. 97 are talking about it. The gap between rhetoric and structural change is where the AI race is being lost — quietly, every quarter. Here's what the 3 understood that the 97 didn't. Mandates are theater. Incentives are policy. You cannot email your way to AI adoption. The CEO who changes one line of the performance review beats the CEO who writes a thousand-word AI memo. Every time. Meta proved it — internal AI usage tripled in 60 days after the policy shift. Frame I'd give every CEO this week: Incentive structure > inspirational mandate. Which side are you on — the 3 who restructured incentives, or the 97 still circulating memos? Drop a number in the comments. I'll respond to the sharpest takes. If your exec team is ready to move from mandate to mechanism, DM "PLAYBOOK" — I'll send the IAW AI-CEO Playbook (free). #EnterpriseAI #AIStrategy #CEOInsights #IntelligentAIWorld

1

Last quarter, my AI inference costs hit $100,000 annualized. I started small. Six months earlier, I was spending $200 a month on Claude. Then I added three agent subscriptions : Codex, Gemini, & Claude Code. I was paying $600 a month. Next I started using AI to transform my todo list into my done list, increasing tasks to 31 per day. $92 daily inference invoices started arriving. Then $400 per month on browser agents. Within two quarters, my inference spend grew from $7,200 to $43,000 to over $100,000 run rate. So I migrated to an open source model. It took a weekend. The key was building the right testing loops : I had six months of historical task data, so I could replay requests through the new model & hill-climb to parity with AI agents working through the night. By Sunday evening, they performed identically. At 12% of the cost. I’m not the only one paying attention to this cost. Technology companies are adding a fourth component to engineering compensation : salary, bonus, options, & inference costs. Levels.fyi pegs the 75th percentile software engineer salary at $375k. Add $100k in inference & the fully loaded cost is $475k. That’s 21% in tokens. The question CFOs will pose : what am I getting for all this inference spend? Can I do it cheaper? If the metric for a new cloud is gross profit per GPU hour, the employee equivalent is : productive work per dollar of inference. For me, the answer is 31 tasks a day at $12k annually. The engineer still burning $100k? They’d better be 8x more productive! Will you be paid in tokens? In 2026, you likely will start to be.

284

70 Comments

Most Web3 projects don’t fail because of the tech. They fail because no one sees them. 📉 I’ve worked with dozens of early-stage Web3 teams over the years, and the pattern is always the same. Smart founders. Solid tech. Strong vision. But zero visibility. Zero momentum. Zero community energy. And it usually comes down to one mistake: "Building in silence." When founders disappear for months, they assume users will magically show up when they relaunch. But Web3 moves too fast. Attention resets every 48 hours. One DeFi client I worked with went from 150 → 4,500 Discord users in 60 days. Not because of ads. Not because of influencers. But because they shared weekly builds, weekly demos, small wins, and even failures. That consistency became their marketing engine. Here’s the simple framework I now give every early-stage team: "The Build-in-Public Framework" Share progress before it’s perfect, so people have something to follow. - Weekly ship updates (even tiny commits count) - Short “what we built this week” posts - Early product demos (rough is fine) - Monthly AMAs or Twitter Spaces - Honest notes about decisions, trade-offs, and lessons You’re not just sharing progress. You’re creating momentum, trust, and community gravity. When you build in silence, you force users to guess. When you build in public, you give them a reason to care. If you're to fix only one thing in your marketing, fix this: start talking to your users long before you’re “ready.” Web3 audiences reward transparency and punish uncertainty. And the earlier you show your work, the faster you attract the right people: partners, contributors, users, and evangelists. If you're an early-stage founder: What’s one update you can share publicly this week? 👇 Read the full breakdown on Medium: [link in comments]

4

1 Comment

Quality beats brute-force slop. Edwin Chen built data-labeling company Surge AI into a $1B+ business without raising a dollar. It outpaces its closest competitor, Scale AI, in revenue, and he did it by obsessing over quality, not scale. In a rare interview, he shared his experience trying to build a sentiment analysis model while at Twitter. The first labelers they hired couldn't grok the slang, the cultural subtext of the tweets. The output sucked. That’s when it clicked: better data beats more data. As AI gets cheaper and easier to ship, the delta between mediocre and great will widen fast. And quality will be the only thing left worth betting on.

10

1 Comment

Jensen Huang founded NVIDIA from scratch in 1993. Now it's valued at $2.687 Trillion🔥 I compiled Jen Huang's 8 of the most practical insights on how to build a successful company👇 1. The only 3 things you need to know about Business Plans: https://lnkd.in/dq58bBxJ 2. What's the most determining factor fo Business Success: https://lnkd.in/d4i3rzSZ 3. How to launch a Successful Product: https://lnkd.in/dARWmAnQ 4. How to maintain the lead and Kill your Competition: https://lnkd.in/d6KuGzTs 5. How to always be the First to Market: https://lnkd.in/dKuMGWqk 6. How to Prioritize your Life: https://lnkd.in/drNAGJcu 7. The Superpower of Entrepreneurs: https://lnkd.in/dYDWHgdk 8. Why AI is really going to create more jobs, not destroy them: https://lnkd.in/dqh9bbRy

30

4 Comments

Apple just made its second-largest acquisition in history, paying $2B for Q.ai, a startup with no revenue.... For context, Q̲.̲a̲i̲. founded in 2022, tackles two issues in human–computer interaction: → Interpreting whispered or silent speech via facial detection → Enhancing audio clarity in noisy environments Before the deal, they were barely known and employed around 100 people. But this wasn’t Apple’s first encounter with Q.ai's founder, Aviad Maizels. In 2013, Apple acquired PrimeSense, his previous company, for ⁓$350M.. That deal helped iPhones move from fingerprint sensors to Face ID. This acquisition follows a similar pattern. Recently, Apple has been facing pressure regarding the shortcomings of its AI. Competitors are moving fast: • Google is embedding Gemini everywhere • Meta shipping AI through Ray Ban glasses • OpenAI partnering directly on new IO devices By acquiring Q̲.̲a̲i̲, they can address these capability gaps.  Particularly in audio processing and human-computer interaction. The deal also reveals their decades-long acquisition strategy: - Datakalab (2023): On-device AI compression - TrueMeeting (2025): 3D avatars for Vision Pro - Pointable (2025): GenAI for spatial content creation Apple engages in vertical integration to create durable competitive advantages. Thus, the $2B price tag sends a clear signal: Apple will deploy capital at scale to ensure leadership.

220

19 Comments

Ex-Google/Meta Execs Raise $100M for AI Hardware Massive news in the AI infrastructure space! Former Google and Meta silicon executives Ofer Shacham, Sha Rabii, and Masumi Reynders have come out of stealth, announcing they've raised a whopping $100 million for their new startup, Majestic Labs. Their mission? To fundamentally reduce the massive data center buildout spend for cloud companies. Here’s the game-changing tech: -- 10x Consolidation: Majestic Labs is developing patent-pending server architecture that they claim can replace up to 10 of today's leading server racks with a single Majestic server. This dramatically reduces footprint, power, and cooling needs. -- Memory Solved: The new silicon design promises 1,000 times the memory of a typical enterprise-grade server, directly addressing the critical memory bottleneck for highly intensive AI workloads—the models that even industry-leading GPUs struggle to handle efficiently. -- Targeting Hyperscalers: While respecting the innovation driven by companies like Nvidia, Majestic Labs is zeroing in on hyperscalers and large enterprises in finance and pharma where data-heavy, memory-constrained AI is a core challenge. This investment, including a $71M Series A led by Bow Wave Capital, signals a huge bet on the next generation of AI infrastructure. It's a clear move to optimize the colossal capital expenditures (currently expected to top $380B this year!) that giants like Alphabet Inc., Meta, Microsoft, and Amazon are pouring into their data centers.

5

Google just put $2.4B behind Google Antigravity - their new agent-first IDE built by the Windsurf team (including CEO Varun Mohan) that competes with Cursor and Claude Code. It combines a standard VS Code-style editor with what they call an "agent-first" approach - multiple AI agents that can operate autonomously across different workspaces. Here's 4 things that caught my attention: 1️⃣ The trust problem with agentic tools is something they're clearly trying to solve. Instead of logs nobody reads, agents generate "artifacts" - screenshots, video recordings, task lists showing what they actually did. You can verify the work without digging through code. 2️⃣ "Manager mode" where you oversee multiple agents running in parallel instead of just prompting one. The developer becomes an orchestrator, not a typist. 3️⃣ The persistent memory is interesting too. The agent stores patterns and reasoning from past work, so over time it learns how you build things. If it works, that's a real multiplier. 4️⃣ Multiple models - Gemini 3 Pro, Claude Sonnet 4.5, and GPT-OSS. Google knows winning this market isn't about one model, it's about owning the workflow. *** Early feedback is rough though. Crashes, incomplete edits, files getting deleted. And it's free for a reason - every click and snippet is data for Google.

29

12 Comments

Will.i.am just showed me his AI car at CES. Trinity runs NVIDIA's DGX Spark as its brain, with reimagined Android fused with multiple LLMs working together. What I learned: The car connects to an in-home DGX Spark that tracks your routes and context. It's designed to be omnipresent AI; in your house, in your pocket, in your car. Will.i.am designed it himself. Inspired by fighter jets. 200 horsepower, 2,000 pounds. This is what happens when a creative mind meets real AI hardware. #AI #CES2026 #NVIDIA #DGXSpark #William

4

6 Comments

𝐓𝐡𝐢𝐬 𝐖𝐞𝐞𝐤 𝐢𝐧 𝐀𝐈, powered by 𝐒𝐢𝐠𝐧𝐚𝐥. Cutting through the headlines so you can see what truly matters for builders, founders, and AI teams. 𝐍𝐞𝐰𝐬 𝐢𝐧 𝐀𝐈 • 𝐆𝐨𝐨𝐠𝐥𝐞 rolled out 𝐆𝐞𝐦𝐢𝐧𝐢 𝟑, their sharpest model yet, powering agentic tools like Antigravity and 𝐕𝐞𝐨 𝟑.𝟏 for video with better audio sync. Demis Hassabis posted how it recreated his ’90s game Theme Park in hours. Live in the Gemini app and Vertex AI. Founders: plug it into prototypes for quick multimodal tests; context holds 𝟐𝐌 𝐭𝐨𝐤𝐞𝐧𝐬. • 𝐀𝐧𝐭𝐡𝐫𝐨𝐩𝐢𝐜 dropped 𝟓𝟎𝐁 on U.S. data centers in Texas + New York, aiming for 2026 launches and ~800 jobs. Pushes against compute shortages, but grids will feel the strain. They also released new 𝐫𝐞𝐝-𝐭𝐞𝐚𝐦𝐢𝐧𝐠 𝐝𝐚𝐭𝐚 for safety. Faster enterprise deploys if your stack aligns on ethics + compute strategy. • 𝐎𝐩𝐞𝐧𝐀𝐈 signed a $𝟑𝟖𝐁, 𝟕-𝐲𝐞𝐚𝐫 AWS deal. Elon Musk noted xAI’s Grok scans 𝟏𝟎𝟎𝐌 posts/day for "unbiased" feeds. Deals like this unlock scale, but who owns the pipelines? Actionable: lock cloud terms early; latency drops ~30% on optimised infra. • 𝐓𝐞𝐬𝐥𝐚 unveiled 𝐀𝐈𝟓, an inference chip hitting 𝟐𝟓𝟎𝗪 for Optimus bots, specialised for their stack, outpacing rivals. Musk calls it “superhuman” execution. Builders: benchmark against it for robotics; edge compute may flip the hyper-scaler script. • 𝐄𝐔 𝐀𝐈 𝐀𝐜𝐭  kicks enforcement in November with fines up to 6% revenue for high-risk lapses. Dario Amodei dubbed it “friction we need” on X. Audit pipelines and transparency mandates bite data flows, but clarity could boost trust. 𝐅𝐮𝐧𝐝𝐢𝐧𝐠 𝐑𝐨𝐮𝐧𝐝𝐬 Totals crossed $𝟑𝐁 this week. U.S. leads; Europe surges on infra bets. • 𝐂𝐮𝐫𝐬𝐨𝐫 𝐀𝐈 (𝐔𝐒𝐀): 𝟐.𝟑𝐁 Series D (𝟐𝟗𝐁 valuation), Nvidia/GV-led. • 𝐏𝐡𝐚𝐢𝐝𝐫𝐚 (𝐔𝐒𝐀): 𝟓𝟎𝐌+ Series B, data-center efficiency AI. • 𝐖𝐚𝐲𝐯𝐞 (𝐔𝐊): 𝟏.𝟎𝟓𝐁 Series C extension, vision-based autonomy. • 𝐀𝐥𝐞𝐩𝐡 𝐀𝐥𝐩𝐡𝐚 (𝐆𝐞𝐫𝐦𝐚𝐧𝐲): €𝟓𝟎𝟎𝐌, privacy-focused LLMs. • 𝐔𝐧𝐢𝐩𝐡𝐨𝐫𝐞 (𝐔𝐒𝐀): 𝟐𝟔𝟎𝐌 extension, enterprise voice agents. Human oversight is selling; weave it into pitches. 𝐀𝐈 𝐟𝐨𝐫 𝐆𝐨𝐨𝐝 𝐆𝐨𝐨𝐠𝐥𝐞 𝐃𝐞𝐞𝐩𝐌𝐢𝐧𝐝’𝐬 𝐆𝐞𝐦𝐢𝐧𝐢 𝐁𝐢𝐨 compresses genomic analysis and experiment planning, now free for researchers. It flags variants in hours that previously took days. Stanford trials reports ~15% error reduction. This is a big win for biotech teams but dataset diversity still lags. Optimism: equitable access could speed therapies. Caution: dataset diversity lags, biases remain. Builders: integrate for bio pilots carefully. What caught your eye? How is this shifting your stack? 𝗦𝗶𝗴𝗻𝗮𝐥 cuts through the noise so you can focus on what actually moves the field.

10

17 Comments

🗓 *Weekly Recap Of Perplexity :* • Signs *$400M Snap Inc. deal* to embed *AI search in Snap Inc.* and expand *consumer reach*. • Launches *Comet Android beta* to grow *mobile AI browser presence*. • Debuts Perplexity Patents* to enter *AI research markets*. 🚀🚀🚀

7

🚨 ElevenLabs CEO says Voice models becoming commodities & He’s 100% right and this is validation of our model ⭐️🚀 As CEO of ID Privacy AI, this has always been our vision and my thesis, that Voices are just rhe surface layer and not the defensible part of the business. - It’s going be the application and orchestration of agents across multi-systems while having the memory, fallbacks, governance, human in the loop, and more interoperability among the various enterprise functions and tools 🧩 🗣️It won’t be the one who sounds the most human who will win - the speech models will inevitably come to the same level or close to human level standards. So hearing it echoed by ElevenLabs only confirms where the market is going, and our devotion to Infrastructure-first was the right move. 🙌 Models were never the moat.🙅‍♂️ 🔥They’re accelerants, if models get better you should always be parallel to that growth; not disrupted. As Guillermo Flor states correctly the real differentiation moved to the Application Layer: 🧠 Orchestration across real systems (CRM, DMS, ERP) - Meaning full bi-directional integration not just a 1-way into the CRM, passing leads in doesnt mean orchestration. 🔁 Persistent memory across voice, SMS, and email -> critical to agent communication and context / journey. 🛡️ Reliability, monitoring, and enterprise safeguards - Full stop. (Ensuring best security practices are in plade) 🔒 Privacy-by-design infrastructure But here’s the part most people miss 👇 There’s defensibility beyond the application layer. It lives in: ( more secret sauce ) but deeper then app layer: 🧩 How agents reason, act, and recover inside workflows and even mieenmmmm Voice is just the interface. Execution is the product. At ID Privacy AI, we don’t sell voices we build agentic systems that actually run parts of your business.👨‍💼 What’s your thoughts on the ElevenLabs CEO statement - on the Voice is just a surface commodity and it’s the application layer that is the real differentiation? Comment below 👇 Cat Valverde Neetu Shaw Menlo Ventures Vinod Khosla Kevin Roose Bob Corbett Michael Tegtmeyer Brian Pasch Jameson Fleming Tracy Fred Randy Parker Ryan Soffa Mark Fussell Katie Barnes Big Purple Dot Chris Baker Yamini Rangan Riana Cook Claudia Marquez #GenAI #VoiceAI #enteprirseAI #PrivateAI #AImodels #AIStartup #startup #Founder

18

8 Comments