Luzia

"That link is not there yet." That's Uber's COO, Andrew Macdonald, on whether burning more AI tokens is actually producing more useful features for riders. His CTO had already blown through Uber's 2026 Claude Code budget by April, and when leadership went looking for what all that spend bought, nobody could draw the line from tokens going in to features coming out. I wrote two pieces pointing straight at this moment (Stop Bragging About Tokens: https://lnkd.in/eizfmvc8 and Tokens are not kWs: https://lnkd.in/eqRq3dv3), so the easy move is the I-told-you-so. I'll skip it, because I'm in the same hole, just a bit shallower. At Luzia, token consumption across engineering is way up this quarter, full AI-coding adoption, almost nobody writing code by hand anymore. The outcomes are not moving together. One engineer's output has climbed for weeks while another's is flat on the same token growth, and I still can't hand you a clean "tokens per shipped feature" for my own team. Macdonald's head-exploding moment is my Tuesday. This is a pain every company is going to walk through, because there's no version of efficient token usage that skips the wasteful part. The engineers who get 10x out of these tools got there by burning tokens on experiments that went nowhere, and the few that landed paid for all the rest. You don't learn which is which by being careful, you learn it by spending. Which is why the timing matters. Uber is having this reckoning at subsidized prices. Almost every token bill today is paid partly by a lab eating the gap to lock you in, and when that subsidy ends the same unanswered question comes back with a figure three to five times bigger attached. Now is the time to waste tokens, while they're cheap and someone else is covering half, not when the meter starts charging the real number. So the scary number in the Uber story isn't the blown budget. Something like 90% of companies have no real AI adoption at all, they're not overspending on tokens, they're not spending on them, period. Uber asking "tokens per feature" is already ahead of the field. The ones who get hurt are the companies doing their wasteful learning after the discount expires, all at once, with finance in the room. I'd rather have my head-exploding moment now, at half price, than in 2027 at the unsubsidized one.

Tokens without a denominator are just an invoice. Álvaro on what to divide by. ⤵️ https://lnkd.in/eZJFywaS

Álvaro wrote the longer version of what we showed last week with the Org Brain. The question underneath it is simple and enormous: if intelligence is no longer the scarce resource inside a company, what is? His answer: context. We agree. We're building around it. https://lnkd.in/eeqNc6dt

Every company past 30 people has the same disease. Knowledge scatters. Wikis that were accurate four months ago. Slack threads with decisions buried in 40 messages. Meeting notes in someone's app that nobody else can search. A Google Doc three people bookmarked and everyone else forgot exists. You can find things if you know exactly where to look. Most of the time, you don't. We built something internally that we've been calling "The Org Brain". An agent that reads our wikis, checks our repositories, knows when a document was last updated, and tells you when something is stale. You ask it a question, it goes and looks. Not keyword search. It reasons about what it finds, cross-references sources, and flags what's missing or outdated. The models are smart enough. That stopped being the bottleneck a while ago. The bottleneck is context: the right document, the recent decision, the constraint nobody wrote down. Get the context right and the intelligence follows. Get it wrong and you have a very articulate system that confidently tells you things that were true in January. We're building this for ourselves first. In the demo video bellow, wrapped in a fancy interface with the latest OpenAI voice model. But ultimately, the pattern is the same at every scale and interface type. The same architecture that helps our engineering team ask "what's the real status of this feature?" is what will let a Luzia user ask "what was that restaurant my friend recommended last month?". Same problem. Same solution. +85 million users' worth of difference in scale. What does yours look like? We're curious how other teams are solving this. 👀

We've been moving the goalposts on AGI for 70 years. First it was chess. Then the Turing test. Now it's consciousness, the soul, things nobody can define precisely but that conveniently keep the finish line just out of reach. Take what's on the table today and move it four years back. Most of the people who set the original thresholds would say we're already there. I said this to Fede Durán for El Mundo this week, in a piece on where AI actually stands. My honest read: the AGI debate is mostly a conversation about definitions, not capability. The definitions shift every time the machines catch up. What I'm less uncertain about: AI is already capable enough to change the lives of hundreds of millions of people. That doesn't require AGI. It requires building the right product for the right person. https://lnkd.in/eT4p5pVw

Built by our team. Used every day inside Luzia. https://lnkd.in/edDuWSrX

We're #hiring a new Product Engineer (Android & Fullstack) in Madrid, Community of Madrid. Apply today or share this post with your network.

We've been working on something... 👀 Everyone has a chat with themselves on WhatsApp. Screenshots of recipes. Links to articles you'll read later. Voice memos with ideas you had at 2am. A restaurant someone mentioned at dinner. The name of that book. It all goes in. Nothing comes back out. Because a chat is a container, not a brain. There's no way to search, no way to ask "what was that Thai place someone recommended last month?" We built Spilo to fix that. Send anything to Spilo on WhatsApp, the same way you'd send it to yourself. AI classifies it, organizes it into lists, extracts the useful details. A recipe becomes a recipe card with ingredients. A Maps link becomes a place card with an address and rating. A voice memo gets transcribed and categorized. Then ask it anything. "What restaurants did I save near Sol?" "When does my passport expire?" It searches across everything you've stored and answers with the actual cards. It's in beta, it's free, and we'd love for you to try it. https://lnkd.in/e2P2MCPy

85 Meta employees. 60 trillion tokens. 30 days. A leaderboard called "Claudeonomics." Two days after the story broke, the dashboard went down. Someone took it offline "at their discretion." That disappearance is the story. A raw token leaderboard measures who burns the most, not who ships the most value. The top spot goes to the engineer who never thought twice before running a model on a whim. The bottom goes to the one who thinks too hard about cost. Alvaro M. argues every company using AI is sitting in a version of this. The bill is climbing, engineering is the early adopter, and the first measurement makes things worse because the measurement has no denominator. Tokens per PR merged. Tokens per feature shipped. Tokens per paying user. Pick a metric the team already lives by and divide. Don't cap usage. Don't run it open. Instrument it. Six to twelve weeks of real use, logged against the right denominators, and by the end you know which teams convert tokens into outcomes and which convert them into nothing. Cloud costs followed the same arc in 2015. It took five years. Tokens will settle in eighteen months. The companies that build the muscle now will be able to answer the board question when it comes. Full piece in the comments ⬇️

Wednesday 29th - No script, no agenda, just a no BS chat about AI. This one you don't want to miss. 👀