Andrej, we don't know each other, but I think I'm speaking for a lot of people when I say I hope you'll post more often. Especially now that the public's view of AI is so negative.
Your ideas in past 2 years like "tabbing", "vibe coding" closed the gap between, like you said, the AI dinosaurs (experts) and the general public (randos like me).
We're moving very quickly and I think we need brilliant teachers to distill the benefits of AI to the public. 🙇🙏🙏🙏
@deedydas Thanks for sharing an internal cultural view. It aligns with what some people who left this year are saying. Google has human values in the right places, and in the long run, that’s what’s really hard to beat!
Congratulations to the @NASA workforce and our partner @SpaceX on another successful crewed launch to the International Space Station.
Crew-13 is now aboard the station, Crew-12 will be home soon, and 2026 is far from over. Still ahead: more science missions, more flyovers, the MAX POWER exposition, cargo resupply, a possible Starliner-1 demonstration mission to the ISS, Artemis III rollout and tanking tests, and the first Moon Base launch before the year ends.
We will never stop going. 🇺🇸
There are people who have not read many books and still show high raw intelligence. They know how to take an open-ended problem apart: name the parts, see which constraints actually bind, and build a path that was not handed to them.
By contrast, many people who have digested far more text still cannot break down a problem that sits outside the distribution of what they have read. They can retrieve a nearby case, quote the familiar move, and stall when neither applies.
Raw intelligence, in that sense, is not a stock of tokens. It is the habit of internalizing structure — the relations, invariants, and causal joints underneath examples — so a new problem can be generated from the inside rather than matched from the outside. Reading helps only when it leaves that structure behind. Otherwise it trains fluency at recall, not the ability to decompose what has not been seen.
@shazcodes This is an important observation. We don’t need a crane just to put a cup in the sink. Startups may be willing to spend fast first to get MVP then optimize, but it does seem like even corporate workflows are run like this.
A startup founder told me yesterday they raised $2M to build an autonomous AI agent for enterprise data retrieval.
I looked at their tech stack. it's literally just a vector database, 3 openai api calls in a loop, and a postgres table.
Their monthly api bill is $14000.
A single indexed sql query would do the exact same job for $20/month on a basic vps.
We are living in the biggest compute burning bubble in human history.
We’re at a point in late stage AI model capitalism where you know the frontier has to win in a bunch of benchmarks to launch. These numbers mean nothing.
The thing to trust is the price.
If they price it high, it’s a good model. If they don’t, it’s benchmaxxed.
@kepano Haha, I had this realization too when helping someone learn how to use ChatGPT. Prompt for system outcomes, not just task outcomes. Then on the OS will gradually emerge a beautiful agentic system
@jgebbia you guys are killing it!! https://t.co/a01EPqCTK0 looks and feels amazing.
holy shit unified government portal seemed impossible just a few years ago