Pim de Witte

there are 2 primary types of cultures at scale: indoctrinated plausible deniability, and brutal, truth seeking honesty. The former often comes across kind, but is low trust and leaves bite marks. The latter appears rude, quiet, or uninteresting, yet is infinitely better long term

@packyM The picture in the article was real, btw ;)

There is a tremendous amount of progress happening in World Models. Multiple labs have raised more than $1B. WMs were the star of GTC. They are a real path to embodied AI. So @PimDeWitte & I wrote a comprehensive 19k word overview of World Models. notboring.co/p/world-models

Also fun fact: the picture in the intro actually happened and was me that night, we were definitely deep world modeling that night :)

Direct link to the piece here: notboring.co/p/world-models

Packy and I spent the past month unpacking world models from first principles. This piece is the result of that exploration. We go into why, what, how, and look out into the future on the implications of our work.

@packyM :)

@PimDeWitte I think we're going to have to write another one next week at the rate things are moving.

Ari and the team are world class and excited to see this grow, we have been happy customers a few times over :)

Imo every AI lab should offer their APIs at cost / for free outside of peak hours for every career related service and educational institution - by far one of the easiest ways to accelerate progress is allowing these to build out systems without $ friction faster (incl students)

@shaneguML Yup

To build a frontier lab, you need the best in product, best in modeling, best in infra, and best in data quality to appreciate, understand, and respect each other's work, despite any firefighting, mistake costing $$$, burnout, and departure. Culture is the hardest.

Ive met Sam, spent a good amount of time with him and dont think this is true. He’s wildly misunderstood, not great at communicating and indeed doesn’t really show emotion well. But I do believe he cares about people. In his head he lives 6-12 months out in the future and fails to translate back to “today-ism” to meet people where they’re really at when talking about AI. Demis does a much better job at this. So does Dario. From my pov this is also the root cause of why a lot of relationship soured over time, because meeting people where they’re are at, and clearly communicating about differences is just really important to relationship building and trust building. With that said, haven’t seen anything to indicate Sam is a bad person, or doesn’t care about people. In fact, strongly believe it’s the opposite. This argument is also taken out of context quite a bit. Their original mission is intelligence too cheap to meter. This is a continuation of that. Which in itself is much better than gating access to tokens for competitive reasons. We should be extremely glad if it ends up being an easy to access utility for everyone to use. And competitive forces will drive the price down and access up over time.

The word for what Sam Altman is is "misanthropic." He hates people, and he sees the arrays of transistors and logic gates he works with as intrinsically more valuable than human life. Its fun to gawk at, but this kind of person is fundamentally and legitimately dangerous.

@francoisfleuret @Ziraax_ @ducha_aiki @Ana_Geneva This is a ridiculously good explanation. Whatever is going on in your latent space today keep it going for as long as you can.

@Ziraax_ @ducha_aiki @Ana_Geneva The problem is that an "idea" is often just a pretty arrangement of words whose meaning can be stretched in any direction. It's why in applied fields execution is [generally] considered the only valid "proof of ownership" for an idea.

You cannot make this thing up

@hoanhle_ Comment is mostly intended for consumer/developer oriented GPUs, systems intended for large scale training and AI workloads are very specialized and different already. To your point. But it will be interesting to see how they will choose to serve inference vs training compute.

@PimDeWitte pic for visual

At a very high level, Nvidia and it’s GPUs are turning deterministic data/compute into tensor ops / mat muls, at increasingly compressed model rates. As such, it’s in a constant race with the GPU’s components itself to be eaten, outside the tensor cores, as it solves for each intermediary stage. You likely won’t need ray tracing cores, etc. World models are the ultimate acceleration of this. This type of consolidation on simple architectures is going to be really great for consumer devices long term, as world models / generative architectures become increasingly general purpose methods for computing.

generalintuition.com/careers, if you prefer the latter also :)

there are 2 primary types of cultures at scale: indoctrinated plausible deniability, and brutal, truth seeking honesty. The former often comes across kind, but is low trust and leaves bite marks. The latter appears rude, quiet, or uninteresting, yet is infinitely better long term

@sainingxie @ylecun @amilabs Right back at you! And indeed, where we're all going, you just need some great collaborators in service of science, progress, and people. Onwards!

@PimDeWitte @ylecun @amilabs thank you so much, Pim. you and general intuition are truly an inspiration to us. hope we can collaborate down the road to tackle some of the hard problems together🚀

i’m joining forces with @ylecun and an incredible group of people to start AMI Labs @amilabs. AMI isn’t a conventional lab. we don’t intend to become one. a lot to say about why this moment matters, but for now we’re heads down building. join us: amilabs.xyz

@invisiblebags You still run into computational limits modeling complex dynamics. Not as obvious as you make it seem. broadly agree that we have a lot of deterministic environments that can be used to create feedback environments for autoresearch style loops. It’s one of many problems though.

The simulators for autoresearch-style loops already exist across dozens of fields: robotics (MuJoCo); autonomous driving (CARLA); drug design (Rosetta), fluid dynamics (OpenFOAM); trading (Backtrader). These were built for labs with massive compute. But now anyone with a single GPU can run narrow experiment slices overnight. We believe there is a billion-dollar opportunity is someone who can plug these simulators into the autoresearch pattern, coordinate fragmented single-GPU contributions across niche verticals, and synthesize the results into real progress. Decentralized research infrastructure is wide open. Research is invisible. If you are building this, talk to us. Vertical specific optimization will win.