For the record: attention, scale, and a sufficiently hard problem is all you need.

get a job that lets you do both

we need Chinchilla-style fixed compute budget trade-off curves across. - model size. - pre-training compute. - RL compute . I suspect they'd show Grok4 spent too much time doing RL, not enough time pre-training. No idea what to expect on model size.

really wish @weights_biases had actual job state callbacks instead of forcing me to write polling infrastructure.

here's another (raw) video of us fooling around with the policy for the first time. it was only trained on a few flag waiving episodes and it generalized the hand-off to all kinds of objects including sparklers (and chip bag, which I didn't record 😭)

Happy Independence Day 🦅🌎🇺🇸. here's another shot of our entirely autonomous interactive policy. despite knowing DYNA-1 can do robust long horizon manipulation, the flag hand-off was a wow moment for me bc it felt like a pet dog (hence the audio 🐶), not alien intelligence

the next few months are going to make abundantly clear the robustness, flexibility, and generalizability of our entire tech stack. everyone squabbling about eval metrics forgets there's one metric to rule them all: ARR 😎.

- new demo task. - new environment with hardware we just setup. - robust to changes in lighting, people visible, etc. - production ready quality and robustness. - real businesses that are interested bc it provides real value. it's a cook or get cooked world and we're cooking.

ready for another beautiful day of semi-automated ML engineering. it's crazy how rapidly things have changed

AV is all about the long tail. there's an ocean of difference between "it almost always works" and "it always works", and that gap is the difference between a fancy L2 and an actual L4 system. and that's why you have e.g. geofenced HD maps. don't be fooled by marketing.

genuinely shocking how bad outsider analysis is of the AV space . it's overwhelmingly. - ideologically possessed. - omitting important caveats. - ignoring the actual open problems.

numerology but for training performance based on SLURM job IDs.

the internet is so full of robot teleop videos or 2 second clips of an "autonomous policy" (that only works on that one setup) that it's hard to put out a video of an *actually* working robot to impress normies. to them a 5 minute video of a robot actually working is just boring.

it's not over for senior+ engineers yet, but wait a year.

I could have shaved *at least* a year off my PhD if I had access to current tools. well designed, clean codebases translate to highly effective agents. in turn, if you guide the design well, you can hit insane research velocity and do far more ambitious experiments with ease.

the capabilities growth of coding agents is crazy. I have stopped coding -- I do design and architecture, the agents do the implementation, and then I read it. it's actually over for junior engineers -- you need the design skills of a senior eng on day 1, and IDK how you get that.

"use a zero vector for noise" doesn't work because, for reasonably high dimensional noise, the model just freaks out because it has never seen a data point from anywhere near the center -- random samples all have extremely large expected radii!.

it would be nice to see more work on controlling diversity in denoising models. in LLMs we can easily move between MAP estimate (temp=0) and highly diverse (temp>>0). currently there's no standard analogous approach in denoising.

this is also why I MIT Licensed everything in SceneFlowZoo -- I want people to actually be able to use it.

it's very annoying that so many research ML model weights and training codebases are non-commercially licensed -- it makes certain lines of work toxic for any company affiliated researchers. if you want your research to have maximum impact, MIT License everything.

this is even more catastrophic in modern LLM online RL algorithms like GRPO (the RL algo behind DeepSeek R1). they forgo explicit state Advantage modeling in favor of average reward over a batch of rollouts -- every state visited will have a very negative Advantage estimate!