craig

@diodeinc | ex-quant at sig

[Sam Altman @sama]

there are a lot of benchmarks that suggest 5.6 sol is the best model in the world right now, but the most reliable way to tell is that elon is obsessed with me again

[craig @craigxchen]

https://t.co/eexL9GI9Kk

[craig @craigxchen]

Wrote a math paper with the help of LLMs, anybody in number theory willing to sponsor my uploading this to arXiv? This was a problem shared with me in my undergrad that always kinda nagged me in the back of my mind, so it’s nice to finally make some progress on it.

Paper below!

[craig @craigxchen]

@yacineMTB This was all I learned from working as a quant tbh just look at it!

[craig @craigxchen]

@om_patel5 I've observed this behavior as well on very long horizon (like 10+ hour) tasks, it seems to grumble and overthink more than opus?

[craig @craigxchen]

@gclue_akira Did you route the board or did opus4.8 also do that?

[craig @craigxchen]

@yacineMTB @YuanmingH There’s decent sim literature from the plasma physics guys too that I’ve started looking at for improving my sim performance

[craig @craigxchen]

@fabianstelzer fable did this too

[Genesis AI @gs_ai_]

Humanoid robots don't need to look human.

Meet Eno, our first general-purpose robot.

Not a machine pretending to be human, but intelligence given a body.

At Genesis, we’re building a future where robots don’t feel cold or distant, but capable, calm, and ready to help.

Available Q4 this year.

[craig @craigxchen]

@yacineMTB solves 6 pendulums and now my entire feed is pufferlib

[Eddie Lee @eddietree]

built a tiny RL rock-stacking experiment

trained on 200k sims via PufferLib (@jsuarez)

observes rock geometry, mass, friction, velocity, stack contour, support intervals, roll torque, balance, etc

[craig @craigxchen]

Can fable stop flagging my RL stuff none of my work is remotely frontier rn

[craig @craigxchen]

@yacineMTB ya this is killing me even for fluid+solid porous interactions

[LeRobot @LeRobotHF]

VLA-JEPA just dropped in LeRobot 🤖

What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics.

During training, the VLA leverages V-JEPA2 by conditioning its predictor. This clever trick adds a world modeling objective to the training, which also allows pretraining on human videos.
At inference, the world model is dropped entirely, keeping only a standard VLA architecture: Qwen backbone and action head.

The demo here was only fine-tuned on 13 examples, showing great pretraining capability and running in real time on @NVIDIARobotics DGX Spark!

VLA-JEPA is the first world model to be ported to LeRobot, and I feel like it won't be the last 🚀

@Thom_Wolf @ClementDelangue

[craig @craigxchen]

I thought my coffee sim was too slow because of algo issues but turns out it was memory locality 🤦

[craig @craigxchen]

@yacineMTB agree though I think we're pretty gated by hardware, so people just threw easy things that worked in other domains at the wall and the VLA happened to stick. once we get cheaper/better tactile + throw compute at jepa then we're cooking

[craig @craigxchen]

Can anyone put me in touch with people working on diamond wafers / using plasma for n-doping?

[craig @craigxchen]

@isosteph I also hear the cheering what is this about

[atharva ☆ @k7agar]

wake up excited to work on robots

robot decided it does not want to work

fix robot, takes whole day

sleep with dream of making the robot work tomorrow

repeat

[craig @craigxchen]

@yacineMTB You should take a look at MPM too, I’m using it for simulating coffee brewing but I might switch to smtg else since I’m running into performance issues

[craig @craigxchen]

the new codex "attach claude" feature is amazing