
hilarious to watch how Opus 5.5 designs a 3d asset, prints it, and controls a robo arm to take it out of the printer

ml research engineer | visual person, evals, robotics prev google, kaggle, tcl @anyactai

hilarious to watch how Opus 5.5 designs a 3d asset, prints it, and controls a robo arm to take it out of the printer

Also in the repo: each session writes a flight record (plans, telemetry, pictures), so you can inspect it in Rerun. Agents connect via MCP, Python or the wu CLI Next I plan on integrating with NVIDIA GEAR's ASPIRE, so its skill discovery can build on these tools

In its own runtime, world-use tests each plan first on a MuJoCo twin, built from the arm's measured joints and the model's scene estimate. If a step would fail, nothing moves The twin grows as the model explores and adds what it measures. Not used in the LIBERO runs

The tracker is EdgeTAM, Meta's on-device model: 16 FPS on an iPhone, runs locally on most machines. The model marks an object once with a point or a box and it is followed from there

Why fewer calls: - An external tracker locates the object in 3D, so the model doesn't guess - One call runs a whole grasp (centre, descend, close, test-lift), not one per motion - Each call reports position error, contact and grip, so no extra looks are needed

Introducing world-use: open-source tools that let agents operate robot arms It packages what I've learned from running LLMs as robot policies On LIBERO-10, GPT-6.2 Sol low solves 30 of 50 episodes with it and 15 without, on 74% fewer model calls 🧵 github.com/anyact-ai/world-u…

coming to LA this weekend, if you're working on robotics or vision RL let's meet!


Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵


MuJoCo 4 is released! github.com/google-deepmind/m…

imagine training a model not just on a finished book masterpiece, but on the author’s drafts and edits that show how it evolved in its final form and then creating many alternative paths that could have led to entirely different books

like yes first responses are getting better but this revision loop feels relatively similar compared to 1-2 years ago i can imagine its exponentially harder to collect training data for that in high quality

when a person writes some text and revises it many times, it usually gains more and more depth here and there with llms, it feels like each revision improves one part at a significant cost to the rest, keeping the overall depth somewhat the same


I think the last month might have been the biggest month in (zeroth-order) optimization history. In the 10 days since we released the Noise speedrun, the time to pretrain a transformer went from 1.76 hours to 1.6 seconds, a 4,000x (!!!) speedup. Dust itself was up to 10,000x more compute-efficient than the previous state-of-the-art. Not only can we pretrain a transformer without backprop, it's much more efficient than anyone would've expected. Search fking works! Now, scaling is the biggest open question. Some of the tricks that brought the time down to 1.6 seconds will simply not scale. To discard those, we'll be scaling the dataset every time it gets overfit, starting with 30M tokens instead of 10M. Huge thanks to @jayden_teoh_, @spicey_lemonade, @RobertTLange, @EmanueleRodola, @ChunyuanDeng, @LuochengHuman, @Alaxarr, @pavleculafic, @IsaacYong13, Euraba and areu01or00 (GitHub), for the contributions. And @cs_serdar for gracefully maintaining the speedrun and @bishmdl76 for helping out. Expect a lot more from us over the coming weeks.