Lucas

Robot action models shouldn't need 256 vision tokens per frame. Pi0.5 spends 400M parameters on SigLIP just to see. We replaced it with a 4.4M encoder that outputs 5 tokens — and action quality barely changes. 91x smaller. 51x fewer tokens. 7.3x faster inference.

Montreal I am in you

there was a post a while ago about how to make a fundamentally better embedding model who did that

@_advaitpatel try to make a jlens on the embedding space on a dataset could do cool stuff

so is the j space paper worth reading that shit long as f

unified action space seems like a great idea actually

This is what all those be teleop data vendors should be doing

if anyone has a lot of dinov3 embedding to do, I can run the 7B model at 2 OOM cheaper

“But can you attract top tier talent?” The question should really be how good are you at identifying talented people who are less recognizable and arbitrage the labor market.

kpop girl groups at icml is just two different religions of optimization finally recognizing each other across the divide

for forty years, you operated the machine. in this demo, it's reversed. i talk, my mac works. acts across my apps, remembers my whole day. fully on your machine, or in the cloud. your choice.

i think they're trying to tell something to SWEs

Game engine devs are a great source of talent for robotics cos

Used dinov3-7b to embed a big video dataset, now training models on the embeddings for interp. Jlens, sae and trying to replicate some goodfire work to tweak the data and train my robots better

Introducing Sparse Delta Memory (SDM) - The first work of my PhD 🎓. SDM combines the GatedDeltaNet update with Product Key sparsity, enabling a recurrent state 3000x larger at the same FLOPs and significantly improving long-context performance. Let RNNs be sparse!

This is the next kind of data center

If anyone else is homeless and grinding, all will be welcome in the new office

Oh you’re trying to sell data to a big lab? You have a data pack right? With an eval right? With pre and post train scaling plots, existing model perf, number of samples, failure case analysis, easy medium hard subsets, ground truth audits, normal and expert human baselines right? You have a target org researcher contact right? Oh you don’t know anybody at the labs and are going to cold email the procurement team? Oh… You have an off the shelf set available immediately and a pipeline with max weekly throughout numbers with scalable QC right? Oh you don’t have adversarial defenses for your outsourced QA team? Oh you don’t even have a ramp target with a dedicated queue manager? Oh…

new hire onboarding: - laptop - email - slack - X account distribution is now part of the job description

“The dirty secret in AI is that everything is a data and an eval problem. The best models have the best data and best internal benchmarks. The mid ones buy a lot of data, not the best, and hillclimb public benchmarks. (you need a lot of compute too)” – Stanford CS Professor

@bill_xby We should collab

anyone needs a dj in mtl hmu

So we can see any models j-space?