
I’m hiring for my robotics inference startup, we’re working on solving fast cheap long context batch invariant Inference for VLAs and WMs, billions will thrive!

billions will thrive robot inference @hf0, ex-quant 2x

I’m hiring for my robotics inference startup, we’re working on solving fast cheap long context batch invariant Inference for VLAs and WMs, billions will thrive!

who's building a wetlab and wants robots?

While the inference optimization space has indexed on Megakernels, what we are finding that the more optimal strategy is creating a larger number of multiple distinct operation fusions. Fusing a subset of operations produces smaller kernels that run near hardware limits/SOL. Of course, the number of useful subsets grows combinatorially with model shape and hardware. Humans can’t write or search that space by hand, but agents can. This is what we are now doing with Together Compiler, and we are getting quite convinced this is a broader view of how systems software will be designed & "grown". A great thread by @dylan__lim.
Last Year We Built Megakernels. This Year We Deleted Them. One year ago, we helped introduce some of the first AI megakernels to the world. Since then, we’ve enjoyed seeing their wide adoption from friends and researchers. However, while whole model megakernels are a helpful abstraction, we’ve found that they overcomplicate and are not always an ideal optimization target. Below, we discuss our journey to understand and push past these limitations. We start with a deep dive into two contrasting paradigms of ML optimization – Torch Compile and Megakernels

Introducing DrivingBench v2 🧵 We wired a Tesla Model 3 so any AI model can drive it, and put 8 frontier VLMs (Claude, GPT, Grok, Gemini models) and 6 fine-tuned robot models (VLAs & WAMs) behind the wheel. A new lot and two new courses, none of which they’d seen. Only one VLM ever reached the finish, and its fastest run took 26 minutes. 🏁

Researchers have used reinforcement learning to teach a 5-tonne John Deere skid-steer loader to perform complex acrobatic manoeuvres autonomously. Trained in simulation and transferred to a real machine, the system can balance on two wheels, navigate rough terrain, spin, and control its bucket before landing safely. The research demonstrates the potential of AI-driven automation for heavy construction and industrial machinery. Video Credit: @leggedrobotics / @ETH Paper: wevlv.co/4kltYys #engineering #technology

DM to try it

we swept schedulers, rtc and seam blending on top: the model call is 1.66x faster and one L40S holds 8 robots with zero starved control ticks


if you need compute DM me

during deployment we kept running into a failure mode where a dropped frame led to model degradation, and when controlling fleets uplink is the bottleneck. So we made our own decoder, which survives symmetric NAT, double NAT, and port-remapping. It was fun to work on!

new transport launch (long time due) three 720p robot cameras cost 46.6 mbps to send to a model as jpeg. as h264 they cost 1.43 here's how it works:

Broadly speaking, the AI safety movement has a huge public champion problem: virtually everyone who speaks up for AI safety has also chosen to be weird in highly distracting ways that break their mainstream appeal. Except for one obvious, photogenic champion who it turns out is exceptionally good under pressure in unprecedentedly contested political races.
People are acting like Jacob Coxon is the selected champion of the safety movement, his tweet was just exactly right-place-right time. It was more about the content than the person.

Does anyone have a space they could let FOUR robots be filmed? Rockets, engines, logistics cool products? Bay area

women in social media marketing: "my team of girly pops and i run our company's social media accounts :)" men in social media marketing: "i'm a member of Growth Team 6, a narrative dominance unit specializing in cyber capabilities and psychological operations. we operate in the trenches of the attention war"


ran an eval in t3, for our models we can now support 8x more clients per GPU, and so inference is now 8x cheaper!! no degradations with our real evals on YAM