Building infrastructure to support large scale physical AI post-training and deployment.
San Francisco, CA
Joined March 2026
- Tweets12
- Following2
- Followers672
- Likes1
Pinned Tweet
Introducing InstinctFlash, the unified open-source physical AI model runtime for fast, efficient inference.
End-to-end robot policies don't fit on the edge. VLAs are already extremely slow to run on-device, but WAMs are even slower, taking 20-50 steps to even de-noise a single clip for an action.
Flash allows you to run generalist models from 8 model families on a single commercial GPU, in real time.
What about dis-assembling?
Excited to share some results (more coming next week!) deploying pi0.7 on a AI hardware dis-assembling task partnering with @physical_int. The robot autonomously removes retaining bars, riser boards, and fans with only 1.5 hours of post-train.
Here is Astra on the same task. The tele-operated ground truth trajectory, video, and text instructions given as prompts into the model's context. Astra took about 2 hours on the task but was not able to successfully complete it.
Stay tuned for our upcoming blog about details on the benchmarks and results from Cosmos 3 and pi0.5:
general-instinct.com/blog
We also tried running the same task on GPT-6 Astra. Astra took a very detailed (~300 words) prompt and the tele-operated ground-truth trajectory + video but wasn't able to one-shot the task. The attempt was 5.75 times slower than GI0 optimized on Flash.
We believe an essential step to deploying real physical AI is to be able to run intelligent models on the edge in real time. InstinctFlash is our first attempt at solving that.
Read more about it and access in the blog:
general-instinct.com/blog/in…
General Instinct retweeted
I made a video explaining why on-policy distillation uses reverse KL divergence youtu.be/IiYT1PECiHU?si=nEle…
