Roei Herzig
2,399
posts
@roeiherzig
Berkeley, CA
Joined March 2017
Pinned
If you want a vision encoder for dexterous manipulation, what should be the most important part to model? ๐ค Current standard models like CLIP, SigLIP, and DINOv2 have an incredible grasp of semantics and spatial details. But they lack the action-centric structure needed for
00:00
I think in-context learning (which is really part of adaptation) is one of the most interesting open research questions in embodied AI. And I believe it may hold the key to physical AI's GPT moment. I've been chasing this question for a while โ back in 2024 at Berkeley, we
Introducing GEN-1.5, a one-shot learner. It can learn new tasks in a few seconds. Show it what to do, and it generalizes. This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.
00:00
The moment Meta fired
@ylecun, its entire research org became irrelevant overnight... Hard to believe this is the same lab that gave us DINO, SAM, and LLaMA. Good times. Let's see if that changes in the future.
Meta fires Yann, Muse becomes good xAI fires their team, Grok becomes good Google fires Demis, Gemini becomes ???
This tweet takes me back to earlier in my career, when I tried reading a few of Noga Alonโs papers. I understood a little, but one thing I knew for sure: they werenโt AI slop :) ๐
I saw someone post a dozen+ AI slop papers purporting to solve niche open problems. This is antisocial behaviour and worse than if the problems stayed open. 0 people understand the solution, and there is no readable writeup. Also, incentive for either is removed.
The hardest problem that even AGI cannot solve ๐
ืื ืื AGI ืขื ืฉืื ืื ืคืืชืจ ืืช ืืืขืื ืืืืืื ืืืืชืจ ืฉื ืืืืืืงืก
