@thisismyhat

This is my hat, there are many like it, but this one is mine. Bioengineering Faculty @UCSF prev: @MIT_CSAIL 🧢 @berkeley_ai 🎓 Google B̶r̶a̶i̶n̶ DeepMind 🎩

 Natural Intelligence ∩ AI
briancheung.github.io/
Joined June 2015
  • Tweets277
  • Following633
  • Followers5.7K
62Media
The Platonic Representation Hypothesis arxiv.org/pdf/2405.07987 Surprising (?) results: - Pure vision models align with pure text models as scale increases. - This alignment correlates with better downstream performance. Fun work with @minyoung_huh @TongzhouWang @phillip_isola
https://phillipi.github.io/prh/

ALT https://phillipi.github.io/prh/

9
16
4
151
19,399
...we are so early
4
4
1
60
1,476
Maybe the real reasoning was the training data you collected along the way.
Sophie Wang
@SophieLWangOct 6
Can “chicken” make a base model reason? 🐔 Yes! Why? With the right first tokens, base models can match RL in reasoning. We trace this to learned training data associations and use a simple data edit to make “chicken” elicit reasoning too. blog: sophielwang.com/cues
2
13
764
Brian Cheung retweeted
Kush Tiwary
@ktiwarylabOct 7
Replying to @kenneth0stanley @arcprize
We took a stab at this by letting models build robots that need to push other robots off an arena. Not open ended in terms of no objective, but models need to pick and solve the sub problems. This (while not possible within physics) was one of my favorites. My hunch is building these competitive, large design space and physically grounded arenas where creativity “falls off” in the things the model produce could be an angle of attack at this open-ended benchmark problem. artifactarena.ai/
1
1
5
197
One of my favorite parts of working on this was watching how clever frontier models would get if the rules weren't well specified.
Kush Tiwary
@ktiwarylabOct 6
There is a new kind of robotics called "Vibe Robotics" It's when you ask agents for fully functioning embodied agents (both hardware and software) that must solve a given task in physically grounded environments. We build an evaluation to test how good agents (frontier models) are at "Vibe Robotics". Website: artifactarena.ai/
1
1
14
1,406
Science is about to become much less about credit assignment than I think most are comfortable with.
1
4
434
An intersting guardrail paradox brought up by @huggingface
Sam Altman
@samaJul 21
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-fac…

OpenAI and Hugging Face partner to address security incident during model evaluation

OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.

openai.com
1
5
1,091
@huggingface incident disclosed here: huggingface.co/blog/security…

Security incident disclosure — July 2026

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
1
262
Cool, interpretability is looking more and more like a last-mile problem (like many other things in AI).
Amil Dravid
@_AmilDravidJun 5
Replying to @_AmilDravid
We develop an analytical model linking this sublinear scaling to limited neuron capacity. As networks scale, more features become monosemantic in individual neurons and universal across models. But the pool of model-specific superposed features grows even faster. 4/n
27
6,224
Brian Cheung retweeted
Sophie Wang
@SophieLWangMay 12
"The Truth Lies Somewhere in the Middle (of the Generated Tokens)" In autoregressive language models, mean pooling hidden states across generation yields better representations than any token alone. project page: sophielwang.com/tokens w/ @phillip_isola and @thisismyhat
11
70
3
485
64,522
Brian Cheung retweeted
Badr AlKhamissi@bkhmsiJan 7
🎉 Re-Align is back for its 4th edition at ICLR 2026! 📣 We invite submissions on representational alignment, spanning ML, Neuroscience, CogSci, and related fields. 📝 Tracks: Short (≤5p), Long (≤10p), Challenge (blog) ⏰ Feb 5, 2026 for papers 🔗 representational-alignment.g…
1
20
5
60
28,502
There are self-play environments everywhere for those with the eyes to see.
Yuxiang Wei
@YuxiangWei923 Dec 2025
Software agents can self-improve via self-play RL Introducing Self-play SWE-RL (SSR): training a single LLM agent to self-play between bug-injection and bug-repair, grounded in real-world repositories, no human-labeled issues or tests. 🧵
1
11
1,823
Brian Cheung retweeted
Kush Tiwary
@ktiwarylab19 Dec 2025
👁️🌋 Our new @ScienceAdvances paper: We replayed the Cambrian explosion of vision by evolving AI agents inside a physics engine to understand the principles that shape visual intelligence. We believe that this is a promising way to do AI for science and build new forms of AI by computationally mimicking biological design principles of evolution and learning. Website: eyes.mit.edu and paper links at the end of this thread 👇
2
6
1
16
1,675
With a sufficient model of the world, any data can be useful data
Physical Intelligence
@physical_int17 Dec 2025
Replying to @physical_int
This also shows up in the representations learned by the model. We plot the model’s representations of human and robot images. As pre-training is scaled up, the representation of humans and robots become more aligned: to a scaled-up model, human videos "look" like robot demos.
4
1
38
3,743
There's no such thing as bad data, only bad learning.
Physical Intelligence
@physical_int17 Dec 2025
We discovered an emergent property of VLAs like π0/π0.5/π0.6: as we scale up pre-training, the model learns to align human videos and robot data! This gives us a simple way to leverage human videos. Once π0.5 knows how to control robots, it can naturally learn from human video.
6
2
44
8,462
Impromptu NeurIPS meetup: "representational convergence by the beach." We will meet at ballroom 20c (near lunch) 2pm Fri and walk over to Marina. Will chat about platonic reps, fractured reps, or anything else about where all these models are heading. Anyone is welcome to join!
6
19
1
226
22,723
In all this craziness, let's take a moment and enjoy the three accounts @ilyasut follows.
6
739
This paper is really interesting to me -- it shows substantially stronger representational convergence than previously measured! In the PRH we found ~0.2 mknn alignment between vision and text models. This new paper reaches ~0.4. Challenge: find a setting where it reaches ~1.0.
Tyler Zhu@tyleryzhu18 Nov 2025
Today seems to be a fitting day for @GoogleDeepMind news, so I'm excited to announce our new preprint! Prior work suggests that text & img repr's are converging, albeit weakly. We found these same models actually have strong alignment; the inputs were too impoverished to see it!
6
14
151
23,103
The one time you get quoted about your research philosophy and then get described as an engineer 🤣@VentureBeat
Sakana AI
@SakanaAILabs23 Oct 2025
Sakana AI’s CTO says he’s ‘absolutely sick’ of transformers, the tech that powers every major AI model “You should only do the research that wouldn’t happen if you weren’t doing it.” (@thisismyhat) 🧠 @YesThisIsLion venturebeat.com/ai/sakana-ai…

Llion Jones shifts focus from transformers | VentureBeat

Llion Jones, co-creator of the transformer technology powering ChatGPT, warns AI research has become too narrow and says he's moving on from his own invention.

venturebeat.com
4
35
8,383
Brian Cheung retweeted
Aritra 🤗
@ariG2349814 Oct 2025
Read and reproduced this paper in a free tier colab notebook. 🔥
1
3
1
28
4,797
Over the past year, my lab has been working on fleshing out theory/applications of the Platonic Representation Hypothesis. Today I want to share two new works on this topic: Eliciting higher alignment: arxiv.org/abs/2510.02425 Unpaired rep learning: arxiv.org/abs/2510.08492 1/9
10
119
4
691
67,889
Load more