Zehao Dou

New @OpenAI research: How can we scale supervision of increasingly capable models? Can we rely on monitoring GPT-7's chain-of-thought? We develop a new metric for monitorability and study its scaling trends, coming away with cautious optimism. 🧵:

💥 We're hiring our first research scientists for OpenAI for Science! As a reminder, our goal is to build the next great scientific instrument: an AI-powered platform that accelerates scientific discovery.

Today, we at @OpenAI achieved a milestone that many considered years away: gold medal-level performance on the 2025 IMO with a general reasoning LLM—under the same time limits as humans, without tools. As remarkable as that sounds, it’s even more significant than the headline 🧵

@bneyshabur @AnthropicAI Congratulations, Behnam!

Thrilled to share that I’m joining @AnthropicAI ! After 5.5 amazing years at Alphabet, including working on Gemini’s reasoning over the past 2 years, I’m looking forward to advancing Claude’s ability to tackle complex reasoning challenges across a diverse range of domains!

🌞My last work at school🌞 Thanks to my excellent collaborators! Our paper focuses on inverse problem solving with diffusion prior and MCMC. Perhaps sometimes a few extra exploration steps can greatly improving your inverse problem solving. 😀 arxiv.org/abs/2409.08551

LLMs have been increasingly used for (sequential) decision-making (as "autonomous agents"), and (to me) quite interestingly, more and more used to "simulate" the "human/social behaviors" with interactions among each other. See the mind-blowing example by @joon_s_pk on 1/n

How to capitalize #GenerativeAI and #diffusion models for modeling complex data and structured optimization? From images to proteins? Check my talk "Diffusion models for Generative Optimization" at @broadinstitute , Harvard, MIT last week. Youtube: youtube.com/watch?v=hDRDx5…

**Training dynamics of attention** 1/📜Introducing our latest paper: "Training Dynamics of Multi-Head Softmax Attention for In-Context Learning: Emergence, Convergence, and Optimality." Link: [arxiv.org/abs/2402.19442] Joint work with @siyuc3141, @HeejuneSheen, and @0920wth

Proving optimization guarantees for transformers is hard, even if just training on seq2seq pairs for which we know some small transformer achieves zero test loss. In practice gradient descent just works. In theory, it's open to prove *any* efficient algorithm succeeds 🥲 1/

Introducing Sora, our text-to-video model. Sora can create videos of up to 60 seconds featuring highly detailed scenes, complex camera motion, and multiple characters with vibrant emotions. openai.com/sora Prompt: “Beautiful, snowy Tokyo city is bustling. The camera moves through the bustling city street, following several people enjoying the beautiful snowy weather and shopping at nearby stalls. Gorgeous sakura petals are flying through the wind along with snowflakes.”

@catherineliangq @ZimingLiu11 @FieteGroup Great work.

Do diffusion models learn semantically meaningful and efficient representation? In our latest work, we explored this question by training a toy model on synthetic datasets and found out that simple diffusion models do not learn factorized representations of independent concepts.

@sunjiao123sun_ @MaxMa1987 @VioletNPeng @jonathanmay @emilio__ferrara Congratulations!!

Today I defended my thesis and became Dr. Sun! 🌞 Thank you my committee members @MaxMa1987 @VioletNPeng @jonathanmay @emilio__ferrara and Dan O’Leary! The slides of my presentation are here: docs.google.com/presentation/d…. Ph.D done but research never ends! Fight on!

@iclr_conf +U+U !

We are nearing the finish line. +U!

My group is hiring a large cohort of interns for the summer of 2024 to work on the Foundations of Large Language Models! Come help us uncover the new physics of A.I. to improve the LLM building practices! (Pic below from our NeurIPS 2023 paper w. interns) jobs.careers.microsoft.com/global/en/job/…

I’m very excited to share our work on Gemini today! Gemini is a family of multimodal models that demonstrate really strong capabilities across the image, audio, video, and text domains. Our most-capable model, Gemini Ultra, advances the state of the art in 30 of 32 benchmarks, including 10 of 12 popular text and reasoning benchmarks, 9 of 9 image understanding benchmarks, 6 of 6 video understanding benchmarks, and 5 of 5 speech recognition and speech translation benchmarks. Gemini Ultra is the first model to achieve human-expert performance on MMLU across 57 subjects with a score above 90%. It also achieves a new state-of-the-art score of 62.4% on the new MMMU multimodal reasoning benchmark, outperforming the previous best model by more than 5 percentage points. Gemini was built by an awesome team of people from @GoogleDeepMind, @GoogleResearch, and elsewhere at @Google, and is one of the largest science and engineering efforts we’ve ever undertaken. As one of the two overall technical leads of the Gemini effort, along with my colleague @OriolVinyalsML, I am incredibly proud of the whole team, and we’re so excited to be sharing our work with you today! There’s quite a lot of different material about Gemini available, starting with: Main blog post: blog.google/technology/ai/… 60-page technical report authored by th Gemini Team: deepmind.google/gemini/gemini_… In this thread, I’ll walk you through some of the highlights.

Transformers are the building blocks of modern LLMs. Can we reliably understand how they work? In our #NeurIPS2023 paper arxiv.org/abs/2312.01429 we show that interpretability claims based on isolated attention patterns or weight components can be (provably) misleading.

Check it out guys. Our paper "Rates of estimation for high-dimensional multi-reference alignment" has been finally accepted by Annuals of Statistics. Thanks so much for all the help and guidance from my advisors Harry and Zhou. arxiv.org/pdf/2205.01847…