🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Demis Hassabis, James Manyika, and I wrote up a (lengthy and illustrated!) overview of the AI research work and advances across Google in 2024. It's a summary of the work of many across Google, covering Gemini advances, Gemma, NotebookLM, generative image and video models like

I don't understand why people say RLHF is a contextual bandit problem. Of course, the exploration is limited and the RL problem is badly formulated. But still, you have to solve the temporal credit assignment problem (updating all tokens) which is not part of c-bandit.

Reinforcement learning once shook world politics, especially in China, with AlphaGo, and now ~9 years later, DeepSeek R1 is doing it again on a much larger scale. RL IS THE FUTURE, as I always believed. It's the ultimate chess move in the game of intelligence. #DeepSeek #AI #rl

SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training Shows that: - RL generalizes in rule-based envs, esp. when trained with an outcome-based reward - SFT tends to memorize the training data and struggles to generalize OOD

I just gave one of the hardest interview questions I had a year ago to OpenAI o3-mini-high and it solved it. Damn.

The head of my MSc lab just won the Turing Award 🎉

Today, we release QwQ-32B, our new reasoning model with only 32 billion parameters that rivals cutting-edge reasoning model, e.g., DeepSeek-R1. Blog: qwenlm.github.io/blog/qwq-32b HF: huggingface.co/Qwen/QwQ-32B ModelScope: modelscope.cn/models/Qwen/Qw… Demo: huggingface.co/spaces/Qwen/Qw… Qwen Chat:

If reinforcement learning models CAN cheat their way to maximizing the return, they WILL cheat!

David Silver really hits it out of the park in this podcast. The paper "Welcome to the Era of Experience" is here: goo.gle/3EiRKIH.

I’m thinking along the same lines these days. If anybody wants to collaborate/brainstorm on this let me know.

the mit 6.S184 lectures on flow matching and diffusion are really helpful for those who want to start with flow matching and in depth intuition behind it

It’s crazy how the demand for training coding models with RL has exploded in just a few months. People from finance to IT are literally throwing 💰 at me! Everybody wants their own specialized coding model now. Wild compared to just a few months ago.