Twitter Beğeni Hilesi

sites.google.com/corp/berkeley.… Excited to share work led by amazing Kevin Lu in collaboration with Aditya Grover and Pieter Abbeel! What’s holding us back from agents that learn in reset-free, lifelong settings?

What are the limits to the generalization of large pretrained transformer models? We find minimal fine-tuning (~0.1% of params) performs as well as training from scratch on a completely new modality! with Kevin Lu, Aditya Grover, Pieter Abbeel paper: arxiv.org/abs/2103.05247 1/8

Complementary blog post to our paper, Pretrained Transformers as Universal Computation Engines, has been released! bair.berkeley.edu/blog/2021/03/2…

Excited to share our new work on applying language modeling ideas to RL policy optimization! Jointly led with Lili and other amazing collaborators.

Really exciting work looking at how to utilize frozen language models for multimodal tasks! Great to see more successes in this direction.

Currently It is challenging to measure progress in Unsupervised RL w/o having common tasks & protocol. To take a step in addressing this issue we release our #NeurIPS2021 paper: (URLB) Unsupervised RL Benchmark! Paper: bit.ly/3bwHhY8 Code: bit.ly/3bAvI1S 1/N

Come chat with us about sequence modeling for reinforcement learning NeurIPS Conference tomorrow (Thurs 12/9) at 8:30-10am PT! gather.town/app/XRWlik7kvt…

LLMs like GPT-3 and Codex contain rich world knowledge. In this fun study, we ask if GPT like models can plan actions for embodied agents. Turns out, with apt sanity checks, even vanilla LLMs without any finetuning can generate good high-level plans given a low-level controller.

In our new work - Algorithm Distillation - we show that transformers can improve themselves autonomously through trial and error without ever updating their weights. No prompting, no finetuning. A single transformer collects its own data and maximizes rewards on new tasks. 1/N

I recently joined OpenAI! Come check out our new model: 82% MMLU at 60 cents per 1M output tokens! openai.com/index/gpt-4o-m…

Excited to release our new small model, developed by a killer crew of team players. Intelligence per $ is very strong with GPT-4o mini. Your turn, developers! omniminiomniminiomnimini (say it 5 times fast)

Exciting Chatbot Arena Update -- GPT-4o mini's result is out! With 4K+ user votes, GPT-4o mini climbs to the top of the leaderboard, now joint #1 with GPT-4o while being 20x cheaper! Significantly better than its early version ("upcoming-gpt-mini") in Arena across the boards.

“OpenAI says that more than 200 million people use ChatGPT each week […] while API usage has doubled following the release of the company’s cheaper and smarter model GPT-4o mini” Has OpenAI API usage really doubled in the past five weeks since 4o-mini? theverge.com/2024/8/29/2423…

Thrilled to release o1-mini, a model near and dear to my heart 💙. o1-mini is an efficient model in the o1 series that’s super performant in STEM reasoning, especially math and coding. I can’t wait to see what you all build with o1-mini!! openai.com/index/openai-o…

A Codeforces contestant used o1-mini in a live contest today codeforces.com/blog/entry/133… and achieved near master-level performance! Agree with the decision to restrict AI in competition going forwards, and it'll be interesting to see how the contest scene evolves.