adaptive test-time compute!
how long to plan for? how to set n_diffusion_steps? how many tokens should LLM spend on my task?
super excited that we improved beyond GPT-5 and RL'ed a model to think adaptively wrt estimated difficulty. with the right incentive, it's all emergent!
adaptive test-time compute!
how long to plan for? how to set n_diffusion_steps? how many tokens should LLM spend on my task?
super excited that we improved beyond GPT-5 and RL'ed a model to think adaptively wrt estimated difficulty. with the right incentive, it's all emergent!
1/n
I’m really excited to share that our @OpenAI reasoning system got a perfect score of 12/12 during the 2025 ICPC World Finals, the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems. This would have…
I copy pasted an unpublished manuscript of mine in ChatGPT and asked it to improve it. I expected that the method we're using was pushed to its limit: gpt-5-pro actually proved it.
Even I did not expect the models to be capable of such things already ...
In deep learning, once you figure out how to go from 0->1, scale usually takes you from 1->100. This is what the past 10 years in deep learning has shown: once things work a little bit, it doesn’t take too long for them to work quite well (see the history of image…
Good news! GPT-5 should now be significantly faster in Cursor.
Big thank you to OpenAI for rolling out improvements to caching and API latency! P95 is ~2x faster versus before.
1/n I’m thrilled to share that our @OpenAI reasoning system scored high enough to achieve gold 🥇🥇 in one of the world’s top programming competitions - the 2025 International Olympiad in Informatics (IOI) - placing first among AI participants! 👨💻👨💻
This is an important and less discussed fact about GPT-5. Not only is it better overall than o3 at tasks like SWE-bench, GPT-5 is much smarter per-token. This is something users will feel.
GPT-5 is just more intelligent per-token in the way it thinks through problems and calls…
This is an important and less discussed fact about GPT-5. Not only is it better overall than o3 at tasks like SWE-bench, GPT-5 is much smarter per-token. This is something users will feel.
GPT-5 is just more intelligent per-token in the way it thinks through problems and calls…
GPT-5 is finally out! Amongst its many improvements, I am personally very excited for everyone to try its enhanced coding abilities — both agenetic and front-end.
And give it your hardest tasks in your favourite IDE or vibe code feature rich and tasteful webapps.
GPT-5 is finally out! Amongst its many improvements, I am personally very excited for everyone to try its enhanced coding abilities — both agenetic and front-end.
And give it your hardest tasks in your favourite IDE or vibe code feature rich and tasteful webapps.
sharing today the gpt-5 prompting guide coauthored by yours truly, with special tricks & tips for agentic coding courtesy of @cursor_ai. wishing you all happy prompting - and remember, when all else fails, just try asking the model to be good™️
cookbook.openai.com/examples/gpt-5…
Along with GPT5, we're open sourcing a new eval, BrowseComp Long Context!
It improves upon existing long context qa evals in data quality and input difficulty. Work with @LK112358, @julieswangg, and our mascot the longham.
A bit more below
we've been testing some new methods for improving writing quality. you may have seen @sama's demo in late march; GPT-5-thinking uses similar ideas
it doesn't make a lot of sense to talk about better writing or worse writing and not really worth the debate. i think the model…
we've been testing some new methods for improving writing quality. you may have seen @sama's demo in late march; GPT-5-thinking uses similar ideas
it doesn't make a lot of sense to talk about better writing or worse writing and not really worth the debate. i think the model…
1K Followers 8K FollowingAI inference, speculative decoding, open source. Built novel decoding algorithms – default in Hugging Face Transformers (150+ ⭐). Making AI faster + cheaper
5 Followers 253 FollowingIn the future, everyone will have their own exclusive large models or agents, and personal data, privacy, etc. will no longer be occupied and fragmented by
35 Followers 1K FollowingBy day I talk about art and investments, by night I’m planning my next trip. Don’t worry, I won’t calculate the ROI of our first date 😉.”
994K Followers 3K FollowingSBF's words (mostly). Shared by a friend.
✉️ (monitored): Samuel Bankman-Fried 37244-510, FCI Terminal Island, PO Box 3007, San Pedro, CA 90733
91K Followers 925 FollowingCEO at @magicpathai 🎨✨
Previously, @AnthropicAI, @brexHQ. @Uber, @Facebook. Creator of Claude Engineer, DesignerGPT, Sequential thinking MCP and more
10K Followers 798 FollowingThinking about whether AI will destroy the world at https://t.co/pMilDvd4ya. DM or email for media requests. Feedback: https://t.co/zGAm1i7SKH
1K Followers 8K FollowingAI inference, speculative decoding, open source. Built novel decoding algorithms – default in Hugging Face Transformers (150+ ⭐). Making AI faster + cheaper
186K Followers 62 FollowingBuilding new freedoms of imagination for the world through pioneering research and design. Try Dream Machine for free → https://t.co/LmWmA4H803
29K Followers 431 FollowingProfessor, CS, U. British Columbia. CIFAR AI Chair, Vector Institute. Sr. Advisor, DeepMind | ML, AI, deep RL, deep learning, AI-Generating Algorithms (AI-GAs)
209K Followers 101 FollowingThe original AI alignment person. Understanding the reasons it's difficult since 2003.
This is my serious low-volume account. Follow @allTheYud for the rest.
2K Followers 172 FollowingResearch Director & Principal Scientist @GoogleDeepMind, Gemini Team | Lead of LaMDA LLM & AI | Worked on Duplex, TensorFlow, Wide & Deep Learning | Hiring!
15K Followers 528 FollowingAsst. Prof. of CS at Stanford, Google DeepMind. Prev: Anthropic, Google Brain. Co-Creator of MoEs, AlphaChip, Test Time Scaling Laws.
No recent Favorites. New Favorites will appear here.