Beff Jezos — e/acc you cannot outaccelerate me

Vicuna-v1.1 is out with improved data processing and tokenization! Want to learn how we trained it for <$300 with SkyPilot even when A100 is so hard to get in the clouds? Vicuna-v1.1 weights: github.com/lm-sys/FastCha… Run the example: github.com/skypilot-org/s…

Skypilot! This is how we imagined the cloud to work: you define an ML job, it will find the cheapest places to run it and then does the work for you. I made an example for how to use it to finetune an LLM with your own data github.com/artidoro/qlora… ( cost $4 all-in )

Latent space modeling of social networks is just creating embeddings for people stat.cmu.edu/~brian/905-200…

Prompt injection is a huge security problem for LLM apps. To study this, we built Tensor Trust: a game where you create and defend against prompt injections. We’re releasing a paper + dataset with 70k unique attacks, 40k unique defense prompts, and new robustness benchmarks. 👉

We thank a16z for the generous AI grant to support SkyPilot's open-source development and cloud expenses. Onwards to building the best OSS framework for AI training + serving on any cloud/on-prem infra!

building is fun; seeing what you've built get used by others is indescribable

🔥Exciting news -- GPT-4-Turbo has just reclaimed the No. 1 spot on the Arena leaderboard again! Woah! We collect over 8K user votes from diverse domains and observe its strong coding & reasoning capability over others. Hats off to OpenAI for this incredible launch! To offer

Meet Reka Core, our best and most capable multimodal language model yet. 🔮 It’s been a busy few months training this model and we are glad to finally ship it! 💪 Core has a lot of capabilities, and one of them is understanding video --- let’s see what Core thinks of the 3 body

🔮

you wouldn’t last an hour in the asylum where they raised me

Check out our latest blog analyzing Llama 3’s performance on Chatbot Arena! My favorite stat: Llama 3 outputs exclamation marks more often than its opponents in over 80% (!) of battles against other top models.

lenin once said that there are decades where nothing happens and weeks where decades happen. every so often, at times like this, i am reminded of the ebb and flow of life - of the uneven pace of change, and the rising and crashing of the waves.

Not all questions need GPT-4! We introduce RouteLLM – a routing framework based on human preference data that directs simple queries to a cheaper model. With data augmentation techniques, RouteLLM achieves cost reductions of over 85% on MT Bench and 45% on MMLU while

So excited to finally share our work on RouteLLM! We investigate techniques for training routers using preference data and show significant cost reductions while maintaining high-quality responses - we even beat commercial offerings. Everything open sourced!

Checkout RouteLLM - a great helper for saving money on tokens using matrix factorization for routing models. Cool to see even more innovations coming out of LMSYS Org youtube.com/watch?v=V_K6PC…

Coding with Claude

Machines of Loving Grace: my essay on how AI could transform the world for the better darioamodei.com/machines-of-lo…

📢We’re excited to share that we’ve raised $100M in seed funding to support LMArena and continue our research on reliable AI. Led by a16z and UC Investments (University of California), we're proud to have the support of those that believe in both the science and the mission. We’re

Introducing the next generation: Claude Opus 4 and Claude Sonnet 4. Claude Opus 4 is our most powerful model yet, and the world’s best coding model. Claude Sonnet 4 is a significant upgrade from its predecessor, delivering superior coding and reasoning.