Introducing *SequenceMatch*, training LLMs with an imitation learning loss Avoids compounding error in generation by: 1. Training against *different divergences* like χ^2 with more support OOD 2. Adding a *backspace* action: model can correct errors! arxiv.org/abs/2306.05426 1/7

Check out our new paper! rail-berkeley.github.io/grif/

Announcing FlashAttention-2! We released FlashAttention a year ago, making attn 2-4 faster and is now widely used in most LLM libraries. Recently I’ve been working on the next version: 2x faster than v1, 5-9x vs standard attn, reaching 225 TFLOPs/s training speed on A100. 1/

We use gestures all the time for specifying targets! How can robots make sense of “gimme that one”? We propose GIRAF, a framework for interpreting human gesture instructions using LLMs. Paper to appear in Conference on Robot Learning: arxiv.org/abs/2309.02721 Website: tinyurl.com/giraf23

Hungry? Let our robot twirl your spaghetti for you! 🍝🤖 Introducing VAPORS: Visual Action Planning OveR Sequences, a framework for long-horizon food acquisition. Project Page: sites.google.com/view/vaporsbot Paper: arxiv.org/abs/2309.05197 To appear at Conference on Robot Learning 1/11🧵

Releasing an open-source PyTorch implementation of DoReMi! github.com/sangmichaelxie… The pretraining data mixture is a secret sauce of LLM training. Optimizing your data mixture for robust learning with DoReMi can reduce training time by 2-3x. Train smarter, not longer!

It's an absolute honor to be a guest on the The TWIML AI Podcast podcast! Sam Charrington and I cover everything under the sun in LLMOps, from hallucinations, RAG to LLM safety. Check out the podcast on the link below!

We are excited to announce the first workshop on Foundation Models, Large Language Models (LLMs), and Game Theory! The workshop will take place at the Center for Discrete Mathematics and Theoretical Computer Science (DIMACS) on October 19-20, 2023. dimacs.rutgers.edu/events/details…

We also call on researchers with ongoing research to submit posters to our workshop. The workshop will provide financial support to a limited number of researchers. Applications for financial support will remain open until September 23, 2023. docs.google.com/forms/d/e/1FAI…

Excited about recent improvements of our NeurIPS Spotlight paper, now even faster with ⚡️⚡️multiprocessing⚡️⚡️! We now get 2x speedup on as low as 50-step DDIM, and 4x speedup on 200-step DDIM! The first version of our paper showed good results, but we wanted even better. -

Eliciting Human Preferences with Language Models Currently, people write detailed prompts to describe what they want a language model to do We explore *generative elicitation*—where models interactively ask for this information through open-ended conversation 1/

I also have no idea what Q* is, but given speculation that it’s a method of self-learning and Monte-Carlo Tree Search (MCTS) in language models, I thought I’d share some recent work on an adjacent idea.

Achievement unlocked ✅, thanks for the shout-out Andrej Karpathy!

We’re hiring for the adversarial robustness team Anthropic! As an Alignment subteam, we're making a big effort on red-teaming, test-time monitoring, and adversarial training. If you’re interested in these areas, let us know! (emails in 🧵)

Made a short video exploring tool use and subagents! (w/ @aaron_begg and everett) Goal: Find the “quickest quicksort” implementation on GitHub by having a larger model orchestrate 100 subagent models Here’s how it works: 1/ x.com/AnthropicAI/st…

At #ICRA24 we've a few papers on 𝗴𝗿𝗼𝘂𝗻𝗱𝗲𝗱 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 of LLMs/VLMs. • Grounded common-sense reasoning via active perception - Minae Kwon's 🧵👇 • Physically grounding VLMs - Jensen Gao's 🧵👇 • Learning from online language corrections - @lihanzha's 🧵👇

Introducing Claude 3.5 Sonnet—our most intelligent model yet. This is the first release in our 3.5 model family. Sonnet now outperforms competitor models on key evaluations, at twice the speed of Claude 3 Opus and one-fifth the cost. Try it for free: claude.ai

I’m taking applications for collaborators via ML Alignment & Theory Scholars! It’s a great way for new or experienced researchers outside AI safety research labs to work with me/others in these groups: Neel Nanda, Evan Hubinger, mrinank 🍂, Nina, Fabien Roger, Rylan Schaeffer, ...🧵

Introducing an upgraded Claude 3.5 Sonnet, and a new model, Claude 3.5 Haiku. We’re also introducing a new capability in beta: computer use. Developers can now direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking, and typing text.

We’re publishing a new constitution for Claude. The constitution is a detailed description of our vision for Claude’s behavior and values. It’s written primarily for Claude, and used directly in our training process. anthropic.com/news/claude-ne…