After a year of team work, we're thrilled to introduce Depth Anything 3 (DA3)! 🚀 Aiming for human-like spatial perception, DA3 extends monocular depth estimation to any-view scenarios, including single images, multi-view images, and video. In pursuit of minimal modeling, DA3

🚀Excited to share our NeurIPS Conference paper Compositional Diffuser (CompDiffuser)! CompDiffuser scales planning horizons at test time, able to construct long-horizon plans while only trained on short-horizon data. Project page: comp-diffuser.github.io Thread 👇(1/n)

🎉 Our diffusion-model monograph 《The Principles of Diffusion Models》 now has an official website! 🔗 Link in thread 💡 Two highlights: 1️⃣ Blog Post: a big-picture walkthrough from diffusion models → fast generation & flow map models; 2️⃣ Teaching Guide: lightweight course

Everything is a world model if you squint hard enough.

To everyone who was questioning the muon results: You are right! The LR was badly tuned, and weight decay was also hampering it. Using lr 1e-3 and 0 wd, got a new best, 3.39 FID @ 400k

Whenever I read about LLMs as a jury, I wonder how different these LLMs really are. Ideally, one would assume their errors are uncorrelated, so that if their mistakes are independent, the error rate becomes epsilon squared. But their base models have all been trained on the

This deserved its own flashcard b/c I've seen bus stop ads from eval vendors encouraging the opposite in San Francisco 🤣 The only thing generic metrics do is waste your time. Links in reply.

One of the best visual explanations I've ever seen for why scaling Transformers works, but is suboptimal, as it's just brute-forcing things, by Llion Jones (co-author of the Transformer) on Machine Learning Street Talk "In the (rejected) paper "Intelligent Matrix Exponentiation", they show

Quite curious about Grokipedia! May be a great opportunity to make it a verifiable large-scale knowledge graph for better context management via graph traversal. Reduces LLM need to memorize the world, enabling a "cognitive core"-like approach (see Karpathy's tweets). Do wonder

We’re launching full-length, on demand practice exams for standardized tests in Google Gemini, starting with the SAT, available now at no cost. Practice SATs are grounded in rigorously vetted content in partnership with The Princeton Review, and Gemini will provide immediate feedback

Self-directed learning through AIs is an autodidact’s paradise.

If science class explained everything with Midwest emo music, we would have a lot more people paying attention. Nuclear power explained. It’s just a fancy way to heat water and make electricity. 🔊

This was an eye opener from Jensen Huang When asked whether he would rather relive his 20s or be 20 years old today, this is what he had to say: "I thought our 20s were happier than these 20s. I think everyone deserves some time to be oblivious, and not wear all of the world's

Anything you can do in Obsidian you can do from the command line. Obsidian CLI is now available in 1.12 (early access).

I have been testing the new Obsidian CLI with Claude Code on my research vault (4,663 files, 16 GB)... I know too many notes!! Early results are significant. Its going to change the way in which Claude Code can interact with Obsidian The way I see it, there are three ways

The next big level up for me wont be a better model, just boring infra. Better container orchestration. Personal compute networking. Secret management. Environment replication. Designed for a world where one person is running 100+ parallel workstreams.

I taught Claude to talk like a caveman to use 75% less tokens. normal claude: ~180 tokens for a web search task caveman claude: ~45 tokens for the same task "I executed the web search tool" = 8 tokens caveman version: "Tool work" = 2 tokens every single grunt swap saves 6-10

Big Update🤩: #paperclip now includes full papers from all of arXiv, PubMed Central and 150 million abstracts!🖇️ You can give your LLM all that knowledge in one line—all optimally indexed for AI agents. Much more thorough and ~100x faster than web search, and free.