Barret Zoph and I recently gave a talk at Stanford on post-training and our experience working together on ChatGPT. Unfortunately the talk wasn't recorded, but here are the slides: docs.google.com/presentation/d…. (If you have a recording, please let me know!)

🎉 Introducing Open Reasoner Zero 🚀 Performance: Matches DeepSeek R1-Zero (32B) in just 1/30 steps! 📚 Full training strategies & technical paper 💻 100% open-source: Code + Data + Model ⚖️ MIT licensed - Use it your way! 🌊 Let the Reasoner-Zero tide rise! 🚢 1/n

Fair enough. Here's my compilation of all results from relevant sources on AIME 2025 performance of Grok and OpenAI models, plus extrapolations of cons@64 for DeepSeek models and o1. I think this is significantly easier to understand than chart crimes of these frontier labs.

🚀 Day 0: Warming up for #OpenSourceWeek! We're a tiny team DeepSeek exploring AGI. Starting next week, we'll be open-sourcing 5 repos, sharing our small but sincere progress with full transparency. These humble building blocks in our online service have been documented,

🚀 Day 1 of #OpenSourceWeek: FlashMLA Honored to share FlashMLA - our efficient MLA decoding kernel for Hopper GPUs, optimized for variable-length sequences and now in production. ✅ BF16 support ✅ Paged KV cache (block size 64) ⚡ 3000 GB/s memory-bound & 580 TFLOPS

It would take a long ass article to articulate this properly but this is not a vageupoast. I have spent the last few months working on some very hard problems (more on that soon). I've been using a combination of R1 and DeepResearch to build and formalize the ideas and proofs.

🚀 Day 6 of #OpenSourceWeek: One More Thing – DeepSeek-V3/R1 Inference System Overview Optimized throughput and latency via: 🔧 Cross-node EP-powered batch scaling 🔄 Computation-communication overlap ⚖️ Load balancing Statistics of DeepSeek's Online Service: ⚡ 73.7k/14.8k

Bro, your post suggests many influential people are never aware of X's several key features: > Highlight - marks important tweets amid sheetposts > X Lists - tracks themed accounts across areas, more organized than following > X Explore - summarized what's happening around >

my 2025 ai twitter experience

I've been saying that DeepSeek will expand from verifiable to general domains, and expected a paper. Here is that paper. Self-Principled Critique Tuning. rule-based online RL. Gemma-2 27b is enough to match R1. This is roughly what Google does for Gemma 3 and likely Geminis.

A couple years of weekly analysis, frontier research, and writing a small book on RLHF was pretty much a long winded lead up to writing this blog post. If nothing else, read it as a favor to me. interconnects.ai/p/sycophancy-a…

VR-CLI is an obscenely powerful RL objective that was mentioned in a paper that wasn't hyped to 1/10th of the degree it deserved. "oh, you can optimize the reasoning traces for next-token prediction in a way that generalizes WAY better..." ...casual bombshell implications.

the last week of launches has highlighted a few things for me: - Progress in LLMs has been amazing but incremental capability gains are clearly closer to log than linear while the corresponding cost for those gains is closer to exponential than linear. scale still works but the

simple "LLM as a judge" protip if you prompt for something like "provide answers to the TRUE/FALSE rubric questions in order, followed by a one sentence justification" this will be worse than the justification coming *before* the TRUE/FALSE marker

being asked how to experience the full version of the Kimi K2 API from our official platform. Simply visit platform.moonshot.ai, click on 'Console' and then 'Recharge.' You will receive a $5 voucher after your first successful payment. We have a clear rate limit schedule, and

just finished the tech report and pushed to github. good night.

many on this app should listen to kimi k2 more