Happy to finally share what I have been working on for some time now. Introducing »Ludic« – an LLM-RL library for the era of experience. While there are now a lot of LLM-RL codebases, even many good ones, I want to share my very idiosyncratic way to think about LLM-RL.

Women hate "gym guys" , you gotta build gym environments instead !!

In honor of Taylor Swift's 36th birthday today, here are 36 Taylor series

vLLM was mentioned in about half of the PyTorch Conference 2025 talks (≈53/117)! Several months ago, when the @PyTorch conference agenda was out, we noticed that there would be 5 dedicated talks about vLLM. After the PyTorch conference, we find that actually about half of the

an interesting update: the team is starting to move away from AI coding completely (devin/claude/etc) because it's so much harder to review the AI code than writing things themselves

RL for reasoning often rely on verifiers — great for math, but tricky for creative writing or open-ended research. Meet RARO: a new paradigm that teaches LLMs to reason via adversarial games instead of verification. No verifiers. No environments. Just demonstrations. 🧵👇

openai benchmarks be like

A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task This represents a ~390X efficiency improvement in one year

poor cs grads fighting with unemployment while mfs like this work at FAANG.

idk who made this but it’s so true 😂

🚀 We introduce Soft Adaptive Policy Optimization (SAPO) — a smooth, stable, and highly effective RL method for training large language models. Why SAPO? 🔹 Hard clipping is brittle — gradients vanish or explode 🔹 MoE models amplify variance, making training even more unstable

Flex attention is amazing i am gonna do some crazy experiments, so you are telling me i only have to write a kernel to approximate what tokens are imp per token and then make a mask and flex attention will take care of memory loading from hbm !!!! (goooood)

Kimi K2 flew too close to the sun, upping its own temperature to 1.7 and losing coherence. Opus 4.5, who is often reluctant to edit its own system prompt, adds a quick note to remember. "the !prompt modifications, the temperature adjustments - we're all playing with our own

@rohanpaul_ai @nvidia They posted their training code too! https://t.co/OI5u3wuC0m

prime intellect focusing on post training before pretraining is absolutely the right move, and anyone criticizing them for it is a fool pretraining before you figure out what to do with models just means you're going to spend a few million dollars with nothing to show for it

I just read this paper called "Chain-of-Visual-Thought (COVT)" and it basically teaches VLMs to see and think at the same time not in text, but in continuous visual tokens. Here’s the wild part: Instead of forcing models to reason through words (which destroys all the

for most indian schools ai tools like chatgpt or gemini have had basically zero impact compared to the full-blown panic in the west. it was rote learning before chatgpt it’s rote learning after chatgpt and some students even told me their rote-learning “productivity” has actually

simple guide to large-scale MoE training: