Whenever I feel whiny about problems I have I remind myself that as an SF tech worker I live in a city with 259 sunny days, make more money at 22 than I ever would have in Germany and am surrounded by people that share the same interests and ambitions as me like nowhere else.

RT @designarena_ai: Kimi K2 by @Kimi_Moonshot and Mistral Small 3.2 by @MistralAI just added to leaderboard . Crown your winner at https://….

SYNTHETIC-2 Datasets are now on Huggingface! We’re releasing an SFT dataset collected from the new R1-0528 as well as an RL Dataset with difficulty annotations from various smaller models. Go train some models 🫡.

Highest leverage thing unskilled engineers can do rn to contribute to frontier AI research is vibecoding RL environments.

First tweet is a banger, I will watch his career with great interest.

We are so back btw (got submitted in final but hey)

great thread, summarizes well why we're particularly excited about RL. higher inference to training compute ratio -> less inter node communication -> better suited for globally distributed training infra with slow connection speeds.

It happens too often actually, I’ve wasted so much money on food delivered to the wrong address.

Been on a rough weight cut for comp for the past two weeks and this is what happens when I want to reward myself

RT @khoomeik: Recently had a good chat with @tamaybes. He thinks we aren’t yet in the GPT-3 era of RL and as it scales, cross-task OOD gene….

I'm wondering to what degree a VLM judge matches human preferences here.

RT @PrimeIntellect: We did it — SYNTHETIC‑2 is complete. A planetary-scale decentralized inference run generating 4M verified reasoning sa….

Prime Intellect is nothing without its child laborers.

slides sneak peek (actually)

I’ll be at the Democratize Intelligence Summit today to speak about why reinforcement learning is ideally suited for a decentralized training setup + share a bit about our work at @PrimeIntellect . Come say hi!.

excited about the training run I'm starting today.

I'm especially looking forward to Deepseek-R1-0528 distilled models that can be trained using SYNTHETIC-2. The distilled model based on Qwen 8B that is out already outperforms Qwen3-32B without any RL (at least on benchmarks), which is quite impressive

RT @khoomeik: there is no excuse for any gpu on earth to be idling right now. every idle gpu can and should contribute to generating high q….

If you want to read more, check out our blog post - we’ll also share more about the data curation pipeline soon!.