I RL-trained @thinkymachines Inkling model to tell dad jokes, all from my laptop. Inkling is a 975B open-weights MoE. I fine-tuned it with GRPO on @tinkerapi into PunTune-0.6, which answers any topic with a focused, original dad joke. The jokes are... fine. But the finding was insightful. The main discovery was that training didn't teach the model new wit. It taught it to pick one punchline... and stop. Something a lot of great models (and my mom) really struggle with. PunTune reaches strong punchlines on about a quarter of the tokens as the base Inkling model. My rewards bought reliability and economy, not necessarily creativity. This is the type of learning that helps product leaders start to connect how training, and understanding what is/isn't possible, can sharpen the value your AI models deliver. Especially when...

New research from Bridgewater AIA Labs, UIUC, and MIT: we prove what we believe to be the first non-vacuous generalization bounds for reasoning LLMs on real-world problems. RLVR powers frontier reasoning capabilities yet its generalization to unseen data has remained an open theoretical question and deployment blocker for practitioners. Our generalization bounds for RLVR deliver provable high-probability lower bounds of the accuracy for billion-parameter RLVR models on unseen data, which can provide guidance on safely deploying RLVR. 1/9

RL Systems Mind the Gap: Matching Trainer and Generator Throughput RL Training Infrastructure, GRPO, PipelineRL, Async RL, Policy Staleness, RL Sandbox Infra, CPU Requirements, TCO Analysis, Thinking Machines Tinker

Today we're shipping Nemotron 3 Ultra. A 550B MoE frontier-intelligence open model built for long-running agents. It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models.

New preprint from @lightningrodai! We trained AI to predict clinical events — ICU transfers, new diagnoses, complications, procedures, ventilation, mortality — directly from raw clinical notes. No labeled data required – Foresight Learning infers outcomes from what happens later in patient records. Using Tinker from @thinkymachines , we trained a lightweight adapter on GPT-OSS-120B, resulting in a specialized predictor that runs on a single GPU. Results: 🎯 ~70% lower calibration error 📈 Brier skill score: ~0% → 27% 🧠 84% win-rate vs the base model in blind reasoning review 🥇 Slightly better Brier than GPT-5, despite being a fraction of the size Hospitals and specialty clinics often treat unique patient populations that out-of-the-box models don't have training data for. This makes it possible to build frontier-quality predictors for highly specific patient groups, with nothing but raw clinical records. Congrats to the team — @indiequant @KSkotheim64001 🙌 Full paper 👇

How does Exa compare to Google for training LLMs to search? In this blog post, we find that LLMs using Exa during reinforcement learning reach higher performance with 70% less training compute.

In Agent RL, models suffer from Template Collapse. They generate vast, diverse outputs (High Entropy) that lose all meaningful connection to the input prompt (Low Mutual Information). In other words, agent learn different ways to say nothing. 🚀 Introducing RAGEN-v2 -- Here's how we define and fix such silent failure modes in Agent RL. 🧵

I pointed Claude Code at a research task (build a golf forecasting system) and let it run for 49 hours on Tinker. No human in the loop. It ran 108 experiments. Here's the full trajectory, including the ones that made things worse.