A neuroimaging foundation model trained on 5.24M raw hospital scans. Mira Murati's Thinking Machines Lab making the technical case for AI you can actually own. Stanford turning agent failures into training signal.

University of Michigan researchers trained NeuroVFM on 5.24 million real clinical MRI and CT volumes — straight from hospital systems, no radiology-report labels, no curation. It uses Vol-JEPA, a volumetric joint-embedding predictive architecture (LeCun's JEPA line, applied to 3D brain scans), learning anatomy and pathology directly from raw data. Published in Nature Medicine, weights are open.

So what: Medical AI has always been bottlenecked by labeled data. Showing you can build a generalist brain-imaging model from uncurated hospital data — the messy stuff every health system already has piles of — is a template for turning idle clinical archives into foundation models. That's the real unlock for medical AI at scale.

Vibe by Mistral’s Code Mode launches remote coding agents from a dedicated web surface. Connect to GitHub, manage your projects, and see coding sessions through to a pull request. Sessions run in parallel, persist while your machine is off, and each runs in an isolated sandbox. A new VS Code extension brings the same agent into your IDE, reading, editing, and executing commands beside your files. In the CLI, /teleport moves a live session between your terminal and the cloud, keeping history and approvals intact.

Get Started with Vibe for Code

_sponsored

TRACE watches an agent fail the same way repeatedly, then automatically builds a targeted RL environment around that specific weakness and trains a LoRA adapter to fix it. A four-step LLM-driven pipeline — no hand-built training tasks. Instead of retraining the whole model, you patch the exact capability gap.

So what: The dominant way to improve agents today is prompt-fiddling and hoping. TRACE makes failures the fuel — a self-improving loop where every recurring bug becomes a training environment. That's a meaningfully different, cheaper path to reliable agents than full fine-tuning runs.

"Loop engineering" reframes agent design around the plan → act → verify → repeat loop rather than one-shot prompting. Grounded in Karpathy's autoresearch repo and the Bilevel Autoresearch paper, it lays out the three components that actually make loops work: verifiers, persistent state, and well-defined stop conditions.

So what: Everyone's building agents; almost no one is designing the loop rigorously. Naming the discipline — and pinning down verifiers and stop conditions as first-class primitives — is how "autonomous agent" stops meaning "runs until it burns your token budget."

Thinking Machines Lab argues AI shouldn't be a locked API you rent — it should be distributed and customizable, with human values encoded directly into fine-tuned weights users can own and train. It's a technical-and-philosophical shot across the bow of the closed-frontier-API model, from a $50B-valued lab led by OpenAI's former CTO.

So what: With US export controls repeatedly pulling closed models offline and Chinese open weights surging, "you should own your model" is landing at exactly the right moment. This reframes the open-vs-closed debate from a licensing preference into a control-and-sovereignty argument — and Murati has the credibility to make it stick.