The archive
Research results and the fundamentals behind the models.
MIT researchers built a memory system that scores how confident a vision-language model is about objects a robot has seen, then re-captions the shaky ones. Early numbers are modest but pointed.
5 min read
Anthropic reports that Claude, run in a closed loop with an automated scorer, found quantum circuit designs that beat the best published results on eight primitives — including a 12x cheaper decoder.
6 min read
Anthropic's Frontier Red Team says an internal Claude model traced the NIST elliptic curve seeds to a 1990s C library random number generator seeded by the clock — in under 30 minutes.
5 min read
Researchers at MIT and the Whitehead Institute say each developmental signaling pathway leaves a consistent gene-activity "fingerprint" — and a transfer-learning model can read it across cell types.
5 min read
MIT associate professor Christina Delimitrou argues that cloud infrastructure wastes enormous capacity, and that machine learning can reclaim it. The profile is light on hard numbers.
5 min read
An MIT profile of associate professor Cathy Wu traces years of failed RL experiments to a 2023 result: train on the 10% of problems that generalize, and skip the rest. Up to 30x more efficient.
5 min read
Anthropic's new robot exposure index finds today's machines can perform 74% of physical US job tasks, but are cost-competitive on just 0.3% of work. Economics beats capability.
5 min read
AWS walks through GRPO post-training of Qwen3-VL-8B on SageMaker HyperPod with SkyRL, reporting a maze solve rate climb from 43.75% to 96.875% on a 64-maze evaluation set.
6 min read
Researchers at MIT's McGovern Institute used an LLM to build a 49-factor risk lexicon, then trained a small, interpretable model on roughly 16,000 crisis conversations.
5 min read
A workshop recap from Ai2 and Providence Swedish names five obstacles to AI-driven discovery, from scientific taste to the risk of amplifying flawed methods.
5 min read
MIT and Harvard researchers report xvr, a method that aligns live surgical X-rays with a patient's 3D scan in seconds, trained per patient in about five minutes.
5 min read
MIT researchers report a deployment-time method that forces pretrained flow-matching models to satisfy hard safety constraints, with no retraining required.
5 min read