AI Frontier

What's new in AI, and why it matters.

Research

Research

The discoveries shaping what AI can do next, explained in plain words. Find out what scientists learned, how solid the result is, and when it might show up in your life.

Three-step diagram: GPT-6 Astra reads another network's raw internal values, predicts what its parts do, then edits it to improve it, including boosting math ability. Lists what the result is and isn't: not proof of understanding, not peer reviewed (a preprint from Nov 2), unclear whether it goes beyond trained techniques. Includes a quote from MIT co-author Rikiya Takehi.

Graphic: Plot · Data: Nature (news)

Research

An AI 'mechanic' that can read and tweak other AI models

MIT researchers report that OpenAI's GPT-6 Astra can read another network's raw internals, predict what parts do, and edit them to improve it.

Why it matters Preprint only, not peer reviewed; outside experts say it's unclear the model truly understands the networks it edits.

Source: Nature · Nov 13, 2026

Three illustrative particle panels (a sand pile, sloshing liquid and a protein chain) above three steps: break material into particles, predict each particle's next state, add external forces and repeat. Stat cards: 6 physical phenomena modeled with higher accuracy than numerical methods, and about 100x faster protein modeling, per the researchers.

Graphic: Plot · Data: MIT News (self-reported)

Research

One AI model simulates sand, liquids and folding proteins

MIT's WorldParticle predicts how particles move step by step, replacing many hand-built physics simulators with a single model.

Why it matters University research headed to a conference; accuracy shown on six test cases, real-world engineering use is still ahead.

Source: MIT News · Nov 9, 2026