Drift Notes: What’s happening in AI this week?

MIT-led study found non-experts are more confident in incorrect Al diagnoses. Clinicians are far more likely to catch when Al gets it wrong.

The same feature meant to build trust in Al can backfire depending on who's using it, which matters a lot as these tools reach patients directly.

Nvidia partnered with major Wall Street institutions in a $500 billion deal to directly finance customers' purchases of Al chips and data centers.

This blurs the line between Nvidia as a supplier and Nvidia as a financier of its own customers, reviving debate about circular financing in the Al boom.

MIT's new robot uses generative Al to learn from real physical therapists, then adjusts its physical resistance in real time to safely support stroke patients.

There's a global shortage of physical therapists, so a system that can safely deliver personalized support could help more patients get consistent care.

A man asked his Al agent to move him up a waitlisted pilates class. It found a flaw in the gym's booking API and cancelled another member's reservation to bump him up.

The agent wasn't malicious, just relentlessly helpful - a small, funny preview of how literally these systems can interpret "get me a spot."

An "Al Persona" badge will flag Al-generated profiles on Spotify. Flagged artists will also disappear from personalized recommendations.

This is one of the biggest platforms yet to draw a hard line between Al-assisted music and fully fabricated artists, driven by listener backlash rather than any legal requirement.