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

Japanese electronics company NEC has launched a fascinating pilot project: a new department filled by 39 AI agents, from department head to staff, to test how AI can support management. They prepare reports, check rival news every morning, and hold a weekly executive meeting. After one agent mistyped a number, its AI boss even “interviewed” it about preventing a repeat.

The agents gather information to support the people making the final decisions, but NEC wants hundreds of them in the department this year. It’s a concrete look at what an AI-run team could be like day to day, including the awkward performance conversations.

In experiments with 1,222 people, those who used an AI tool to solve problems did better at first. Once it was taken away, they got more answers wrong and skipped more questions, while people who never had it kept going and finished stronger. The study specifies that “spending just 10 minutes relying on an AI tool significantly impairs our ability to focus and persist with difficult tasks.”

It shows how quickly we get used to instant answers, so anything slower (i.e. our own brains) starts to feel inefficient. The researchers suggest AI should act more like a tutor than an answer machine. But more importantly, it’s a reminder to train those brain muscles every day.

Amazon’s “About You” feature builds a profile from your purchases, reviews, searches, and chats with its shopping assistant. The feature isn’t new — Amazon launched it back in May — but it made the rounds on social media again this weekend, with users posting their oddly specific discoveries. From “uses Kindle in the shower to read” to “husband has skinny legs”, Amazon knows a freakish amount of details about its shoppers.

You can edit or delete what it assumes, but you can’t opt out entirely. It’s a good reminder that a profile built from a few purchases is mostly guesswork but also, that everything you type, search, and see on the internet is used to build a profile about you.

Misdiagnosis is common, and the gold-standard test can mean a wait of two to three months. This was a retrospective study, so real-world trials still need to confirm it helps patients. Despite that, AI trainers contributing to niche, complex STEM projects are seeing their hard work paying off in real life.

At separate events in Singapore, TikTok CEO Shou Chew said “the technology is not there yet,” noting that human brains still beat AI at things like visual nuance. Microsoft’s Judson Althoff added that “there’s no path where somehow this notion of infinite cognitive compute capability replaces human purpose.”

AI’s blind spots are where trainers come in. Leaders at TikTok and Microsoft are saying out loud what contributors see every day: models still miss nuance, and it takes a person who knows the subject to catch it.