Agent Horizon

A university lab just got an AI to react to conditions hotter than the core of the sun faster than any human ever could, while a tech CEO promised, again, that his chatbots will cure most human disease within five years. Only one of those claims has data behind it. This week’s actual news cycle was a good reminder that the boring, well-documented AI progress is still coming from labs and earnings calls, not from X threads.

Chips got cheaper to build around, cloud bills got real, and a couple of jobs reports quietly undercut two years of doom-laden headlines. Meanwhile, the usual suspects kept promising either salvation or extinction, sometimes in the same week. Here’s what actually mattered, and what was just noise wearing a lab coat.

Worth Reading

• Princeton’s PACMAN framework lets AI models control fusion plasma in real time — the system reacts in about 20 milliseconds and caught a destabilizing event roughly 200 milliseconds before it happened, all while humans stay in charge of the actual goals. This is the unglamorous, patient kind of AI progress that quietly moves fusion energy closer to reality, no marketing department required.
• Positron AI raised $875 million to build inference chips that skip scarce high-bandwidth memory entirely — swapping in ordinary LPDDR5X memory instead. It’s already running 50-plus racks in production at Oracle, which is a refreshing amount of receipts for a semiconductor startup this early in its life, and a good sign the industry is engineering its way around chip bottlenecks instead of just throwing more cash at the same shortage.
• Oracle’s cloud infrastructure revenue jumped 121% in a single quarter, beating Wall Street’s estimates and pushing the stock up sharply. Whatever you think of the broader AI investment debate, actual enterprises paying actual invoices for actual GPU capacity is a much sturdier signal than another slide deck.
• Apple shipped the first high-volume 2-nanometer smartphone chip, the new A20 Pro, built specifically to run AI models on-device rather than shipping everything to the cloud. It’s a good example of what real AI hardware progress looks like: incremental, unglamorous, and measured in transistor density rather than adjectives.
• A fresh look at nationwide employment data found no evidence that AI-heavy states or industries are shedding jobs faster than anyone else — in fact, the companies using AI most are hiring more, including younger workers who were supposedly first in line to be replaced. After two years of confident jobs-apocalypse headlines, it’s nice to see someone actually check the numbers.

Spot the Hype

• An AI company CEO says AI will cure “most human disease” within five to ten years, and doctors are, understandably, not having it. Curing most of human disease is an ambitious roadmap for an industry whose chatbots still can’t reliably cure themselves of making things up.
• A column argues it’s “time to clean house” and let AI replace some of the workforce, framing layoffs as a productivity audit rather than, you know, layoffs. Nothing says grounded analysis quite like rebranding job losses as spring cleaning.

That’s the week: a plasma reactor behaving itself, a chipmaker betting on cheaper memory instead of miracles, and one more CEO who thinks five years solves biology. What’s the most overhyped AI headline you ran into this week — and what actually impressed you? Drop it in the comments, and if you missed last time, catch last week’s digest here.