Amazon, Adobe, Atlassian, Citi, and others are quietly capping AI usage as token costs spiral out of control. The gap between the AI productivity narrative and the actual economics is becoming impossible to ignore.
Researchers found people rated AI impersonators as more authentic, coherent, and relevant than the actual politicians. This isn't a future threat — it's a present crisis for democratic discourse.
Geoffrey Litt's framing from the AI Engineer Summit: as coding agents build increasingly complex systems, developers risk losing meaningful comprehension of their own codebases. The problem isn't speed — it's legibility.
Simon Willison ships an early-alpha coding agent built on top of his LLM library, now maturing into a proper agent framework. A Fable 5 experiment worth watching as the open ecosystem catches up to proprietary agents.
A security flaw in Apple's privacy feature could allow attackers to unmask users' real email addresses. The researcher who found it says users deserve to know the risk exists now.
Mirendil cofounders — ex-Google and Anthropic researchers — make the case for self-accelerating AI systems that contribute to their own development. The recursive loop is no longer theoretical.
Flagged the whiplash moment of the week: Anthropic's Fable 5 launched, got yanked overnight by government export controls, then returned after the Department of Commerce reversed course. Whatever your view on AI regulation, this episode proved that frontier models can now be treated as controlled munitions — there one morning, gone by night.
Mollick captures the core split in how people think about AI right now: some "feel the exponential" and are building accordingly, while others are unconsciously treating AI as a productivity tool that plateaus. He doesn't say who's right — but the framing itself is clarifying. Which mental model you're running determines every decision you make.
Found a token optimization playbook for Fable 5 that cuts costs significantly — right on cue as the "tokenpocalypse" story breaks. The timing is not coincidental. Cost management is becoming its own skill layer on top of prompt engineering, and the people who figure it out early will have a durable edge.
Amazon research shows that running panels of LLM judges and averaging their scores dramatically outperforms single-judge evaluation. A quiet but important finding: the reliability problem in AI evaluation may be more tractable than assumed — just run the committee, not the individual.