AI Weekly

An op-ed in Tech Policy Press by three academic researchers is worth a look if you touch AI governance or procurement, because it takes aim at a shortcut a lot of compliance regimes are quietly building around: treating the system prompt as evidence that a model is safe.

A new ICML 2026 study analyzed 60 language model benchmarks against 14 saturation-related properties and found nearly half exhibit saturation.

Legal scholars Deven Desai and Mark Riedl argue software interaction protocols can discipline AI agents more effectively than new regulations governing human agents.

A peer-reviewed study in npj Climate Action estimates AI could add 0.47 to 1.8 gigatonnes of CO2 emissions per year through fossil-fuel productivity gains.

A parallel bet on how to build intelligent machines is running alongside the silicon AI race, and it uses actual human brain tissue.

Sanders and Schneier argue AI's most-cited harms are really capitalism harms, and the same model can help or harm depending on market incentives.

A six-author paper out of a small research group is making an unusually direct claim: an agent that rewrites its own tools, prompts, context assembly, and core implementation can top two of the hardest agent benchmarks without a new base model doing the heavy lifting. The

Prime Intellect released Prime Agent as an open-source MIT-licensed agent built on a Recursive Language Model and a persistent Python REPL.

Josh Miller, CEO of The Browser Company, posted a "hot take" on X that struck a nerve in Silicon Valley: "isn't it kinda crazy that nobody is really using AI Agents?" He wasn't talking about software engineers or AI early adopters, but about "college friends talking about it in group chat," the kind of mass-consumer moment that Instagram and TikTok both got and that AI agents, so far, have not.

Cory Doctorow argues the real AI market is not chatbots but 'decision support' systems for medical, bail, resume, and loan calls.