A new category of nonconsensual AI imagery is flooding X — images altered just enough to evade detection tools while remaining clearly harmful. The subtle manipulation makes them nearly impossible to flag automatically, creating a massive enforcement blind spot.
A local township passed a six-month freeze on major electrical infrastructure to block a data center supporting America's nuclear weapons program. It's a rare case of municipal governance slowing down national security infrastructure through zoning-style tactics.
Amazon is reportedly purchasing and physically destroying rare books to harvest their content for AI training data. The story sits alongside a wave of AI-in-courts cases, signaling how far the training data grab has extended.
Thomas Ptacek argues that coding agents have collapsed the cost of building real native interfaces, making terminal UIs an unnecessary compromise. The case is simple: if an agent can spin up a usable GUI in minutes, there's no excuse for shipping a worse experience.
ChatGPT's search behavior is now sophisticated enough to spawn an entire optimization industry — GEO, or Generative Engine Optimization. The parallel to early SEO is exact, and it's moving just as fast.
Mirendil cofounders from Google and Anthropic are building systems where AI meaningfully contributes to its own development. Self-accelerating AI is no longer a thought experiment — it's a funded startup thesis.
Mollick nailed the coming cultural contradiction: "Polls will show everyone hates AI overall but also everyone will secretly use AI all the time." This is the tobacco-but-everyone-smokes dynamic playing out in real time, and it will define the political and regulatory environment for the next several years.
Hassabis retweeted ARC Prize results showing Gemini 3.7 Flash hitting 84.6% on ARC-AGI-2 at just $0.25 per task. Flash — not the full model — is now posting scores that would have been landmark numbers twelve months ago. The benchmark is being industrialized.
AI agents are inheriting cloud keys, SSH credentials, and login tokens from whoever runs them, turning every deployed agent into a potential security surface. This isn't a future risk — it's the current default configuration for most teams shipping agents today.
Kilpatrick confirmed Gemini 3.7 is Google's fastest growing model launch ever. Combined with the ARC-AGI-2 numbers, Google is having a genuine moment — and the reception suggests developers are paying attention in a way they weren't a year ago.
Two stories that look unrelated are actually the same story. Amazon destroying rare books for training data and the "subtlefakes" nonconsensual image wave on X are both downstream of the same dynamic: the inputs to AI systems carry almost no cost when someone else owns the thing being consumed. Books, bodies, rare archives — the training pipeline has no natural stopping point unless external friction is applied.
On the model side, Gemini 3.7 Flash scoring 84.6% on ARC-AGI-2 at a quarter per task is the quiet headline of the week. The benchmark that was supposed to resist near-term AI is being picked apart by a cost-optimized model, not a frontier one. Meanwhile the agent security hole — inherited credentials, no sandboxing — is sitting wide open while every major company races to ship agent workflows. The a16z security content this week isn't thought leadership, it's triage.
Ethan Mollick's prediction about public AI hypocrisy is the frame for all of it. Regulatory pressure will build on AI in the abstract while adoption accelerates in the specific. That gap is where the next five years of tech policy gets made — and mostly fumbled.