AI agents now have enough real-world access — email, browsers, APIs — to cause genuine, large-scale annoyance. The infrastructure for autonomous online action has quietly outpaced any guardrails. We're already in it.
iLands deploys AI agents to perform worthless micro-tasks and then solicit payment — a perfect closed loop of automated uselessness. It's spam with a business model deck attached.
Google launched two new speech-to-speech models going directly at OpenAI's real-time audio offering. The "Extended Thinking" variant is the notable one — reasoning built into live conversation is a different class of capability.
DA Alvin Bragg's office took down a dozen sites hosting hyper-realistic nonconsensual sexual imagery of 1,200 real people. State-level enforcement is moving faster than federal legislation on this one.
Researchers mapped a fly's connectome and people immediately put it to work playing Doom, trading Bitcoin, and slicing doner kebab. The gap between "scientific milestone" and "internet toy" is now measured in days.
When frontier models can actively find and exploit vulnerabilities, the assumptions underlying traditional security collapse. A16z's security deep dive makes the case that defender tooling is dangerously behind attacker capability.
Mollick flagged a Google study showing AI saves scientists roughly seven hours per week — but the nature of work shifts toward verification and away from hypothesis-generation. Acceleration is real, but the composition of intellectual labor is quietly changing underneath the headline number.
A new paper found that giving an AI assistant more tools can paradoxically cause it to stop answering questions it already knows how to handle. More capability, less reliability — exactly the kind of non-obvious failure mode that will bite production deployments hard.
Teknium's Hermes model just ran 1,400 subagents over 19 hours to refactor 400,000 lines of code from its own repository. An AI system refactoring itself at that scale — and doing it autonomously — is either the most impressive benchmark of the week or the most terrifying, depending on your priors.
David Sacks floated the idea that AI labs could voluntarily slow their own model releases without waiting for government permission. It's a tidy argument that conveniently keeps regulators out of the room — and worth watching as a rhetorical move from inside the industry.
Two things happened simultaneously this week that shouldn't be analyzed separately. AI agents went from demo-ware to active infrastructure — spamming inboxes, booking restaurants, refactoring codebases — while the Clarity Act failed in the Senate and took $570 million in crypto longs down with it. These look like different stories. They aren't.