AI #174: You're It

TL;DR

• 1

  Claude Tag is a real workflow shift, not a Slack bot: Anthropic says 65% of its product team's code is now created by its internal version of Claude Tag, which spins up per-thread sandboxes, learns channel context, and works asynchronously across Slack.

• 2

  GLM-5.2 looks like the new top open model, with caveats: Zvi calls it the best open model right now for private or fully local agent use, but says it is expensive for its class and not obviously the right fit for most deployments.

• 3

  AI-written text is getting caught in public with almost comic obviousness: Pangram flagged a Jacob Helberg op-ed and a Harper's Bazaar UK short story winner as fully AI-generated, and Zvi's point is not that detectors are perfect, but that many institutions are failing at basic quality control.

• 4

  Security norms around coding agents are already absurdly lax: Marius Hobbhahn's framing lands hard here. Organizations are giving agents internet access, secret internal context, and long autonomous runtimes, then often not checking for malicious behavior afterward.

• 5

  The Midjourney scanner fight is really a fight over whether more medical data is good: Zvi sides with the bull case that cheap, safe, repeated scans and time-series diffs could be highly valuable, even if the current healthcare system is bad at handling incidental findings.

• 6

  Dean Ball joining OpenAI is framed as a serious policy upgrade: Zvi says OpenAI's new Strategic Futures team could improve the company's thinking on catastrophic risk, labor, and government relations, though he worries Ball may still be too AGI-pilled and not ASI-pilled enough.