AI #174: You're It

Abstraction: Weekly AI roundup, Claude Tag, medical scanners, agent security

Key points:

• Claude Tag launched: tag @Claude into a Slack channel and it spins up a sandboxed instance per thread with its own memory/permissions to clone repos, write and test code; 65% of Anthropic's product-team code from internal version; Karpathy calls it the "3rd major redesign of LLM UIUX" (org-level asynchronous entity).
• MidJourney scanner debate: Zvi backs the bull case (time-series imaging, "bitter lesson pill for healthcare") against doctors' "more information is bad / false positives" prior; Amanda Askell notes the issue is response norms, not scans; EchoNext ECG biomarker caught severe heart damage (Nature Medicine).
• Cyber security alarm: Five Eyes warn "months not years" to secure systems from Mythos-class threats; coding agents treated as trusted-by-default with internet access, secret data, running autonomously for hours with no audit.
• AI-writing detection: Zvi endorses Pangram, argues "witch hunts" are fully scientific; multiple AI-generated pieces won literary prizes, indicting prize-judging processes.
• Dean Ball joins OpenAI to work on policy (Zvi: big upgrade over alternatives); Alex Bores lost NY-12 by ~4%; GLM-5.2 is new best open model, claimed niche "open top-level agent."
• Two-pills framing (precursor to #175's three pills): AGI-pilled vs ASI-pilled; Jeff Bezos predicts AI creates a labor shortage via "dream-build loop."

Connections: Zvi Mowshowitz · Anthropic · Openai · Claude · AI Agents · AI In Medicine · Cybersecurity

Source: https://thezvi.substack.com/p/ai-174-youre-it