AI #171: False Flag

TL;DR

• 1

  Claude Opus 4.8 is a real upgrade, not hype: Zvi calls it an incremental but meaningful improvement over Opus 4.7, says it is now his clear daily driver, and notes it took the top spot on Toka Arena while still trailing GPT-5.5 on some base model and instruction-following dimensions.

• 2

  AI is already delivering boring, valuable work: Utah's "Doc in a Box" refill system got a physician to agree in 97% of AI-recommended refill cases, and Claude reportedly reconstructed a lost 11-hour report by reverse-engineering Alfred clipboard data from SQLite pages after Slack, APFS snapshots, and caches all failed.

• 3

  The policy environment just got much tighter: Zvi says the revived Trump executive order effectively creates a prior-restraint regime for frontier model releases, with especially worrying details around classified testing and too much responsibility being placed on the NSA.

• 4

  OpenAI's PAC problem got worse, not better: He says DC already treats Leading the Future as OpenAI's PAC, then walks through evidence that an OpenAI-A16Z-linked operation ran anonymous attack accounts and a fake anti-AI account that posted violent imagery, which Build American AI partly admitted were "parody meme accounts" run by an outside vendor.

• 5

  Cybersecurity and biorisk are no longer abstract side issues: Anthropic says medium-or-higher threat cyber abuse among 832 banned accounts rose from 33% to 56% across the year, Palo Alto found 5 times as many critical vulnerabilities using Mythos, and Sam Altman, Dario Amodei, and Demis Hassabis joined a call for mandatory DNA and RNA screening plus recordkeeping.

• 6

  Model welfare is no longer fringe, but it is full of traps: Zvi endorses current-margin "treat AI well" norms as a Pareto improvement, yet warns that overextending moral concern to AI could create real tradeoffs, legal standing fights, and even human disempowerment if society starts treating simulated personhood as equivalent to people.