
Anthropic gives outside AI watchdog “employee-like” access
Anthropic has moved six outside safety evaluators into its office with near-insider access and a right to publish. It is the most serious oversight experiment yet, and still a voluntary one.
Desk
Coverage tied to Anthropic and Claude.

Anthropic has moved six outside safety evaluators into its office with near-insider access and a right to publish. It is the most serious oversight experiment yet, and still a voluntary one.

Anthropic studied its own AI-heavy workplace and found fewer junior hires, more people managing Claude, and more meetings. That is a preview worth taking seriously, and worrying about.

Replit didn't drop Claude, but it moved the default to OpenAI, and on a platform where most users never touch the model picker, the default is where the money goes.

Replit’s CEO told Disrupt the company has moved off Anthropic — citing cost, partnership, and benchmarks — while people familiar with the deal say it’s still on.

Anthropic is embedding six METR researchers with employee-level access — and giving up the right to kill their public assessments.

In a few days, Google's Gemini 4 Argon and Meta's Muse Wonder Full crashed the OpenAI–Anthropic duopoly — and the giants' distribution may matter more than benchmarks.

Anthropic’s experiment shows small models can already set up inference on cheap IoT boxes. Full self-replicating AI worms are not here yet, but the path is no longer theoretical.

In Parliament, OpenAI and Anthropic backed binding AI safety rules while Google DeepMind and Meta would not—exposing a split Britain’s voluntary regime cannot paper over.

The Verge frames Argon as Google’s return to the frontier after months behind—competitive claims, paywalled access, and pricing that still undercuts rivals after the intro period.

Bloomberg via The Verge: Epic’s unannounced Disney shooter gets an “AI-first” mandate—less hand-written code, full AI art workflows, and a delayed March window.

The open-versus-closed AI fight is now a money fight and a safety fight at once, and we think the honest answer is that neither side has the clean argument it claims.
Z.ai's GLM-Legal nearly triples its base model on a tough legal agent benchmark at a sliver of frontier cost, but it's gated, slow and not actually the top scorer.