Four AI labs back a slower frontier pace

Between 10 a.m. and 7 p.m. ET on Saturday, Dario Amodei, Elon Musk, Sam Altman, and Demis Hassabis endorsed a slower pace for frontier-model development. The first checkable commitment is independent evaluators with employee-like access inside Anthropic and OpenAI.

Amodei proposes a three-stage pacing system

Dario Amodei via Techmeme

Amodei's proposal starts with embedded evaluators, then calls for coordination among democracies and with authoritarian governments. Anthropic says it will implement the evaluator step on its own.

OpenAI commits to employee-like evaluator access

Sam Altman via Techmeme

Altman says OpenAI will adopt the same evaluator model and provide more details soon. That turns a general endorsement into a commitment that can be checked against access terms and named evaluators.

“Committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”

Hugging Face asks to join the evaluator program

Clem Delangue via Techmeme

Hugging Face launched an Open Alignment Initiative led by co-founder Thomas Wolf and asked to participate in Anthropic's embedded-evaluator program. The pledge now has a volunteer, but no disclosed selection or access process.

Susan Zhang questions evaluator independence

Susan Zhang on X

Zhang asks whether the industry can establish evaluation that is both independent and technically competent. The answer will determine whether employee-like access produces external scrutiny or a lab-selected review process.

The FRONTIER Act would put audits in statute

Rep. Lori Trahan on X

Trahan points to the bipartisan FRONTIER Act as a statutory route for independent audits. Its requirements would be set in federal law rather than left to voluntary agreements between labs and evaluators.

David Sacks challenges the labs to slow down first

David Sacks via Techmeme

Sacks argues that Anthropic and OpenAI can pace their own development for business reasons without waiting for a preferred regulatory system. Voluntary action is now a direct test of whether the labs' public commitments change their release behavior.

OpenAI takes a 2026 IPO off the calendar

Axios

Altman says OpenAI won't go public this year because it has safety and alignment work to complete. The decision postpones one expected listing on the same day the labs endorsed a slower development pace.

“Right now would be an ill-advised moment to go public,”

Xi proposes AI cooperation across BRICS

CNBC

Xi Jinping says China will help expand artificial-intelligence collaboration among developing economies. Amodei's international pacing proposal depends on coordination beyond democratic countries, while China's stated priority is broader development cooperation.

Real-SWE tests agents on private enterprise code

Specific

Real-SWE evaluates model-and-harness combinations on licensed private production codebases. Private tasks reduce the chance that public benchmark solutions appeared in training data, while testing work tied to existing products and business rules.

Amp adds a free bring-your-own-key option

Quinn Slack on X

Amp now offers free use when developers bring their own model-provider keys. The change separates the coding-agent interface from bundled inference pricing.

AgentsDock builds an IDE for agent research

AgentsDock

AgentsDock is a dedicated development environment for agentic AI research. It packages agent experimentation into an IDE instead of treating it as an extension of a general-purpose coding workspace.

Woof 1.1 targets on-device agents

Sigil Wen on X

Underdog's Woof 1.1 is a four-billion-parameter model intended for an on-device AI operating system and uses less than 2.5 GB. The release pairs a small local model with a defined agent environment rather than a standalone chat interface.