Two federal proposals target frontier AI access and safety

CNBC reports that the White House is considering Gold Eagle, a system for controlling access to frontier-model releases. A separate report describes a FINRA-style AI-safety regulator linked to the SEC; both remain proposals rather than official policy.

A FINRA-style regulator enters the debate

Techmeme

A reported proposal would put AI-safety oversight in a new federal body linked to the SEC. Its authority and review process remain open questions.

Maggie Eastland examines the regulator proposal

Maggie Eastland on X

Eastland discusses the proposed FINRA-style structure and the institutional choices it would place around model oversight.

Miles Brundage questions policy consistency

Miles Brundage on X

Brundage places the proposals against existing AI-policy positions and points to tensions in how federal intervention is being argued.

Anthropic may rent up to $10 billion of Meta compute

CNBC

CNBC reports that Anthropic and Meta are discussing a multiyear compute arrangement worth as much as $10 billion. The talks would add another external source of capacity for Anthropic.

Japan plans a 27,500-chip Rubin build

Techmeme

According to the report, a state-backed Japanese effort involving Noetra, SoftBank, Sony, and NEC plans to secure 27,500 Rubin chips for a domestic robotics model.

Apple's complaint reaches OpenAI's hardware plans

TechCrunch

Following last week's first reports on Apple's suit, TechCrunch considers how the allegations could affect OpenAI's device ambitions. Any effect on a possible IPO remains speculative.

Deployment controls move closer to agent behavior

AI Engineer

This engineering talk covers deployment controls that can stop harmful behavior, constrain the affected cohort, and support replay during an agent rollout.

Checkpointing preserves agent state

AI Engineer

A second engineering talk focuses on checkpointing and state management so long-running agents can recover without restarting their work.

Proof-or-Stop gates completion on evidence

arXiv

The paper proposes an evidence-gated control layer for autonomous-agent lifecycle states. Completion depends on verifiable proof rather than the agent's own declaration.

Open-weight models narrow a reported cyber gap

Techmeme

New benchmark reporting says leading open-weight models have narrowed the cyber-capability gap with closed models. The result measures benchmark performance, not observed attacks or every open model.

Ethan Mollick connects cyber results to release policy

Ethan Mollick on X

Mollick connects the model results to the policy debate over open releases and competition between the United States, the United Kingdom, and China.