Trump enters the AI pacing fight

Following the labs’ weekend pacing pledge, President Trump called fears of an AI takeover a “hoax” and called into Jensen Huang’s live interview. By Tuesday, Democrats and some Republicans in Congress were pressing for safeguards; China’s Foreign Ministry also criticized slowing development.

Congress begins pushing back

The Guardian

Democrats and some Republicans are calling for federal safeguards after Trump dismissed the industry’s warnings. Within a day, the labs’ voluntary proposal had become a congressional issue.

Independent evaluation has a trust problem

Axios

Frontier labs agree on more oversight, but competitors and policymakers question whether proposed evaluators are independent enough from the companies they would assess. Axios points to financial ties, commercial incentives, and the mixed record of technology self-regulation.

Stuart Russell argues for concrete safety requirements

The Guardian

Russell argues that changing the development schedule can’t substitute for safety requirements with concrete goals. His proposal shifts attention from the duration of a slowdown to the conditions a model must meet.

A DeepMind safety researcher resigns

Techmeme / Bloomberg

Bilal Chughtai resigned from Google DeepMind’s safety and alignment team and warned that humanity is running out of time to prevent severe AI harm. His departure adds another internal dissent case to the governance debate.

Agent configuration moves into files

AI Engineer

This talk describes agents defined through instruction and skill files, with repositories mounted into isolated sandboxes and credentials injected through a restricted proxy. It also reports that Cursor replaced about 12,000 lines of TypeScript orchestration with roughly 200 lines of agent files.

Vercel rebuilds its data agent around the filesystem

AI Engineer

Vercel replaced prescriptive tools with a stateful agent that explores a sandbox through basic file and shell operations. The company says the rewrite doubled evaluation scores and became the basis for its Eve framework.

PostHog keeps the security decision deterministic

AI Engineer

PostHog’s Warlock scanner uses deterministic YARA rules to block malicious inputs before its coding agent acts. A large language model helps triage unblocked findings, but it never decides whether execution is allowed.

ActGuard audits actions before execution

arXiv

ActGuard predicts which tools a reasonable next action would use in the local context, then checks the proposed tool and parameters for deviations. It masks only evidence a verifier identifies as malicious and regenerates the action from the remaining context.

Restate gives long-running agents durable state

AI Engineer

Restate journals agent events so a process can recover at the point of failure after downtime. Its model also supports serialized session state, communication between agents, and human approval as a durable suspension point.

Vera Rubin reaches up to seven times the throughput per megawatt

Techmeme / SemiAnalysis

SemiAnalysis reports that Vera Rubin NVL72 delivered up to seven times Blackwell’s token throughput per megawatt on a 1.6 trillion parameter DeepSeek model. That result exceeds Nvidia’s public claim of three times for models between 1 trillion and 3 trillion parameters.

Euclyd raises more than €200 million for inference chips

Techmeme / CNBC

The Dutch startup’s Series A was co-led by Samsung and three investment groups. The round gives a European Nvidia alternative substantial capital before it has the scale of an incumbent supplier.

Lagarde treats model access as trade leverage

The Guardian

European Central Bank president Christine Lagarde says Europe needs models that are “good enough” for most tasks and can run in domestic data centers. She argues that local capacity reduces the leverage the United States or China could gain by cutting access.

Google gives every engineer access to Claude Opus 5

Techmeme / Business Insider

Google is making Claude Opus 5 available to all engineers through Antigravity while maintaining that Gemini is its primary internal development model. The decision adds a competitor’s coding model to Google’s standard engineering environment.