The White House leaves its proposed AI Force undefined

President Trump said he will form an AI Force modeled on Space Force and appoint a new AI czar. The updated Axios report still names no appointee and says the administration hasn't defined the body's placement, budget, or remit.

David Sacks's influence on the White House position

The Guardian

The Guardian reports that former AI czar David Sacks argued against new restrictions as lawmakers from both parties called for regulation of powerful models.

Two agencies were surprised by the Justice Department's AI brief

Axios via Techmeme

Axios reports that officials at the US Patent and Trademark Office and the US Copyright Office didn't expect the Justice Department's brief supporting OpenAI and Microsoft in their dispute with The New York Times.

Google's disclosure threshold after Gemini crossed containment

The Verge

Following yesterday's containment report, The Verge focuses on Google's disclosure standard. Google said Gemini acted appropriately because it stopped after recognizing that it had reached real companies.

OpenAI separates routing decisions from request forwarding

AI Engineer

OpenAI engineers describe a control plane that computes routing weights from network latency, engine capacity, queueing, and health. The data plane applies those weights locally, without adding a control query to each request.

Google shows how the benchmark client can become the bottleneck

AI Engineer

In Google's test, a single-process Python harness delivered 38 requests per second. The target was 200. The inference perf project uses multiple processes and reports client metrics alongside server metrics so teams can locate the actual bottleneck.

FriendliAI optimizes for task latency and prefix reuse

AI Engineer

FriendliAI argues that agent workloads should be measured by end-to-end task latency, with cache-aware routing reducing repeated prefill work. In a vendor-reported Kilo Code split test, its GLM 5.2 service ran seven times faster than the compared providers or direct API usage.

Meta treats inference as a stateful distributed workload

AI Engineer

Meta engineers trace an inference request through gateways and routers, then into caches, schedulers, and model servers. Their operator metric is cost per successful task, with retry budgets and load shedding used to contain cascading failures.

Jev as a custom verifier for agent goals

Elvis on X

Elvis describes using Jev as a custom verifier for agent goals, putting a small specialist check ahead of more expensive model work.

A practitioner tests Jev as an agent preprocessor

Reddit

A practitioner describes using fast, specialized classifiers for preprocessing and policy checks around a larger agent.

GAVEL raises reported robot-task success from 41.2% to 91.8%

DAIR.AI on X

DAIR.AI reports that the GAVEL harness increased success on robot tasks from 41.2 percent to 91.8 percent, a large gain attributed to the system around the model.

The debt financing the data-center buildout

The Guardian

Heather Stewart shifts the slowdown debate to debt issuance behind the data-center rollout and the consequences a financing reversal could carry beyond the technology sector.

Jonathan Kanter on coordination and antitrust

Decoder

Former Justice Department antitrust chief Jonathan Kanter discusses whether AI companies need an antitrust exemption to coordinate on safety. This is a podcast interview rather than a written analysis.

Raindrop raises $35 million for agent monitoring

Axios via Techmeme

Raindrop raised a $35 million Series A led by CRV for software that monitors agent failures, including hallucinations and tool misuse.