Federal DeepSeek order reaches third-party services

DHS has ordered federal civilian agencies to remove DeepSeek products, citing national-security concerns. The order also covers third-party services built using DeepSeek's technology.

DeepSeek opens its production inference stack

DeepSeek · GitHub

DeepInfer releases the serving software behind DeepSeek's production models. The code targets faster inference, better chip utilization, and lower compute and energy costs.

vLLM adds push-based key-value cache transfer

AWS

AWS explains vLLM's push-based key-value cache transfer for deployments that separate prompt processing from token generation. The change targets both time to first token and time per output token.

Mercury 3 makes a diffusion-based speed claim

Inception Labs

Inception claims its diffusion-based language model generates more than 1,000 tokens per second. It also reports beating GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks.

OpenAI reports gains in legacy programming languages

OpenAI

OpenAI reports stronger coding and migration performance in COBOL, Fortran, and Visual Basic. These languages remain in use at banks and governments, making modernization a concrete application for the reported gains.

Leaked documents describe self-propagating Army agents

The Intercept via Techmeme

The Intercept reports that ECHOTRIBBLE, a nascent Army Cyber Command program, seeks agents that autonomously spread across networks to find and eliminate targets.

Microsoft expands enterprise agent containment tooling

Indian Express

Microsoft is expanding Project Perception to monitor and contain rogue agents inside enterprises, according to the Indian Express.

Shared Meta Muse chats expose personal and business details

Ars Technica · Dan Goodin

Ars reports that a single query can find thousands of shared Muse conversations listed by title, including chats containing personal and business secrets. Sharing a transcript can make sensitive content discoverable beyond its intended audience.

Harnesses can differ on cost even when eval scores converge

AI Engineer · 18 minutes

This talk separates runtime state and execution from context and tool orchestration. It argues that cost per task can differ substantially across harnesses even when coding evaluation scores converge.

A deployment sequence for policy-driven agents

AI Engineer · Evan Cormos · 17 minutes

Coinbase's Evan Cormos describes a progression from manual verification to sampling, then routine automation with human exception handling. He ties the progression to policy thresholds, repeated evaluations, and trace logs.

Goldman projects power demand growth against long grid waits

Axios · Madison Mills

A Goldman Sachs report shared with Axios projects a 170% rise in global data-center power demand from 2025 to 2030. It puts the median grid-connection wait at roughly five years. Local opposition accounts for 76 of the 86 canceled or significantly delayed projects since January, according to the report.

Nscale and Anthropic reportedly discuss another $45 billion deal

TechCrunch

Nscale and Anthropic are reportedly discussing a second compute deal worth about $45 billion, following a similarly sized agreement announced a little over a month ago. The talks come as both companies pursue IPOs.