vLLM adds expert RAM offloading for local frontier models

A new White House task force will produce a report assessing risks posed by artificial intelligence, per the Wall Street Journal. The report will also address what responsibility the US federal government should have over the technology.

Willison argues pay-by-usage services need hard caps that cut off usage and return errors after a set monthly spend, not soft warning emails. He notes AWS launched spending limits a few weeks ago, and says AWS is the service he most wants this from.

The release fixes Bash deny/ask rules being skipped under sandbox auto-allow when a bare variable assignment precedes the command, and stops user plugins from rewriting org-managed MCP server sign-in tool descriptions. Agents gain agent.spawn for shared teammates, with unified agent IDs and clarified idle/waiting states.

A developer hosted a private WoW server and built a browser client playable on PC or mobile for free at jankcraft.xyz, then added a custom MCP and agent harness to drive it. Qwen was used to play the game via the client.

A Hugging Face blog post titled "The Agent Said It Was Done. The Database Disagreed." examines the gap between an AI agent's self-reported task completion and actual database state. No further details were available in the source snippet.

LFM2.5-Encoder-350M is a multilingual bidirectional encoder built on the LFM2 architecture, described as the larger variant for maximum downstream quality. It uses full bidirectional attention as a masked language model, designed for fine-tuning into task-specific models such as classification.

Nebius Token Factory's Dylan Bristot and developer advocate Sujee Maniyam walk through the trade-offs of running open LLMs in production, noting open models have nearly caught up with proprietary ones while fast, cheap serving remains the hard part.