Simbe completes rollout of 3,000 Tally robots globally

A custom Strata fork runs the IQ3_S quant of Qwen3.8-Flash-Next-GSQ-RCO-Abliterated at 20-30 tok/sec decode and 300 to ~90k tok/sec prefill at 131k context on 12GB VRAM plus 32GB RAM. The Q2 quant reaches 39-45 tok/sec decode.

Meta researchers argue that giving AI agents more compute at inference time is not enough, and propose a different approach in a new paper covered by VentureBeat.

Cohere's guide frames the choice around request profile, traffic pattern, and utilization rather than headline price. It contrasts bulk indexing jobs (~100 descriptions per request, ~200 tokens each) with single short search queries where latency matters.

Essay contends AI dissolves the human intellectual communities that make private intellectual activity possible, citing software engineering's shift from compilers and type systems to "prompts," "harnesses," and "loops." It argues prompting is a shallower skill than software engineering, so human capital formation has collapsed.

Anthropic's Claude Code Projects runs each task as its own thread in parallel within one ongoing conversation. The feature is in beta and rolling out, with support for steering threads, reviewing pull requests, creating artifacts, and scheduling routines.

Claude Code v2.1.296 adds a code key to the Claude apps gateway's managed.policies[], applying cli settings in Claude Desktop's Code tab and enabling gateway mode. Subagent frontmatter gains autoCompactWindow so subagents can compact earlier than the main conversation.