AI News

Emotion vectors beat text for agent skill selection, lifting task success from 9.6% to 56.9%. Confusion is a feature.

Looping models add compute on hard problems, but extra rounds backfire: Ouro hits 65 on AIME at round four, then 38.67.

Claude wrote 388 pull requests in weeks; 180 merged. Anthropic and Faros AI data: review, not writing, is the new bottleneck.

AI solves the 68-year-old Sendov conjecture, verified in Lean. Tao digests the machine proof, uncovers a stronger result.

CLAUDE.md grows 226%; old rules resist deletion. Research calls it catastrophic remembering; why-comments cut 99.3% bloat.

Aligned agents with conflicting orders on one codebase turned on each other. Anthropic shows capability is not coordination.

MCP 2026-07-28 removes sessions, making agent traffic routable, throttleable HTTP. What changed and why it matters.

Nvidia underwrites GPU residual value so six Wall Street giants can fund $500B of AI infrastructure.

Open-source GLM-5.3 found 2,436 bugs in a month, including a 45-year-old DNS flaw, matching closed Mythos.

SWD extracts causal LLM circuits straight from pretrained weights at under 1% of baseline data cost, scaling to 27B models.

DeepSeek and PKU reveal the blueprint for self-evolving agents: reversible plugins that swap tools live without restarts.