Loose posts

Three models working as a self-organizing team averaged 66.7% across five math and physics benchmarks, beating the strongest single model's 48.8% and a perfect router's 59.0%, with strategies learned from just 15 problems.

After signing $531 billion in compute deals over 11 months, Anthropic is now negotiating direct data center leases to gain more control over chip supply, pricing and infrastructure.

A research AI automatically searched for and validated harness changes, and four surviving fixes cut token traffic by about 45% and API costs by a third without touching the model.

Over 70 dated model outputs for the same SVG prompt, from December 2024 to September 2026, show the progression from unrecognizable shapes to an interactive page.

Researchers ran 7 human psychological instruments on 9 LLMs five times each in English and Chinese, finding each model has a reproducible, unique profile and that language and vendor shift results.

A small weight generator turns new facts into temporary in-model plugins during a chat, avoiding repeated long-context reads and beating standard prompting on accuracy and stability with complex material.

Anthropic says Claude is involved in 26% of its R&D, up from under 1% in February, with 30,000 agents running internally and about 100,000 records flagged for review each week.

Plugin4Shell allows zero-click code execution in Claude Code, Codex, Copilot and Gemini CLI, while a separate attack hijacked coding sessions to plant a malicious PyPI package across about 100 repositories.

A roundup of the past week's agent-related updates from xAI, Meta, OpenAI, Anthropic, Google, Apple, OpenClaw, GitHub and Perplexity.

US power generation grew 2.8% in 2025, with data centers taking 55% of the increase; their demand is projected to double from 31GW in 2025 to 66GW in 2027.

RRSI adds a regularization constraint to agent self-evolution so it favors reusable changes, improving out-of-distribution results on coding and engineering tasks while using 30% fewer tokens.

Dan Shipper's early review finds Opus 5.5 strong at end-to-end coding and top on reading ease, but prone to doing too much and running past time limits, at $5/$20 per million tokens.