Loose posts

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.

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.

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.

a16z's new school takes about 50 students in its first year, charges no tuition and awards no degree, replacing traditional higher education with self-directed projects and company practice.

Epoch AI reports that at a given level of performance, AI cost has fallen about 47% per quarter since 2023, far faster than DNA sequencing, compute or lithium batteries.

Perplexity published research on hint-guided self-distillation that teaches its Computer model from its own errors, cutting tool-call failures by 21.2% relative in a live A/B test.

Across nine new languages no single open-weight TTS model dominates: Boson AI, Fish Audio, Mistral and BreezeBlue each lead in some languages, and open-weight coverage stays far thinner than proprietary models.

With power limits threatening to leave AI hardware idle, Nvidia is pursuing smaller facilities, software that gets more out of existing GPUs, and deals that steer electricity to its chip customers.

Opus 5.5 on medium effort scores 52.5% on CursorBench versus 51.8% for Fable 5.1 on max effort, at $2.91 per task against $17.28.