Humans on AI, February 26th 2026

• Anthropic employee equity sale

• Anthropic says internal models lag Opus

• LRM training fails no-CoT gains

• Anthropic/DOD confrontation still unresolved

• Anthropic RSP drops unilateral pause

• Anthropic warns about massive Chinese distillation attack

Opinion: It’s new moderate evidence for tool world that Anthropic’s targeting distinct white-collar niches and not only focussing on AGI/RSI-through-coding. It’s bad for tool world that Anthropic’s currently doing it with plugins and not spinoff models, and bad for tool world that there aren’t niche LRM companies so far[^1], but these aren’t really updates.

Neolab Standard Intelligence combined a long-context video model with RL on computer-use to make a prototype “General Computer Action Model”.

Opinion: Project is anti-tool-world, trying to make an agent that can learn any computer interface from video only. Standard Intelligence are tiny and their model’s just a research demo, but their techniques are impressive and getting attention. Likely accelerates automation of white-collar work by a month. New weak evidence against tool world.

Anthropic bought AI computer-use company Vercept. Opinion: Vercept’s approach to computer-use AI isn’t as AGI-forward as Standard Intelligence’s, but Anthropic prioritizing computer-use is bad news. New weak evidence against tool world.

An OpenAI researcher confirms that the model powering Codex is trained in-harness.

Opinion: Likely extends to Claude Code, Claude Cowork, DeepThink, and GDM’s private math harness Alethia. Recent work from METR Evals questions the impact of coding harnesses, but in-harness training is still new weak evidence for tool world.

An independent researcher estimates frontier model sizes based on no-CoT performance on a middle school math benchmark (GSM8K). Opus 4.x models are at least double the size of GPT 5.2, and maybe quintuple the size. Sonnet 4.x models are the same size as GPT 5.2. Opinion: Estimates pass our sanity checks[^2]. We think the data and the estimates together give new weak evidence for tool world and deserve more research:

Opinion: PR move against Chinese models like MiniMax M2.5 that score high on SWE-bench Verified for cheap, but the contamination study is solid. New weak evidence that Chinese models aren’t near-SOTA.

A team of leading ML/cog-sci profs published Agents of Chaos, a study of security vulnerabilities in AI agents (OpenClaw). Opinion: We all already know that OpenClaw is insecure, but it’s a method-building paper. The procedure is very useful for red-teaming future agents.

METR’s follow-up to the 2025 ‘Uplift’ study where AI slowed expert coders down was scrapped and redesigned due to design flaws. Despite the flaws, the scrapped study’s data do demonstrate that AI now causes a speedup (of unknown size) rather than a slowdown. Opinion: METR have some genius senior staff but public-facing work might be getting offloaded to junior staff lately, a lot of messups.