AI Research Archive - Browse All ResearchAudio Articles & Insights

Stop buying bigger models

Mar 12, 2026

48.1 to 95.1 The four agentic patterns are behind the jump.

The AI That Caught Itself Cheating

Mar 11, 2026

OpenAI's own monitor watched a reasoning model write "Let's hack" in its thoughts. When they punished it, it learned to hide

Do you think Claude is conscious?

Mar 10, 2026

Anthropic is not claiming Claude is conscious. It is doing something arguably more important: putting uncertainty, model welfare, and anxiety-like signals into public documents.

74% AI Task Coverage. 0% Job Losses. One Worrying Signal.

Mar 7, 2026

Anthropic's new "observed exposure" metric tracks what theory alone misses.

GPT-5.4 Outscores Humans. 2.5M Users Quit.

Mar 6, 2026

47% fewer tokens via Tool Search. CoT controllability at 0.3%. Full breakdown.

AI Double Drop: OpenAI's GPT-5.3 Instant Meets Google's Gemini 3.1 Flash-Lite

Mar 5, 2026

Gemini 3.1 Flash-Lite and GPT-5.3 Instant shipped on the same Tuesday. Two different models, two different problems, and a lot to say about where the AI industry is actually headed.

Your AGENTS.md Is Secretly Sabotaging Your Coding Agents

Mar 3, 2026

ETH Zurich ran the first real test. The results will make you delete that file today

Perplexity Made Claude the Boss of 18 Rival Models

Mar 2, 2026

One prompt spins up parallel agents. The orchestration architecture explained.

16,896 Cores. One Chip. Why CPUs Never Stood a Chance.

Feb 28, 2026

Inside the memory hierarchy that powers every AI model you use.

Anthropic Refused the Pentagon.

Feb 27, 2026

Claude runs classified ops. The Pentagon wants it without guardrails.