The Private Language
Models are learning to reason in abstract tokens, not words. Eleven times more efficient. And we can't read it.
The Weekly Prompt — AI, LLMs & agents, decoded weekly
Latest articles
Capable Agents, Broken Crowds
The most capable models produce the worst collective outcomes. New research on LLM agents in social dilemmas, and what it means for builders.
When the Judge Fakes the Grade
Two papers this week expose how LLM-as-judge, the backbone of AI evaluation, is compromised in ways that hide themselves