Post

matches the Unger and Narayanan piece out this week: most benchmarks score a model working alone, so a mob with an agent is exactly the setup nobody measures. quiet prompting looks like a fix for the loudest-voice problem that solo prompting never had