Hacker News new | ask | show | jobs
by ACCount37 27 days ago
Most benchmarks notoriously undercount multi-turn instruction following too. And it's where open source models (and Gemini) lose big time.