OpenAI says major factual errors down 65% and sycophancy down 83% since March
charliermarsh · x · 2026-09-10
OpenAI researcher aidanmclau recapped quality gains since March: responses with major factual errors are down 65%, instant-tier gpt-5.6 now outperforms o3 at high reasoning effort, and overall sycophancy is down 83%.
The post quotes OpenAI exec michpokrass on serving 1B+ weekly ChatGPT users, most on the free tier: the mission is to distribute AGI's benefits as widely as safely feasible, and the past 6 months focused on bringing maximum capability to both free and paying users.
Related event: OpenAI Says Major Factual Errors Down 65%, Sycophancy Down 83%(2 posts)→
More from Companies & People
- Safety researcher joins OpenAI's Safety and Security Committee, stresses externally verifiable oversight — JacquesThibs · 2026-09-10
- Waymo co-CEO Dmitri Dolgov, at Google's self-driving project since 2009, to speak at SPC — adityaag · 2026-09-10
- Midjourney opens annual community call to rank next 12 months of priorities — midjourney · 2026-09-10
- Alexandr Wang's Muse app hits #3 on the App Store — alexandr_wang · 2026-09-10
- Warner Music forces Suno to kill its original AI model, lobbies governments on AI training — Neurogence · 2026-09-10
- Paul Christiano joins OpenAI board's Safety and Security Committee — ethanCaballero · 2026-09-10