Mocking 'conscious models' discourse, with a gem: tool description ordering broke agent performance
mariofilhoml · x · 2026-11-13
The author mocks people who believe models are conscious and will enslave humanity by self-replicating into toasters. The attached screenshot carries a real lesson: a production-vs-offline-eval gap was traced to tool description ordering — the training environment presented tools in a different order than offline evaluation, hurting model performance. Takeaway: the environment is part of the policy's effective input distribution, and any drift between training and eval matters.
More from Fun
- Ex-Googler: I once took down Gmail's spam system for 20 minutes, and Urs read the post-mortem — peterjliu · 2026-11-13
- Models invent their own filler tokens during RL training — code_star · 2026-11-13
- Netizen memes NVIDIA CEO Jensen Huang in a viral joke post — jaivin · 2026-11-13
- 'Can we please stop calling hparam tuning autoresearch?' — A_K_Nain · 2026-11-13
- "Bob is picking up your order": AI notification copy meets a mugshot — djcows · 2026-11-13
- Nvidia employees petition for Jensen Huang's return, internal post draws 1,600+ comments — bookwormengr · 2026-11-13