Training data distribution caps AI agent performance, developer argues

Developer @gerardsans argues that training data distribution sets a hard ceiling on model capability—like a model trained on a phone book never writing Shakespeare—and that after release, only harness tweaks remain, which cannot rescue a weak model.

2026-11-15 ~ 2026-11-15 · 2 related posts

• "Phonebook-trained models can't write Shakespeare": data distribution beats any harness — gerardsans · 2026-11-15
• Once a Model Ships Training Is Frozen: Harness Tweaks Can't Rescue Weak Models — gerardsans · 2026-11-15