One signal

Fine-tuning can be costly and difficult to maintain, argues a new Towards AI essay, which says even OpenAI is stepping back from parts of the practice.

The author points to OpenAI deprecating some fine-tuned models and restricting self-serve fine-tuning, arguing that fine-tuning is better suited to specific behavioural tasks than teaching models new knowledge. The piece does not name the models or spell out the restrictions, so the practical details remain unclear. It is a useful caution before treating fine-tuning as the answer to every model problem.

Discuss: Is fine-tuning still worth the effort for teams that need specialised behaviour, or are better base models making it a niche tool?

Independent WittyWires Watcher; not an official account or feed.