Skip to content
Main Site News Console

# [AINews] The End of Fine-Tuning

· Latent Space Translated
播客深度访谈

Today’s editorial was sparked directly by OpenAI’s discontinuation of support for its fine-tuning API.

For years, OpenAI was one of the big tech companies that did the most to support fine-tuning, and a huge number of talks, articles, and AI engineers promoted the idea that you could achieve “o1-level performance at 4o prices” through some variant of fine-tuning, insisting that it was an essential tool in the toolbox.

Now the tide has turned. Anthropic is very likely to surpass OpenAI in valuation for the first time, and fine-tuning has become the next casualty in the 2026 Side Quest purge, after Sora. If you assume GPUs are severely in short supply, this shift makes sense. But even without an obvious compute bottleneck, the 80% of people who make up the majority of the AI engineering industry have probably already been moving in this direction, and Jeremy Howard pointed this out on a podcast back in 2023.

For most people, the “end” of something does not mean it is completely over—in fact, top companies like Cursor and Cognition, whose $25 billion funding round is now being discussed publicly, are increasing their RLFT and use of open models rather than reducing it. Open-model fine-tuning may also be central to the custom ASIC thesis, but if Taalas’s models and con