flâneur

Anatomy of a Modern Finetuning API

benanderson.work · 1,567 words · saved by 4 readers

After months of silence followed by a few very cool blog posts, Mira Murati's very expensive AI lab, Thinking Machines, just released its first product. That product is a language model fine-tuning API called Tinker. Tinker is very different from any fine-tuning service I've seen, and as someone who spent the better part of a year working on an LLM fine-tuning, I wanted to share my initial thoughts. Despite the provocative subtitle, my reaction is positive—Tinker is unique, shows promise, and most importantly, reveals some interesting details about how frontier labs offer fine-tuning services at scale. When I built an LLM fine-tuning service in 2023, we did it the "regular" way. Reinforcement learning wasn't mainstream yet, LoRA and QLoRA had just been introduced, and the main paradigm was supervised fine-tuning, which teaches a model to mimic pre-written responses from a training dataset. To use the service, a customer would upload a dataset, with columns like "prompt" and "response",

--> Anatomy of a Modern Finetuning API Anatomy of a Modern Finetuning API Tinker, Tailor, Soldier... why? Posted Oct 4, 2025 by Benjamin Anderson After months of silence followed by a few very cool blog posts, Mira Murati's very expensive AI lab, Thinking Machines, just released its first product. That product is a language model fine-tuning API called Tinker. Tinker is very different from any fine-tuning service I've seen, and as someone who spent the better part of a year working on an LLM fine-tuning, I wanted to share my initial thoughts. Despite the provocative subtitle, my reaction is po

saved by

related reading