Keeping a Fine-Tune Alive Across Model Releases
Daniel wrote in with a plan: fine-tune DeepSeek Flash 4.1 on several hundred hand-edited episode transcripts, deploy the checkpoint to Modal, and repeat every few hundred episodes. Then Hannah asked what happens when 4.2 ships. We dig into what fine-tuning is actually for, where style lives inside a transformer, and why a LoRA adapter trained on one version often collapses on the next. Plus the DeepSeek V4-Pro routing episode from this month, and why a model name is a contract. Episode 5520 — myweirdprompts.com/n/5520 (permanent short link: myweirdprompts.com/128586)
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.