> Let's look at the data: 72% of enterprises are now fine-tuning models rather than just using RAG (22%) or building custom models from scratch (6%). This isn't a trend, it's because fine-tuning works when other approaches fail.
Where did that data come from? My mental model is still that most companies find fine-tuning an LLM isn't worth the effort compared to promoting with better chosen examples or setting up effective RAG. Am I out of date?
On reading further: it looks like this series of posts is specifically about building voice assistants that run on a mobile phone, which need TINY models. From what I understand getting tiny models to perform interesting custom tasks is a challenge that fine-tuning is well suited for.
They surveyed Fortune 500 types for it. The numbers above were from a survey of 70 "AI decision makers" and the question concerned "How are enterprises customizing their models?"
> Let's look at the data: 72% of enterprises are now fine-tuning models rather than just using RAG (22%) or building custom models from scratch (6%). This isn't a trend, it's because fine-tuning works when other approaches fail.
Where did that data come from? My mental model is still that most companies find fine-tuning an LLM isn't worth the effort compared to promoting with better chosen examples or setting up effective RAG. Am I out of date?
On reading further: it looks like this series of posts is specifically about building voice assistants that run on a mobile phone, which need TINY models. From what I understand getting tiny models to perform interesting custom tasks is a challenge that fine-tuning is well suited for.
I think I found the source: A16Z in March 2024: https://a16z.com/generative-ai-enterprise-2024/
They surveyed Fortune 500 types for it. The numbers above were from a survey of 70 "AI decision makers" and the question concerned "How are enterprises customizing their models?"
I am curious why function calling and not MCP server, don't they serve the same functionality?
From building your own Siri, now you learn the boring dataset part that you cannot skip!