[HN Gopher] Data preparation for function tooling is boring
       ___________________________________________________________________
        
       Data preparation for function tooling is boring
        
       Author : andreeamiclaus
       Score  : 4 points
       Date   : 2025-05-13 21:16 UTC (3 days ago)
        
 (HTM) web link (thehyperplane.substack.com)
 (TXT) w3m dump (thehyperplane.substack.com)
        
       | andreeamiclaus wrote:
       | From building your own Siri, now you learn the boring dataset
       | part that you cannot skip!
        
       | simonw wrote:
       | > 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 Am I out of date?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.
        
         | simonw wrote:
         | 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?"
        
       ___________________________________________________________________
       (page generated 2025-05-16 23:01 UTC)