Get the latest tech news
First impressions of early-access GPT-4 fine-tuning
GPT-4 fine-tuning is starting to roll out in limited early access, and we've been experimenting with it for a while now. Here are our learnings about the performance leaps, the cost and latency of the latest-and-greatest fine-tuneable general purpose LLM.
We use LLMs to answer our users’ natural language questions, with the goal of providing a great starting point for further deep dives into the data. Due to the complexity of the task and the need for a custom output format (a DSL that the AI uses to effectively construct app UI state), we found that fine-tuned models tend to perform significantly better than the base ones. Note: the original version of this article showed significantly worse latency stats for GPT-4: our first benchmarks only achieved 5.8 tokens/s, but OpenAI has since greatly improved the service's stability and speed.
Or read this on Hacker News