Hamel Husain breaks AI failures into three types: specification, comprehension, and generalization. Most failures are actually comprehension problems that look like generalization. Before fine-tuning or upgrading your model, fix your prompts, retrieval, and data analysis first.
When it comes to analyzing your AI product failures, there's generally 3 buckets of failure: specification, comprehension, generalization. Generalization is when you have given your model all the context possible that it needs to succeed, and you've also done a lot of data analysis to know that those instructions and that prompt is good, as well as your retrieval and everything else is good. It's just, frankly, the model is not carrying out your instructions, even though it has fairly pristine information. Now, it is the case that failure modes that people tend to experience are in the other categories. However, if the model is still not doing what you want, there are some things you can do like fine-tuning or exploring more powerful models. But in my experience, everyone always has some room to grow when it comes to doing better data analysis or refining their prompt in their retrieval.
