Thanks for the mention. I forgot to get back to your earlier comment, so I thought I'd clarify here. The principle of don't give your model capabilities you don't want it to have goes beyond just the data.
For eg. teams are now looking for a way to remove hallucinations, have GPT do precise computations, and generate charts without any error. My supposition is that the tool doesn't do any of the things well. Instead of training it with a lot more resources, look at alternative techniques (use a reader model on specific contexts instead of entire documents, pass additions along to a computing function, and call matplotlib). Where Precision is a must- consider regular expressions for processing, rule engines for decision making. Instead of trying to take a tool, and mold it to fit the problem statement- define your needs first. Pick the tool by your need. That's generally what I mean by good design
Hey Logan,
Thanks for the mention. I forgot to get back to your earlier comment, so I thought I'd clarify here. The principle of don't give your model capabilities you don't want it to have goes beyond just the data.
For eg. teams are now looking for a way to remove hallucinations, have GPT do precise computations, and generate charts without any error. My supposition is that the tool doesn't do any of the things well. Instead of training it with a lot more resources, look at alternative techniques (use a reader model on specific contexts instead of entire documents, pass additions along to a computing function, and call matplotlib). Where Precision is a must- consider regular expressions for processing, rule engines for decision making. Instead of trying to take a tool, and mold it to fit the problem statement- define your needs first. Pick the tool by your need. That's generally what I mean by good design
This is an excellent point. Thanks for clarifying!