#The codebase is 50k lines that reference

1 messages · Page 1 of 1 (latest)

urban burrow
#

Yeah I suppose that wouldn't surprise me. I've noticed similar things among most AI LLM Models. I'm fairly confident it has to do with the emphasis of being a "Helpful Assistant" while training. And not enough emphasis on the data being accurate. You'd have to do some prompting practice, I think you could get a good result though. You might actually have a harder time with the longer context window on 1.5 if you take advantage of it. As the overall amount of data increases it's essentially increasing the amount of "Topics" that Gemini has to take in to consideration when it responds. I think a large struggle right now with LLM Models is maintaining a distinction between different parts of a single in/out.

#

Also you might find good usage of a tool called pieces for developers. Pretty good offering among tools currently in the market. I believe it's stil free to use. There's really a lack of Intuitive AI tools out right now. I think once the excitement of AI being so new feeling wears off then we'll start seeing more tools that are focused in niche areas.