3 ms·
I can see how “divide and conquer” can get the job done for many standard tasks/modules, but it seems like it invites a lack of coherency across units as the mo
by swatcoder 4y ago
I can see how “divide and conquer” can get the job done for many standard tasks/modules, but it seems like it invites a lack of coherency across units as the model stumbles into a different style and approach each time you engage it.
This sounds like it would bring huge maintenance and optimization challenges to any large effort.
What are you seeing so far and how are you planning to keep it under control? Or is the utility of this always going to be fundamentally bounded by the context size?
- immortalloom 4y agoYea you're totally right! So the more I learn on compilers vs GPT-3, the more of a maintenance issue things become. So as much as possible I'm using GPT-3, and then using compilers where it makes sense. I have had good results so far feeding in things like directory structures, and asking the model a question, and getting good answers as to what files and folders should be added or modified for an outcome. For example, a prompt might be "given the following directory structure, where should I put my prisma schema file?" and GPT-3 tends to answer this correctly