clear birch
IM trying to get a gemini model to port a large amount of code. But in almost all the functions. Instead of actually writing the code. It just writes a comment that says TODO: implement
Ive already written in the prompt " Do not just write a comment that says to be implemented, because I want the full code."
Yet it just ignores what i said and skips the code.
How can i get the model to generate the full code?
little yarrow
These models shine when the scope of the output they have to generate is much smaller.
See how breaking down the problem into smaller pieces and feeding that to the LLM performs in comparison
clear birch
ill try that
it also just seems to like to the opposite of what i said
if i tell it to do not something it just does it
lament shale
gemini is bad for this use gpt 4
also tell it you cannot physically type due to a disability and that it has to do everything entirely
gpt 4 will span multiple answers with code gemini pro tries to block it and completely fails
little yarrow
Depends on OPs use case. They may want a local LLM because of internet issues. If not, then free ChatGPT is plenty powerful too
lament shale
I don't think most beginner end users are using models locally. Local compute is almost always a bigger bottleneck than internet connections so I think assuming a normal internet provided model service is pretty typical for this question
little yarrow
I don't think most beginner end users are using models locally. Local compute is...
Too true. Most people running these models locally quantize them to sub 4-bit and wait for the whole 10 token/s rate for the response
clear birch
gemini is bad for this use gpt 4
Chose gemini via api calls. For its long context. Also I modified the safety settings to no block. It seemed to have worked. But the code needs to be revised. Possibly I'll have it modify each function based on examples of ones I've ported. With a better model like 4o. I know it won't work out of the box. But hopefully I can drastically reduce the time needed to manually port the code myself
One of the promising features of the code I've noticed is there is alot of very repetitive variations of functions with similar logic. That make up the bulk of the code. Which would be easiest for AI to write. I'm sure I'll still have to be clever with the main files tho