Making coding deterministic? going in circles resulting in wasting time and getting blocked

Prompting
, ,
1

One aspect when trying to build code,

When dealing with creating GPT’s the Ai model is nondeterministic - I find myself going in circles trying to get it to guide it when coding. It often gives me entirely different code adding things or missing things out and I end up getting blocked which is really frustrating especially when in a creative flow for an idea.

How to make it nondeterministic?

What prompts can I use to make sure it follows my instructions perfectly and doesn’t imagine stuff and keeps on track, it always seems to steer totally off course and I need to keep correcting it.

What workflow works for you?

I often hit a the “limit brick wall”

Example, I had an invoice format I wanted it to look at in PDF and this was my first experience making a GPT

asked it to look at certain elements and extract those specific elements and firstly put into a table to show me how the data structure will look.

I also told the GPT that I wanted to make it understand the elements step by step that was included in the PDF

First it could not take the PDF at all saying something along the lines of, I cannot read PDF’s I’m sorry etc etc.

I then converted into an image and asked it if I uploaded 30 images can it process them for me, it said sure! send me them, I also asked can I send PDF’s in a zip, it said yes, I asked can you accept 10 at a time in the zip. it said YES.

Then tried to upload the zip file, then told me I’m sorry I can’t open zip files i don’t have that function.

you see were I’m getting at, by the time I tried to break things down I was blocked just by trying to upload a few invoices.

I ended up trying again, making a new one sometime later, i then started instead with images, then it said that I can upload in PDF etc, then guess what? It could read PDF’s all of a sudden, so I trained it so far then started to change all the ideas I had previously trained it on only outputting half the data or less and not completing the job properly. I think I managed to get it to complete the job ONCE perfectly, even told it ok that’s great, that is exactly what I have been training you to do and I want you to output the WHOLE PDF information like this in future files I send, didn’t work. only 2 or 3 lines of data and when it did try to output the table in full didn’t happen again and started to give me just two rows or columns in the table. I also asked just to output as a csv file so I could then import into google sheets, yet again it only outputted one line.

2 Likes
2

I’m not sure I fully understand but the LLM’s confusion might be that it’s true LLMs don’t directly take PDFs or ZIP files, because the LLM itself has to be sent just text. But maybe your GPT has it’s own (external to the LLM) ability to convert those into text before sending to LLM. So asking an LLM about what kind of files it supports will give random incorrect results perhaps.

3

I totally agree. Its so easy to get lost in the question and answers hell, because the model loose the previous answers, and start hallucinate in the middle of the conversation.

I think it would be wonderful we could ‘pin’ some answers or prompts as a key reference for the topic, in a way that every time the LLM would need to remember something, it could start on the pinned topics as the source of truth.

Also another simple idea that is simple to implement and would aggregate a lot to make the conversations more focused is to delete some Q/A pair.

With this two simple additions to the interface the application would gain a lot in usability helping us to keep focused and avoiding to getting lost on long conversations.

I hope that someone in OpenAI would read this, and pass it to the dev team.

4

yeah you tell it strict instructions via training the Ai such as - GREAT you got that absolutely perfect the output you just gave me, ok now lets develop the next steps together, but do not change the outputted example apart from extending it as we develop it together.

It agrees confirming that.

it just basically doesn’t listen after that and starts going in all random directions even after you gave it strict rules drives me mad!!! hahaha

therefore resulting in going in circles, once I get to that stage I need to start again! with a new GPT by that time I’m locked out for hours upon end and I pay for this, just doesn’t sit right for coding right now. I could probably code it manually faster in Visual Studio.