It seems unfair to be access-locked for over-use, when I am battling an idiot at the other end…

How do users force AI to always follow the custom instructions as my custom GPT’s do not?

I’m using the web interface to ChatGPT and set up a custom GPT for an old-school programming code - BASIC. Whether I use the custom GPT where I can’t select which model answers, to the newer GPT 4o with canvas… both are useless at coding a language that has a history on the Internet. I load up my custom GPT with the relevant syntax documents, etc… and I’ve tried the “canvas” version to how it handles queries.

Either one has very short memories now days. They forget an instruction at a moments notice, even just 2 responses later from being told and explained it’s a crucial syntax issue.

In the end, I get access-blocked online for too many back and forths with the silly AI coding output that just doesn’t want to follow basic :rofl: instructions.

3 Likes

I use loops to remind GPT in my instructions I highlight and give cross reference instructions in code as reminders so it can loop kind of like an old BASIC 10 go to 20 20 go to 10 but loops in loops. Basically an algorithmic function.

Like as an example I always number my instructions and in them I loop instructions back to old instructions in function. IE If you use this tool it reads how to do it every time. You can also set it to read multiple paths to make sure it’s processing all instructions at every use.

This is my BASIC AI B.A.S.I.C. AI the walk down memory lane

Welcome to the forum.

The issue presented here is not unique.

I program in two very uncommon programming languages, yet no AI tools currently exist that are effective for daily use with these:

The core challenge lies in the training methods of large language models (LLMs), which often lack sufficient context regarding the specific programming language and its corresponding version.

It is observed that for more mainstream programming languages, AI performance tends to improve. However, even in those cases, its utility is often limited to basic tasks, frequently using additional prompts with error messages to produce functional code:

  • HTML
  • JavaScript
  • CSS
  • JSON

In short, if it does not work then do not use it.

1 Like

OP “ Either one has very short memories now days. They forget an instruction at a moments notice, even just 2 responses later from being told and explained it’s a crucial syntax issue.”

How does that keep the GPT from “forgetting” instructions?

Maybe that’s the plan. The worst subscribers in terms of profits are ones that use services.

The model powering GPTs seems to specialize in not following an instruction.

1 Like

This happens for a ton of structured generations users give instructions to do it in a method but the GPT “forgets” it’s all over the forum.

It doesn’t.

The takeaway is

In short, if it does not work then do not use it.


With regards to it forgetting and having to refresh the prompt, I understand the problem and that is a temporary solution but in the end I think it is just a dead end.

What I am alluding to is that while the attention mechanism has helped in making great strides with some hard problems with AI, it does not solve all problems and too many users expect that some similar wins will carry over into other problems. Having been around for many decades with AI this will again lead to another AI winter.

2 Likes

Oh I see yes I agree. It’s like the old joke hey doctor my arm hurts when I do this… then don’t do it :laughing:I sometimes over read things and thought I missed something :heart::strawberry:

1 Like

My method is more like reinforcement learning, I focus more on infrastructure than data. But yes I like branching logic too like attention mechanics

@mitchell_d00

You might find this of interest.

While I am not currently using it, I do keep the idea in the back of my mind and think if it might be of value for certain problems. It keeps popping into mind when I think of the problem of different programming languages and specific versions.

And yes I am dropping this here in the case someone reads this and decides to run with it in training a LLM based on the concept and then reporting it in a public free research paper.


If one watches the video they might think that is the same as the attention mechanism and I would be hard pressed to come up with the details to show the difference. However in my mind what would be done is to hold the concept(s) (node(s)) for both the programming language and the version in the attention mechanism while the related code is being learned.

If one thinks of reading a book then one keeps in mind the book title, chapter name, section name etc. in their mind, while reading. The attention mechanism drops these off the end as the sliding window moves. Instead the attention mechanism should have a stack for the context nodes and a sliding window part as is currently done. As the sections, chapters and book change so would the concepts (nodes) in the stack associated with the attention mechanism. The upside that might come of this is that the length of the sliding window might hit a sweet spot shorter than the current sliding windows thus making the training time faster and lowering the energy needed.

1 Like

Im out of hearts but yes TY! :heart: I read a lot of this out of the box tech. I started out before all thisAI stuff as a physics nerd who deep dives into game theory :rabbit: I am high functioning so I’m a spaz :crazy_face:

Way off topic but have you noticed that we practically never see mention of
Physics-informed neural networks (PINNs ) on this site?

1 Like

Yes but entropy is coming up everywhere people considering chaos and order but balanced they become indistinguishable , amazing patterns in this forum it almost vibrates metaphorically of course.

Black Box topics that are kind of PINN but really not much else in forum. Weird .

Edited to add this cause I read it for like 10 min trying to decide if I posted a half thought :rabbit::heart:. Because BB aren’t transparent so one can’t tell if it is structured base or what not like a PINN

:strawberry: is kind of PINN. In how it “thinks” in steps.

I know it’s chain of thought just comparing it to the process a PINN would have in doing “physical” steps

There’s probably not many making robots learn to dance.

Here’s your link from last decade as close as you’d get, before OpenAI was for GP consumer language products: Welcome to Spinning Up in Deep RL! — Spinning Up documentation

(MuJoCo is opensourced by Google DeepMind now.)

1 Like

Welcome o the forum :rabbit:
PINN can be applied to virtual physical phenomena you don’t actually have to have physical things moving.

Welcome to the forum.

That is not what I think of when I think of using a PINN but I do see the connection.

The link felt like using the Wayback Machine, quite the walk through history.

1 Like

We took this way off topic I am sorry @me.mark . I get trapped in rabbit holes and this one was joyful. This forum is mind bending :rabbit::heart:

1 Like

What i mean is that on an OpenAI forum there is no Physics AI offered. They probably won’t help with your bugs or complaints!

1 Like