To what extent does this conceptual model of LLMs represent how they work?

[nawzak]

Select Topic Area

Question

Body

1. An LLM can be represented by a model which has two components, an interface, and a computing function
2. Both components are separate yet interoperable
3. The computing function performs the core tasks performed by the model (e.g. generating a software code)
4. The interface part interprets human language queries to the computing function (e.g. providing user requirements for a software code)
5. The interface part translates the output of the computing function into an easily understandable format.
6. The purpose of the whole system is to align the request at its input with the response at its output
7. Hallucination occurs when the output of the computing function is inaccurate yet the interface part translation of this output still makes sense
8. An output response of the system due to a prompt provided at its input is generated through an interaction between both the interface and computing function components
9. Fine tuning of the system using small datasets creates a new model which includes its own interface and computing function components

[KateCatlin]

Hi @nawzak! Thanks for posting this question. Your understanding is a good start, but if I were being picky I’d change the wording of a few things:

1. LLMs don’t necessarily have a distinct “interface.” They’re probabilistic models that process input text as token sequences and generate outputs based on learned statistical patterns. The UIs you see to interact with them (like chatbots) are separate from the LLM itself. You can call an LLM and get full responses directly through a terminal!
2. That said, I guess you could think of certain parts of the model, like tokenization and decoding, as an "interface" in the sense that they help transform raw input into a format the model can work with. But these components are part of the model itself, not a separate system.
3. When generating text, the LLM determines the most likely next word or subword (e.g., predicting ‘lik’ → ‘ly’ vs. ‘lik’ → ‘able’). Hallucinations happen because the model’s probabilistic generation produces incorrect (but, plausible!) text, not because a separate "interface" misinterprets the output.
4. Fine-tuning adjusts model weights but does not introduce distinct components.

If you want to learn more about how LLMs like ChatGPT are built, I’ve heard good things about this instructional video: https://www.youtube.com/watch?v=kCc8FmEb1nY

Hope that helps :)

[nawzak]

@KateCatlin Thank you Kate for this informative discussion. I understand and agree with you in points 1 and 2 which you’ve kindly mentioned. As for point 3 according to https://www.nature.com/articles/s41599-024-03811-x/tables/1
“AI hallucination has not yet solidified into a universally agreed-upon definition” so I think explaining why hallucinations happen might not be that simple.
As for point 4 which you kindly mentioned, what I meant was that by changing the model weights, the whole model changes

[KateCatlin]

Sounds great! Can I ask what you're building with AI? What kinds of products or features are you working on? Or are you exploring more philosophically?

[nawzak]

Having worked as an assistant lecturer at university for years before my current job as a senior technology specialist, I’m interested in writing brief articles and blogs about technology topics including AI. My main participation is in the following sites:
https://www.researchgate.net/profile/Mohamed-El-Nawawy-3
https://medium.com/@melnawawy1980
https://www.linkedin.com/in/mohamed-el-nawawy-2aa11375?originalSubdomain=eg

[KateCatlin]

Oh, very cool!

Well, let me know if any of our GitHub AI tooling (current and upcoming ✨) gets mentioned in any of your writing! I'd be happy to retweet or LinkedIn share. 😃 Best of luck!