I have some data science background, and I kinda understand how LLM parameter tuning works and how model generates text.
Simplifying and phrasing my understanding, an LLM works like - Given a prompt: Write a program to check if input is an odd number (converts the prompt to embedding), then the LLM plays a dice game/probability game of: given prompt, then generate a set of new tokens.
Now my question is, how are the current LLM’s are able to parse through a bunch of search results and play the above dice game? Like at times it reads through say 10 URLs and generate results, how are they able to achieve this? What’s the engineering behind generating such huge verbose of texts? Cause I always argue about the theoretical limitations of LLM, but now that these “agents” are able to manage huge verbose of text I dont seem to have a good argument. So what exactly is happening? And what is the limit of AI non theortical limit of AI?
Edit
Disclaimer: : All of my LLM experience is with local models in Ollama on extremely modest hardware (an old laptop with NVidia graphics) , so I can’t speak for the technical reasons the context window isn’t infinite or at least larger on the big player’s models. My understanding is that the context window is basically its short term memory. In humans, short term memory is also fairly limited in capacity. But unlike humans, the LLM can’t really see (or hold) the big picture in its mind.
But yeah, all you said is correct. Expanding on that, if you try to get it to generate something long-form, such as a novel, it’s basically just generating infinite chapters using the previous chapter (or as much of the history fits into its context window) as reference for the next. This means, at minimum, it’s going to be full of plot holes and will never reach a conclusion unless explicitly directed to wrap things up. And, again, given the limited context window, the ending will be full of plot holes and essentially based only on the previous chapter or two.
It’s funny because I recently found an old backup drive from high school with some half-written Jurassic Park fan fiction on it, so I tasked an LLM with fleshing it out, mostly for shits and giggles. The result is pure slop that seems like it’s building to something and ultimately goes nowhere. The other funny thing is that it reads almost exactly like a season of Camp Cretaceous / Chaos Theory (the animated kids JP series) and I now fully believe those are also LLM-generated.
You can improve the novel writing by using agents. First you generate just an outline with the plot points to every chapter. Then you chop that up and feed it to several agents to flesh out individual chapters. Finally the generated chapters are verified against the outline and overall plot. If that doesn’t fit, the agents are tasked with a rewrite. Repeat that until you have something serviceable.
As you point out, there exists plenty of bad writing in TV series. These often have a number of different authors, who don’t necessarily know the other episodes very well.
I will say that while most of these models are non-deterministic their training data was very similar so if you did something like this I can guarantee you if you churned out enough you would start to see the common threads.
Sure. Lots of fiction, especially TV stick to well established tropes, regardless of a human writing it or not.