ctdinjeu2
It's a subjective reasoning problem, not an objective reasoning one - we've got the latter with LLMs.
But it only takes 2 seconds to output the result. There will be no "agent working while I sleep" since it only takes 2 seconds before I'm needed again.
Said another way: The calculator doesn't do my taxes.
To enable an LLM to work longer, autonomously, to be "an agent" on my behalf, I'd need it to approve things and make decisions too. For that, I'd need subjective reasoning.
For that, the LLM needs to have a model of me too, and of my cultural environment. This should be possible in the ad age (no pun intended), but nobody is doing it yet.
Subjective reasoning is I think a different modality, in the way that code-writing is a different modality from question-answering.
Sometimes correct means factually and objectively so. But sometimes it means the most aligned with my preferences, or what might be good for me, or inspiring.
To invoke the tax analogy again: Say you get the calculator to do your taxes - now get it to do it how you'd do them, regardless of what's generally correct, ethical, yielding the highest return and so-on. You need a subjective reasoning model, or if it's a training data problem only, then a language model with a subjective reasoning modality.
In the early 00s, nobody even knew what the hell we were doing. Most people asked things like “do you do IT?” nobody really got it.
The idea of photoshopping a mockup, or the idea that Wordpress is how you build the website for their business and why they now need “HTML5”, the cloud and advent of IaaS/PaaS, I digress:
All the esoteric aspects of the work and knowledge have been curricularized.
There are a million Joe React Developers now, everybody can use Figma, anyone can manage a sprint, and with AI forget it - anyone can play now.
Buuut not so fast you say, and yeah I agree with you!
And yet, I know that you know that they know that we know that they know. You know?
LLMs can’t retain most codebases nor even most code files accurately - they start making serious mistakes at ~500 lines.
Paste a ~200 line React component or API endpoint, have it fix or add something, it’s fine, but paste a huge file, it starts omitting pieces, making mistakes, and it gets worse as time goes on.
You have to keep reminding it by repeatedly refreshing context with the part in question.
Everyone who has seriously tried knows this.
For this reason alone the LLM “agent” is simply not one. Not yet. It cannot really drive itself and it’s a fundamental limitation of the technology.
Someone who knows more about model architecture might be able to chime in on why increasing the context size will/won’t help agents retain a larger working memory to acceptable degrees of accuracy, but as it stands it’s so limited that it works more like a calculator that you must actively use rather than an autonomous agent.