Comment by riskable
9 days ago
Context size helps some things but generally speaking, it just slows everything down. Instead of huge contexts, what we need is actual reasoning.
I predict that in the next two to five years we're going to see a breakthrough in AI that doesn't involve LLMs but makes them 10x more effective at reasoning and completely eliminates the hallucination problem.
We currently have "high thinking" models that double and triple-check their own output and we call that "reasoning" but that's not really what it's doing. It's just passing its own output through itself a few times and hoping that it catches mistakes. It kind of works, but it's very slow and takes a lot more resources.
What we need instead is a reasoning model that can be called upon to perform logic-based tests on LLM output or even better, before the output is generated (if that's even possible—not sure if it is).
My guess is that it'll end up something like a "logic-trained" model instead of a "shitloads of raw data trained" model. Imagine a couple terabytes of truth statements like, "rabbits are mammals" and "mammals have mammary glands." Then, whenever the LLM wants to generate output suggesting someone put rocks on pizza, it fails the internal truth check, "rocks are not edible by humans" or even better, "rocks are not suitable as a pizza topping" which it had placed into the training data set as a result of regression testing.
Over time, such a "logic model" would grow and grow—just like a human mind—until it did a pretty good job at reasoning.
Upvoted, as it basically 99% matches my own thinking. Very well said. But I, personally, would not predict a breakthrough in this direction in the next 2-5 years, as there is no pathway from current LLM tech to "true reasoning". In my mental model LLM operates in "raster space" with "linguistic tokens" being "rasterization units". For "true reasoning" an AI entity has to operate fluently in "vector space", so to speak. LLM can somewhat simulate "reasoning" to a limited degree, and even that it only does with brute force - massive CPU/GPU/RAM resources, enormous amount of training data and giant working contexts. And still, that "simulation" is incomplete and unverifiable.
I would argue that the research needed to enable such "vector operation" is nowhere near the stage to come to fruition in the next decade. So, my prediction is, maybe, 20-50 years for this to happen, if not more.
Wasn’t this idea the basic premise of coq? Why didn’t it work?