I've been diving into the topic of LLMs and their reasoning capabilities, particularly referencing the "illusion of thinking" paper. It seems like there's a split between what's known as System 1 (fast, instinctual responses) and System 2 (slow, reflective reasoning) in cognitive science. I wonder if instead of pushing one model to do both, we could create a system where a frozen LLM acts as the instinctual layer while pairing it with an adaptive System 2 that monitors and guides it.
The idea would be to have System 2, a flexible layer using something like Kasparov-Arnold Networks, that could critique and influence the outputs of the frozen LLM without changing its weights. Over time, this could help the system adapt and make better decisions, theoretically moving us closer to AGI. I'm hoping to hear thoughts on whether this approach has merit and if there are existing efforts in this direction!
4 Answers
I appreciate your take on this, but don't forget that the concepts of System 1 and System 2 originated with Kahneman, which influences how we see reasoning in AI. It's true we're using static weights in LLMs, so your idea of an adaptive System 2 is worth exploring. Current systems can learn from interactions, but whether they're achieving true reasoning is still up for debate. I think merging these concepts could potentially fill gaps in how we model AI thinking!
Interesting concept! I can see why you'd think a frozen LLM could serve as a good instinctual base. Just remember that using a static model definitely has its limitations, since it lacks real-time learning. But the way you're suggesting pairing it with an adaptable System 2 could really optimize decision-making over time. It's like having a great reflexive engine that’s guided by smart reasoning! It's a new spin on things that might lead to some cool advances in AI.
I get what you mean! The separation of fast and slow thinking could result in some really exciting outputs. Let's just hope that combining these models doesn't lead to overcomplication.
If System 2 can correct System 1, then why wouldn’t we just skip using System 1 altogether? It seems like right now we’re layering models without addressing the core reasoning challenges, and I’m skeptical we’ll see true AGI anytime soon with just transformers. We might just create a system that mimics reasoning.
I didn’t read the paper, but I have to ask, why do we assume we need to replicate human heuristics for AI to work? Isn't it more productive to find a unique approach?
That’s a solid point! I don’t think we need to replicate everything, but human intelligence offers a valuable template. It’s about leveraging what works from our own cognitive processes to refine AI, since that’s the only true general intelligence we have as a reference.
Exactly! You've hit it on the nail about Kahneman. I think we still need a separate layer that actively tunes how our current models operate to really get to the next phase.