Ever since we accelerated the scientific method and an objective understanding of the universe, we have relegated divine interference to the parts of our perception and experience that science cannot (yet) explain. Hence was coined the phrase "God of the Gaps".
I've often been reminded of this phrase over the last four years. Every advance in the models was met with "but they cannot do this", "these are mere stochastic parrots", "oh, but these don't reason like a human", etc.
And yet the Large Language Models (LLMs) have gotten better, the harnesses have improved, and all the scaffolding around these models have meant that more and more of cognitive tasks have come under the circle of competence of LLMs.
The ardent skeptics remain unmoved, though their rationalizations move to an entirely different zip code every few months. "These models will never do X". X turns into Y and then into Z as the frontier capability advances.
I also have my version of skepticism, so might as well fess up to it now. While I predicted the shifting sands of copium when it comes to AI, I have maintained that these models don't reason like a human, and that artificial general intelligence (AGI) is a pipe-dream, that cannot be achieved unless we have 1) embodied AI, and 2) some more architectural tweaks to unlock performance like we did with attention on autoregressive neural nets and looping model calls.
Like a good Bayesian, my beliefs should update along with the evidence. So as I sit in this coffee shop, thought I'd reflect on my current beliefs and where I think this whole thing is going. Even if it makes me uncomfortable.
First, I don't mess around with consciousness. There is furious debate on whether these models are conscious, or can ever be. I kind of don't care. Consciousness isn't a verifiable quantity, even in humans - as far as I'm concerned. It's semantics judo-karate that I don't find interesting. So let's move along.
AGI. My main contention was that these models don't reason like people. They are next-token-predictors. But what if they do reason like people, because that is exactly what humans do? I didn't and don't want to believe it. I want to believe that something more than mere prediction is what makes us human - but I'm starting to waver on that. Psychology and neuroscience in recent decades have made it rather plain that prediction and updating our beliefs when we predict something wrong is exactly what goes on in our brains and minds - and that is exactly what LLMs are doing, just via different mechanisms.
I thought general reasoning was a gap that will forever belong to humans. Claude and its ilk are closing in on that gap. In the next decade or so we will embody this artificial intelligence as well. Kind of wild that we built the artificial mind before the artificial body - but this removes doubt for me that we will get there. Then I don't see any structural and durable gap between the competence of humans and that of AI. Only that the AI will be much better at various tasks.
So, where is the silver lining?
...
Exactly.