We can't prove that humans do actual reasoning, but we can be relatively certain that the current models don't (yes, I know there are some who claim otherwise,

We can't prove that humans do actual reasoning, but we can be relatively certain that the current models don't (yes, I know there are some who claim otherwise, but it's just not particularly convincing). The debate over whether or not "reasoning" is just an emergent process isn't necessarily a bad one. But if it is an emergent process, there's little evidence that AlphaCode has emerged it. If "reasoning" does just boil down to repetition, then humans are repeating patterns on a much deeper and more sophisticated level than these models are, and it's still worth questioning if the model's current approaches and the strategies we're using to build AI will be good enough to get to that level of mimicry on their own without a lot more innovation. Philosophical zombies and perfect imitations of humans are interesting and useful thought experiments, but it's not necessary to prove that humans have qualia/reason in order to prove that much simpler algorithms and statistical methodologies do not, at least not at their current scale. And from a practical perspective, the distinction does matter, because we're trying to develop practical tools and it's useful to have a rough understanding of what the current models are and aren't capable of -- ie, they're not currently capable of explaining why they actually made certain decisions. That affects the way that we use them as practical tools. ---- I guess if it helps, sub out "actual reasoning" for "pattern matching to such an advanced degree that it can generate consistent responses that look like reasoning, and that are useful in the real-world as real-world explanations for why a choice was made, and that (importantly) can be communicated to humans, built upon, affirmed, or corrected in real-time." In other words, if non-pattern-matching logic/reason doesn't exist, fine, it could very well be an abstraction or an emergent property. But the machine isn't good enough at pretending to have that, and that limits its usefulness to scenarios where you don't care about it being able to imitate "actual reasoning". As an analogy: folders don't exist on your computer; it's all ultimately just bits on the drive and lookup tables. But there is still a really big practical difference between an operating system that has a hierarchical filesystem and an operating system that doesn't. In other words, abstractions can be purely imaginary, and it can still be useful to distinguish between things that adhere to those abstractions and things that don't. Similarly, even if "actual reasoning" is a made-up abstraction, there's still a big difference between an AI that can explain why a decision was reached and an AI that can't -- even if both AIs ultimately end up working the same way under the hood. As far as I can tell, AlphaCode is a fairly long ways away from hitting that point where we can even pretend to ourselves that it's thinking about what it's doing.