Unfortunately it’s pretty clear from the article that Cory does not have much familiarity with the research going on in the field of machine learning, and is cr

Unfortunately it’s pretty clear from the article that Cory does not have much familiarity with the research going on in the field of machine learning, and is creating a straw man. Quite a lot of work is being done on causal inference, out-of-distribution generalization, fairness, etc. Just because that is not the focus of the big sexy AI posts from Google et al does not mean that the work isn’t being done. I’d also point out that humans can infer causality for simple systems, but for any sufficiently complex system we also can’t reason causally. But that does not mean we can’t infer useful properties and make informed, reasonable decisions. I’d also point out that not all models are “theory-free”, as he describes it. I specifically do work in areas where we combine “theory” and machine learning, and it works very well. And finally, his point about comprehension does not really fly for me. There is no magical comprehension circuit in our brain. It’s all done via biological processes we can study and emulate. Will that end up being a scaled up version of current neural nets? Will it need to arise from embodied cognition in robots? Will it be something else? I don’t know, but it’s certainly not magic, and we’ll get there eventually. Whether that’s 10 years or 1000, who knows. Are current paradigms going to lead to AGI? Frankly, I’d just be guessing if I even tried to answer that. My gut instinct is no, but again, that’s just a guess. Can current methods evolve into better constrained systems with more generalizable results and measurable fairness? Absolutely.