Post #1490151

2026-04-18 17:41 UTC

I just started reading From Deep Learning to Rational Machines: What the History of Philosophy Can Teach Us about the Future of Artificial Intelligence by Cameron Buckner. I can already tell this is going to be an interesting read for me. I'm quite frustrated by how much DL research seems to think getting high scores on benchmarks means they have "solved intelligence," without really stopping to think about what they've actually done and how it relates to what we call "intelligence" in living beings. When they do reference the relevant philosophy, they do so shallowly to make themselves look credible. This book frames the discussion in terms of two sides: one that assumes humans are naturally endowed with instinct and a capacity to reason that DL cannot replace, and one that assumes humans learn by experience alone, and DL will, too, if we can only give them enough data. The book attempts to break down this dichotomy, and explore the fertile ground in between.

Replies (1)

@ngaylinn@tech.lgbt 2026-04-18 17:41

I'm very much in favor of finding a middle ground! However, I'm alarmed to find myself in the "nativist" camp, as Buckner defines it. Or, at least, I'm reluctant to say that DL learns from experience alone starting from a blank slate, or that evolution doesn't provide us with innate bias that shapes and enables our intelligence. So, I might find parts of this book quite frustrating, but hopefully in a useful way for refining my own arguments! I say the problem is we think of brains having an innate supply of facts, which seems mostly untrue. DL suggests it's also unnecessary. However, I think our innate endowment is not facts, but constraints. So, when we set up a particular ANN architecture, dataset, and training regime, that is us giving the ANN an innate endowment. That is what makes it able to learn the relevant facts. When we claim that such models learn from a blank slate, we neglect that we have done the same work that evolution does for living systems to make learning possible.

Open ##2384208