[resf]

People said the same thing about Go. That it was far too difficult and that a pro-level AI was 10+ years away.

What happened was new mathematical tools and new hardware were developed, and suddenly it was all too possible.

It's clear that with our current tools, general AI is out of reach, and new tools must first be developed. But because nobody has any idea what those new tools are, it could happen overnight or over 100 years.

[lambdadmitry]

It didn't come from nowhere, the current streak of ML achievements rides on the back of deep learning, which is an elaborate pattern matching at its core. What makes Go "harder" than chess is that it's difficult to estimate how good or bad a particular position is, so we employed a "magical box" of deep learning and learned to estimate how good a particular move is. That's pretty good, but let's not forget

- it took a ton of very hard work

- it's not transferrable to other domains per se ("elaborate pattern matching" can be, but it's not even an AI)

- this has nothing to do with qualia, consciousness or the theory of mind.

Programming is not about elaborate search or pattern matching at the end. It's about formalizing a domain, stripping it from some subset of real-world complexity, and inventing a solution to a problem in that domain. A rift between beating someone in Go and deducing a fact that doubles wouldn't do well in financial calculations is immense.

This sort of over-extrapolation of current trends is surprisingly prevalent in the tech crowd to be honest. It's like folks in mid-20th century who saw both airline and car industry exploding and made a "logical" guess about flying cars being the obvious next step. Guess what, physics doesn't work that way and flying cars are a dumb idea. The current AI craze seems very, very similar to me.