Past Performance is Not Indicative of Future Results (2020)

That's how I felt too. Most of the article is trying to pull us with an emotional attachment (mostly to racist things a computer will do if tasked to do important things). While that criticism is welcome, it's not specifically meaningful towards an argument against AGI. The only part that was seemed to be that statistical inference is not a path to AGI which is somehow backed up by the emotional stuff.

What deep learning seems to step into more and more is time-based statistical inference.

AGI is not:

seeing that a girl has a frown on their face.

seeing that a girl has a frown, because someone said "you look fat"

seeing that a girl has a frown because her boyfriend said you look fat

seeing that Maya has generally been upset with her boyfriend who also most recently told her she is fat.

But keep going and going and going and we might get somewhere. Do we have the computer power to keep going? I don't know.

Covid is a bio-weapon. Fauci is the dealer.

Not really adjacent to what we do today.

I view A.I. as dual to "neoliberal M.B.A. culture". Just as the business schools taught that managers should be generalists without craft knowledge applying coarse microeconomics, A.I. that we have created is the ultimate pliant worker that also knows nothing deep and works from statistics. In a bussiness ecosystem where analytics and presentations are more important than doing things, they are a perfect match. Of course, a bunches of statistician-firms chasing each other in circles is going to exhibit the folly, not wisdom, of crowds.

I think solution is to face reality that more people need to learn programming, and more domain knowledge needs to be codified old school. https://www.ma.imperial.ac.uk/~buzzard/xena/ I thus think is perhaps the best application of computing, ever.

Training A.I. to be a theorem prover tactic is a great way to make it better: if we can't do theory and empiricism at the same time, we can at least do meta-empiracism on theory building!

I think once we've codified all the domains like that and been running A.I. on the theories, we'll be better positioned to go back to the general A.I. problem, but we might also decided the "manually programmed fully automated society" is easier to understand and steer, and thus less alienation, and we won't even want general A.I.

Cybernetics and control theory, broadly speaking, involve the design of data feedback loops to govern simple machines or complex socio-technical systems. For instance, an organization might instrument a feedback loop to use qualitative survey data to inform decision-making. That isn't ML, but it is cybernetics. And, based on Peter Norvig's definition, it is a form of AI.

Consider that "autopilot" was invented in 1914, long before digital computers. From this perspective, Artificial Intelligence might even be seen as an ancient human practice— present whenever humans have used artifacts to govern complex systems.

I feel like this would have felt more relevant maybe five-ten years ago when there was more of a feeling that deep neural nets was the end all. He mentions correlation vs causation but seems to have missed that causal inference is one of the most active and interesting fields of research today.