In 2018 or 2019 I saw a comment here that said that most people don't appreciate the distinction between domains with low irreducible error that benefit from fancy models with complex decision boundaries (like computer vision) and domains with high irreducible error where such models don't add much value over something simple like logistic regression.
It's an obvious-in-retrospect observation, but it made me realize that this is the source of a lot of confusion and hype about AI (such as the idea that we can use it to predict crime accurately). I gave a talk elaborating on this point, which went viral, and then led to the book with my coauthor Sayash Kapoor. More surprisingly, despite being seemingly obvious it led to a productive research agenda.
While writing the book I spent a lot of time searching for that comment so that I could credit/thank the author, but never found it.
> A great misunderstanding accounts for public confusion about thinking machines, a misunderstanding perpetrated by the unrealistic claims researchers in AI have been making, claims that thinking machines are already here, or at any rate, just around the corner.
> Dreyfus' last paper detailed the ongoing history of the "first step fallacy", where AI researchers tend to wildly extrapolate initial success as promising, perhaps even guaranteeing, wild future successes.
https://en.wikipedia.org/wiki/Hubert_Dreyfus's_views_on_arti...
> A puzzling characteristic of many AI prophets is their unfamiliarity with the technology itself
> After reading these books, I began to question whether “hype” is a sufficient term for describing an uncoordinated yet global campaign of obfuscation and manipulation advanced by many Silicon Valley leaders, researchers, and journalists
AI Snake Oil – by Arvind Narayanan and Sayash Kapoor
Nexus – by Yuval Noah Harari
Genesis – by Henry Kissinger, Craig Mundie, and Eric Schmidt
The Singularity Is Nearer – by Ray Kurzweil
A big part of the problem, the authors maintain, is confusion about the meaning of artificial intelligence itself, a confusion that sustains and originates in the present AI commercial boom.
Narayanan and Kapoor are particularly worried about the conflation of generative AI, which produces content through probabilistic response to human input, and predictive AI, which is purported to accurately forecast outcomes in the world, whether those be the success of a job candidate or the likelihood of a civil war. While products employing generative AI are “immature, unreliable, and prone to misuse,” Narayanan and Kapoor write, those using predictive AI “not only [do] not work today but will likely never work.”
2. The idea that using ML to predict outcomes "does not work" is so obviously wrong that I don't really feel the need to argue against it. Perhaps weather models, content moderation systems, NLP analyzers, spatial modelers, and the vast universe of other examples are all not really AI in the first place, in their book? In that case, what is "predictive AI"? Just a few cherry-picked examples of local governments trying to cheap out on bureaucratic processes, I guess?
After this brief intro, we arrive at the meat of the article. Picking on a Harari book seems like beating a dead horse, but y'know, sometimes that's fun! Still, the specific criticisms fall flat:
[Harari] offers the example of “present-day chess-playing AI” that are “taught nothing except the basic rules of the game.” Never mind that Stockfish, currently the world’s most successful chess engine, is programmed with several human game strategies
Harari fails to explain that while machine-learning models assemble a template of solutions to a specific problem (e.g., the best possible move in a given chess position), the framework in which those problems and solutions are defined is entirely constructed by engineers.
Harari bungles straightforward issues and ideas concerning artificial intelligence.. But Harari, attempting to argue that the alignment problem is a timeless conundrum, applies [the alignment problem] to historical events that did not materially involve artificial intelligence
Finally, there's the criticism about the COMPAS system that ProPublica uncovered (the true GOATs in any story). But what exactly is the criticism there? "He was critical, yes, but not critical in exactly the way I prefer"? That applies to pretty much every book ever in some way or another...
I'll skip going through the other two as closely--because I'm on the anti-markdown site, where walls of text are the only option--but it's all just the same tired assumptions wrapped in a condescending attitude. The writers of Genesis are far from experts in AI, but regardless, the criticisms of both them and Kurzweil come down to variations on one theme: "these people think AI is a big deal, which is obviously wrong, because it's not". I don't think you need me to tell you that this is not a solid argument.
I mean... Ugh. Criticizing the idea of a technological singularity as an "imaginary event" that "consists almost entirely in extrapolation" is again technically true, but the implied pejorative usage of these terms is completely unfounded; it is no more imaginary than climate change, nuclear war, or the simple empirical assumption that the sun will rise again tomorrow.
It's especially tiring to read this when we're literally in the middle of the singularity right now, which is quite obvious if you hear the real meaning of the term ("a point where our models must be discarded and a new reality rules"[3]), rather than the somewhat-bungled description here that relates more to Intelligence Explosions (" sufficiently advanced machine intelligence could build a smarter version of itself, which could in turn build an even smarter version, and that this process could continue to the point of vastly exceeding human intelligence"[4]).
The only people who still think the future of AI('s effect on humanity) is predictable post-2022 are the ones who are dogmatically certain that computers as we know them will always be crappy tools at best. I implore you, privileged reader: do not fall into this comforting trap. Face the future with us, despite the terror. Posterity is counting on us.
[1] https://github.com/official-stockfish/Stockfish/commit/af110...
[2] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Sorry, why is it misleading to recognise that all neural networks are created manually by human engineers?