Five claims about AI futurism
A continuation of the same email thread with William Eden, this time after he had asked Alyssa for her own AGI timeline. Rather than offering a single year, she sketches five specific claims she thinks are underappreciated in popular AI futurism — covering DeepMind’s PR strategy, the gap between academic benchmarks and real economic value, the difficulty of forecasting downstream social effects, the technological assumptions baked into modern democracy, and how little of the current “chatbot” industry actually uses deep learning at all.
I don’t really have a number for AGI anywhere in my head, other than it being pretty unlikely to happen soon but reasonably likely to happen in the next century. There is too much basic research and ontology shifting between here and there; it would be like physicists in 1800 trying to give a date for the invention of quantum mechanics, when no one even knows what amplitudes are yet. I think exact timelines (if done well) can help out through 2030 or so, and contingent timelines can flesh out a lot of planning (eg. A implies B implies C), but if Omega came and told me that AGI would happen in 2042 vs. 2054 it wouldn’t really change my actions at all.
My writings on AI futurism tend to center around some specific conclusions regarding particular facts. Some major ones here are: Demis Hassabis is extremely good at PR, and almost everyone doesn’t properly adjust for this when deciding how impressed to be. Eg. one might notice how, for all of his most impressive accomplishments, he has carefully chosen things which have a very high ratio of fame to commercial value. For example, most educated people know what Go is, but there is basically no money in it. Ditto Atari games. Ditto chess matches, etc. Since he has a huge amount of funding (both directly and indirectly via Google infrastructure), this lets him cheaply acquire lots of PR points; it’s like grabbing all the culture techs in Victoria II, if you’ve played that. (Stockfish, previously the top chess program, was designed for hardware costing maybe a few thousand dollars, because who would ever pay more than that?) That’s not to say that his team aren’t heavy-hitters etc., obviously they are, just that the PR needs to be treated skeptically.
The “hot” applications in academia are heavily skewed towards a small number of problems which are easy to compete on. Take ImageNet, for example. It’s very easy to measure performance on, just download and run. But what good is it? Who has a bucket of millions of random images that they need to sort into generalized categories? There are lots of somewhat related problems that have real value — eg., being able to process a set of video streams to reliably identify particular object identities and locations has obvious military applications. But that is much harder to test, so it doesn’t get the recognition. (This also leaves big pots of money open for sharp entrepreneurs, although I personally am not focusing on vision problems right now.)
The consequences of any major AI technology (even ones distant from AGI) will be specific, diverse, and hard to predict. Computer hardware, and even more generally industrial stats like amount of oil produced, can be nicely quantified on a continuous scale. But AI is software, and any software is going to be some specific set of techniques adapted for particular purposes. For example, suppose one had software that could semi-reliably estimate (in dollar amounts) the expected legal costs of any action. This would completely change how law works (at least civil law), since attorneys now never think in terms of statistical distributions and probabilities and expected utilities. But it would not be a uniform change — each type of company would be affected in very different ways. To figure out what would happen, one would have to game out what the most important changes in each area would be, and then what governments would themselves do in response to those changes. I could take a decent crack at it given a few months to research, but it would be vastly more complicated than just “lawyers now unemployed!” or whatever people would say on Reddit. (Argh, there is so much dumb political propaganda mixed up in these discussions….)
Somewhat relatedly, social institutions are going to be seriously disrupted by any important AI, which everyone underestimates because there’s very little reliable knowledge on why these institutions work how they do to begin with. For example, average people talk about “democracy”, and they tend to assume that “democracy” simply means electing politicians to a government. However, what most people de facto mean by “democracy” is “how Western countries operate in the late 20th century”, and these operations are all based around an essentially 1960s-70s level of technology. Trump is not just “Americans elected someone stupid”, or even “lots of Americans hate immigrants” (which has always been true), but “we should systematically expect Western governments to work differently now”. (You may have noticed that many of the proposed contenders to run against Trump in 2020 are themselves celebrities with no political experience.) We should expect further changes, and not just in the direction of “more Trumpiness”, but in many possible directions depending on what types of AI are developed and how those AIs are used.
Despite all of the hype around “chatbots” and “intelligent agents” and “personal assistants”, the major real breakthroughs of deep learning (image processing, speech recognition, etc.) have nothing to do with being an “agent” at all; they are supervised machines which produce a fixed type of output from a fixed type of input. The overwhelming majority of “bots” now operate on rule-based principles not fundamentally different from those used by ELIZA in 1966. I had to do a substantial amount of digging before finding anyone who was even trying to use new tech, and when I finally did (Yoshua Bengio’s MILABOT), it was an academic project which had received no press and (to my knowledge) is not being applied commercially.