Superintelligence: The Idea That Smart People Refuse to Think About

I could be misrepresenting Maciej here (idlewords, please correct me if so), but my understanding of that talk was that "superintelligence is not worth worrying about" was only part of it. A lot of it was more about the deleterious knock-on effects that seem endemic to AI worriers, developers, SV bigwigs, and the rationalist community. A lot of the shitty behavior of these groups, that Maciej has been hammering on for some time now in previous talks, is closely correlated to belief in AI X-risk. He seemed to be claiming that this correspondence is not just correlative but causative. That may or may not be true, but even if not we should focus on improving it all the same. To wit:

1. Stop treating every problem in the world as an engineering problem to be solved with clever software, and realize that people and social problems are more complex than that. This is one of the motivating factors behind wanting AI (so we can turn social problems into programming problems) and leads to megalomania on our part and deep resentment on everyone else's part.

2. Similarly, stop thinking us developers are the noble heroes who are going to save the world from the ignorance of the masses. This is both dead wrong (we are in a rapacious capitalist industry that mostly wants to eliminate anything that stands in its way, whether or not the impediment has good reason) and again is building resentment toward us.

3. Start worrying more about the extremely serious problems that are being caused by data retention and machine learning right now, like the permanent loss of privacy, increasing wealth consolidation in the hands of a few capitalists, and the "laundering" of human bias through supposedly-but-not-actually-neutral algorithms.

4. Realize that whatever neurons in our brains are lit up by religious belief don't necessarily require a "religion" in the traditional sense of the word, and that a lot of this eschatological transhumanist and AI X-risk stuff seems to slot into the same pathways, especially for very smart engineers. That doesn't make them wrong necessarily, but it does mean their claims should be analyzed with the same bias toward extreme skepticism that most of us would automatically use with religious claims.

I suspect Maciej would be much less distressed by AI worry if the AI worriers also concerned themselves more with 1 through 4. At that point it would be at worst a harmless distraction, and maybe even beneficial if AI X-risk is in fact a legitimate concern. That, I think, would be best for everyone.

I think the argument about AI overtaking humanity is wrongly framed. Sci-fi has taught us to be on the look out for an all-powerful machine, robot, algorithm, turning against humanity. A single, united swarm of machines that only needs humans for energy. This is unlikely to happen soon.

But if we borrow an idea from Richard Dawkin's The Selfish Gene, things look differently. He argues evolution is not about the success of individuals or even species, but about the proliferation of their DNAs. The organisms are just the vessels for DNA's replication.

Let's look at the corpus of AI technology as an evil machine's DNA. It already claims a lot of human energy and capital for its evolution and proliferation. Our top researchers invest their time into making it better. We burn fossil fuels to provide the power for its computational cycles. What's the primary fuel for its evolution? Money. Which comes from stock markets and VCs, which use AI models. I'm sure a lot of fundamentals-based financial models, based on very different lines of reasoning, end up valuing AI-based companies higher. It is in no way a deliberate conspiracy, just a result of the AI doing it's job well.

Of course, self driving cars, painting funny moustaches on your faces, or playing Goare all benign, non-threating abilities of AI. But it is easy to see how falling into the wrong hands, they might lead to not-so-great results. Our stereotypical AI researcher is surely very careful about AI's potential dangers. But do you think Uber, a notoriously profit-driven company, would invest this much care if it doesn't help the bottom line?

The cat in a cage argument is surely convincing, but AI is already taking a lot of decisions in place of humans. Be it mortgages, crime detection and prevention, or what information reaches us through our news feeds. And it is easy to see how a single imperfectly-configured algorithm can lead to real-world effects.

For example, imagine a threat assessment AI, as already used in some shape and form by police forces around the world, which is configured not to minimize deaths and injuries, but to minimize damage to the economy by an individual. A benign sounding setup. Imagine such a system has to evaluate the threat level of a known researcher, who specializes in AI safety, thus reducing the profit of a bunch of the world's biggest companies. Is he a threat to the economy? Should he go to jail?

I think the fundamental difference of opinion--between those who worry about the existential risk of gAI and those who don't--is not whether or not gAI can pose a risk to humankind, but rather whether it will, and if so, whether there's anything meaningful we can do about it now to prepare for it.

This article is written as a rebuttal to another article (http://idlewords.com/talks/superintelligence.htm) adapted from a talk. That article reconstructs the gAI-is-risky argument and then provides a series of counter-arguments that show that the gAI future is not guaranteed--that there are many ways in which it might not come to pass. The point it's trying to show is that the gAI-dystopia is not a given, and therefore we should question its plausibility. In particular, it raises the notion that there may be something about very high-level thinking ("being smart") that makes causes very knowledgeable people to overestimate the probability of gAI-dystopia. It then suggests that the indirect consequences of promoting the gAI-dystopia theory might outweigh any possible benefits we get from worrying about it right now.

Then this article responds to each of those arguments (about how it might not come to pass) with an argument about how it might come to pass. It closes with a remark that any engineering discipline should ask what the consequences are if we succeed.

Obviously, the two camps could go back and forth all day, both being completely correct, neither making any progress, because these arguments are based in rhetoric, not analysis. There's nothing wrong with rhetorical arguments, but any two sufficiently skilled rhetoricians can push their point infinitely far; the real question is which premise you accept:

After hearing the arguments put forth by the gAI-dystopia group, do you think the dystopian scenario is likely and that we can meaningfully prepare for it at this time, or do you think it's unlikely or that we can't meaningfully prepare for it?

The gAI-dystopia crowd is not necessarily wrong, but it's a minority of the community because it can't attach any meaningful risk analysis to its claims without making shaky assertions for the prior probabilities. There are infinite ways we could end up in a gAI dystopia, and there are infinite ways we could not. Existential threats are hard to accept, so until we have more evidence that such a future is at all likely, it will be hard to convince any significant number of people. In this sense, it is just like a religion (but I don't say that as an insult).

This isn't a dupe. It's a response to the other article with a similar title. I've reposted under a less-confusing title.

How exactly does "refuse to think about" fit with the existence of OpenAI[0] and the Singularity Institute[1]?

[0] https://en.wikipedia.org/wiki/OpenAI

[1] https://en.wikipedia.org/wiki/Machine_Intelligence_Research_...