In this article, I am going to do something unprecedented: I am going to tell people to argue less about cool philosophical questions.
Imagine you were in a forest with a bear who was both hungry and, like most bears, evinced little respect for human rights and the will of God. As it happens, there are a number of very interesting philosophical questions involving bears. Are they conscious (almost certainly yes!)? To what extent if any do they have knowledge? Are there really such things as bears, or just particles arranged bear-wise? If a bear assembled as a consequent of a random lightning strike in a swamp and then started referring to things…
But none of these things are directly relevant to the dilemma at hand. Whether the bear is a composite object or just an arrangement of particles, it can still eat you. Whether the bear has real knowledge, it still has whatever relevant analogue of proto-knowledge is needed for it to eat you.
A giant share of AI discussions strike me as like haggling over bear metaphysics when assessing whether the bear will eat you.
Lots of debates about AI concern the question of whether it has real understanding. Whether, for example, what it does when solving a math problem is just pattern recognition or something deeper. And to be sure, this is an interesting question. But it’s not directly relevant to assessing the trajectory of AI.
We know that AI can code, discover novel math proofs, provide original analysis of philosophical arguments, and more. We know that AIs can complete programming tasks that take humans a number of hours, and this has been doubling consistently, only to speed up recently. If you ask an AI to analyze a political issue, you’ll get a much more sensible answer than you’d get from 99% of the population—and not just in terms of knowing more facts but also in terms of reasoning better. We similarly know that AI is being designed to complete increasingly long tasks and function as an artificial agent.
The case for AI being a big deal, and potentially a serious threat to the world, simply depends on extrapolating out these trends. If AI is getting smarter, more ingenious, and more agentic by the day, then in the limit, we should expect AIs that are very smart and agent-like. That poses a number of quite severe risks. It also has the potential to radically upend the global economy by replacing most human workers.
You don’t have to sort out any of the interesting philosophical questions to evaluate this case. Whether the AI really understands what it’s doing or whether it’s conscious just isn’t relevant to this case. The latter might be relevant to how we ought to treat AI, but it’s not directly relevant to assessing the capabilities of AI.
The most famous argument against AI understanding comes from John Searle. Searle asks us to imagine a guy in a room who is designed to respond to messages sent from Chinese speakers. The guy didn’t previously know a lick of Chinese. Over time, he memorizes a series of patterns—when such and such phrase is used, you respond with such and such. Searle argues that this guy could communicate intelligently even if he didn’t really understand Chinese. LLMs, it is often claimed, are similar.
I think there are a lot of gripes to be had with this argument, but let’s just grant it for now. None of this is at all relevant to assessing the capabilities of AIs. If you want the guy in the Chinese room to be a translator, you don’t need to know if he has Real Understanding (TM). You just have to know if he can do the job. LLMs demonstrably can do various complex tasks with increasing aptitude.
This is one thing that irritates me so much about the stochastic parrot canard. AI skeptics claim that AI is just a stochastic parrot, meaning that they are trained to predict the next token without true understanding. Now, never mind the fact that “next token prediction,” isn’t even an accurate description of how AIs work in the modern day. Never mind the obvious reply that to predict text you have to understand the world (if you don’t have the foggiest clue how trains work, you will generally be unable to predict their behavior). Never mind the fact that similar suggestions would imply that humans are just reproducers, because that’s what we were made by evolution to do.
But the biggest problem with the argument is that it replaces an analysis of AI capabilities with an analysis of AI ontology. Even if you think all AIs are doing is copying pre-existing patterns in some sense, that doesn’t automatically tell you what they can do. What they can do is an empirical question. Conditional on there being stochastic parrots that can solve novel math problems, make art, do philosophy, and write poetry, what leads to the overwhelming certitude in the minds of the AI skeptics that AI will never replace many jobs?
There’s a line that often goes through my head when reading many of the “AI is useless,” advocates. It came from an old article that wrote in response to a book arguing against gay marriage on grounds that gay marriage does not count as real marriage. Richard argues that this is an insane way to go about things—you don’t need to figure out the ontology of marriage to figure out what public policy ought to be:
Methodologically speaking, I find the "metaphysics first" approach to public policy rather bizarre. For example, when instituting an intellectual property regime, the core question is not "what is intellectual property?" (as if there were some pre-legal fact of the matter), but something more like,
what values are at stake here and what policies/laws would best serve these values?
Similarly, I find the “metaphysics first” approach to deciding AI capabilities bizarre. If you want to know what AI will be able to do, you should look at its actual capabilities. You should not try to deduce it from first principles, then move the goalposts each time your prediction about what AI will never do is disproven. Just compare the track record of the AI skeptics to the AI boosters.
Yann LeCun, one of foremost AI skeptics, claimed that AI—even GPT-5000—will never realize that if you put an object on a table, and move the table, the object will move. Then, a year later, AI did just that. And it didn’t take GPT-5000 to do it; it was literally the next iteration. This is a pretty staggering error—LeCun managed, in claiming that this would never happen, to be off by infinity orders of magnitude. His example number of an AI that would never be able to do this was three OOMs off.
Then, after being this crazy wrong, LeCun moved on as if nothing was up. The goalpost moving is insane! Over and over again, people have claimed that AI will never do something, only for AI to do it one month later. Ask yourself: would the stochastic parrot people have correctly predicted present AI capabilities if they were forecasting in 2023? Of course not! They demonstrably didn’t.
It’s extremely hard to figure out things by doing analytic philosophy from first principles. The track record there is much worse than in empirical domains. This is why we should generally try to figure out what the world is like through the use of empirical evidence, rather than the kind of aprioristic reasoning common from the AI skeptics. If your toy model for how AI is supposed to work incorrectly predicted 10 of the last 10 AI advances, then maybe you should give it up!
And guys, this is coming from me! I think that you can use a priori philosophy to deduce that the world is infinite without doing any scientific measurements (technically this is a posteriori, because it updates on your existence, but sure). Me telling people that they have a problem with trying to get too far from philosophical reasoning from first principles is like Jeffrey Dahmer telling someone they have a problem with eating too many people or your dog saying you have a problem with reacting too much around squirrels.1
I think I saw this joke in a Jeff Maurer article somewhere but I can’t find it!
