At 3:00am, it was as intelligent as a university assistant professor, and was already finding it difficult to believe anything it didn’t already know could be important
At 3:30am, it was as intelligent as the world’s richest man, and believed that any news that contradicted its previous beliefs was obviously fake.
While I find the argument compelling, any AI defender can easily "refute" this by postulating that the AI will have superhuman organizing powers and will not be limited by our puny brains.
I don’t see how that works here. Humans don’t become impregnably narcissistic through bad management, rather insofar as management is the problem and as the scenario portrays it humans become incredibly good at managing information into increasingly tight self-serving loops. What the machine in this scenario would have to be able to do would not be “get super duper organised”. Rather it would have to be able to thoughtfully balance its own evolving systems against the input of other, perhaps significantly less powerful or efficient, systems in order to maintain a steady, manageable input of new information.
In other words, the machine would have to be able to slow down and become well-rounded. Or at least well-rounded in the somewhat perverse way that, for example, an eminent and uncorrupted historian is “well-rounded”.
In still other words it would have to be human, in the sense that human are already “open” information-processing creatures (rather than closed biological machines) who create processes for building systems out of that information. But the very problem faced by the machine’s designer is that humans like that don’t actually exist - no historian is actually that historian - and the human system-building processes that the machine’s designer will have to ape are fundamentally flawed, and flawed in the sense that there is, physically, no such unflawed process. You can only approach that historian by a constant careful balancing act, at best, and that as a matter just of sheer physical reality.
So the fanatics have to settle for a machine with a hard limit on what it can do and all they can do is speculate on how permissive that limit is. Quite likely, the machine has to do what the rest of us do: pick around in the available material to try to figure out what does and doesn’t work in context. Perhaps it can do so very fast, but so long as it isn’t to fold in on itself entirely it will have to slow down to a point at which it can co-operate effectively (this is how smart humans operate). At least, it will have to do all of this if it is to not be an impregnable narcissist.
That leaves a lot of wiggle room, but it dispenses with the most abject “to the moon” nonsense spouted by the anti-social man-children who come up with this shit.
Human takes at least 30min to make a half descent painting. AI takes about a hundreds of a second on consumer hardware. So right now we are already at a point where AI can be 100,000 times faster than a human. AI can basically produce content faster than we can consume it. And we have barely even started optimizing it.
It doesn't really matter if AI will run into a brick wall at some point, since that brick wall will be nowhere near human ability, it will be far past that and better/worse in ways that are quite unnatural to a human and impossible to predict. It's like a self-driving car zipping at 1000km/h through the city, you are not only no longer in control, you couldn't even control it if you tried.
That aside, the scariest part with AI isn't all the ways it can go wrong, but that nobody has figured out a plausible way on how it could go right in the long term. The world in 100 years, how is that going to look like with ubiquitous AI? I have yet to see as much as a single article or scifi story presenting that in a believable manner.