Some people asked “why should we care about GPT-3?” — this long (30 or so tweet) thread explains my reasoning.
To some people, who were arguing with me about these ideas thirty years ago, or heard about them in the interim, this will be very old news. To others, this will be surprising, or perhaps unbelievable, and perhaps even more reason to question my sanity.
In our 200,000 years on earth, humanity has created more and more capable tools with time to augment our natural abilities. Tools have the interesting feature that they may be turned towards the creation of yet more sophisticated tools.
Roughly speaking, the more general a tool, the more important it is, and the more broadly it can be applied towards creating new tools. Sewing needles, for all their utility, have not had the impact of wheels, because their domain of application is too narrow.
The really big and history-changing inventions have been things like writing, artificially powered motors, etc., which have no single specific use, but which have very, very broad applicability.
In recent decades, humans have invented two sorts of “final” tools. There will be at least two more of these “final” tools to come. These tools are characterized by the fact that they are as extreme as possible in their generality.
The first was generalized (“Turing Equivalent”) computing machines, limited only by the creativity of the programmers and the mathematical limits imposed by the theory of computation. Computers aren’t for anything in particular, and thus are used everywhere.
Originally people imagined computers would be used for things like calculating ballistics tables, or doing payroll accounting, and of course, you can use them for those things, but those are hardly the most interesting uses of computers.
The second tool was generalized communication infrastructure. The internet is not for any particular kind of communication, and thus is for every possible kind of communication. It’s an extremely general communications mechanism, perhaps as general as you can design.
Coupled with computers (which it requires to exist, and which it synergistically augments), the Internet has also become ubiquitous; it’s uses, too, are mostly limited by creativity, not by technology
People originally imagined this second tool would allow for things like remote terminal access to distant computers, and so it does, but again, that application is now clearly one of the least interesting things you can do with it.
The third tool does not yet exist, but is entirely feasible. It will be the machine that can construct anything at all — any possible configuration of ordinary matter that can be envisioned — and do so atomic precision, every bond in its place in an arbitrary sized object.
Such a device will be able to construct any machine you can imagine, including copies of itself, and will be limited only by the creativity of the designers, the laws of physics, and the availability of energy and appropriate types of atoms.
This third tool will permit fully automated construction of truly arbitrary objects — anything from aircraft with spars of pure diamond to molecularly precise machines that can enter and repair damage in individual human cells.
Note that these two are likely to be some of the least interesting applications, just as when computers and networks were envisioned, their most obvious applications were also ultimately some of the least interesting.
The final tool, which is likely to precede fully generalized manufacturing, will be creating automated systems that display the sort of intelligence and flexibility that humans display, though of course, once you’ve built such things, you need not stop at mere human capability.
AGIs may be set at any task humans can perform, from writing poetry to designing starships. They can be built rather than being painstakingly raised over decades, and will learn information and skills other AGIs know instantly rather than over months or years.
Importantly, AGIs can be set at the task of creating and manufacturing faster and more capable AGIs. Relatively tame speculative engineering designs tell us that constructing AGIs *at least* millions of times faster and more capable than humans is feasible.
Artificial General Intelligence is truly the final tool, because it eliminates the creativity barrier presented by the limited availability and speed of human minds. If you need more minds and faster ones, you simply build them.
So at some point, the world will be dominated by intelligent entities that can conduct millions of years of engineering and design work in months, and are coupled with devices that can construct literally anything that isn’t prohibited by physics, including more such entities.
Once such entities are ubiquitous, the world of humans, at least as we understand it, and as has existed relatively unchanged for 200,000 years, will cease.
Why? Because at some point, and that point will happen very, very soon after AGIs are designing better AGIs, there will be AGIs whose intelligence exceeds that of humans as much as human intelligence exceeds that of ants.
Note that I’m not suggesting humans will necessarily cease to exist. Ants still exist, of course, but they don’t have much control over the course of history.
I know some smart people who think this is all garbage and that there’s no sign of progress towards AGIs. I would strongly argue that such people haven’t been paying attention to the work being done on the topic.
(I know some people who have odd psychological barriers to the notion of artificial intelligence entirely; they seem to have some sort of deeply ingrained belief in vitalism. I don’t see why one should imagine solution phase carbon chemistry is important to intelligence.)
I know others who do research on the topic and think it’s nowhere close to happening; I think in most cases, they’re too close to the problem, and they see only how hard the problems immediately in front of them are and not the arc of progress being made.
Regardless, in my opinion, essentially any human problem you spend most of your time worrying about will cease to be a recognizable issue some time “soon”. “Soon” could be years, it could be decades, but I’d argue that matters little.
Whether “soon” is two years or one hundred, it’s nearly instantaneous compared to the timescales of human history. It might matter a fair bit to you personally of course, but not very much to the (posthuman) scholars of the future.
Anyway, that’s why GPT-3 is significant. Sure, it’s pretty limited. If you know what you’re doing, you can ask it a question like “who was the U.S. President in 1650” to which it will produce a coherent and obviously wrong answer.
Sure, you can ask it to add two six-digit numbers and it will fail. Not really the point. If you had handed GPT-3 to a researcher a decade ago, he would have called it a fraud and assumed he was being hoaxed.
GPT-3 is a harbinger, and a harbinger of something much more significant than the current crisis (no matter what you think the current crisis might be.) And that is why I rank it in importance far above the things currently dominating our headlines. /end
