If I have but one take to belabour regarding AI, it is that non-general AI is real intelligence, and permanently world-altering because it is freed from the constraints of biology. I think that opinions on AI which do not grant this—though understandable—represent failures of imagination or 'terror management'. I wish to expand on this in the next few posts.

Part I

There's this idea in literature that an author cannot write a character who is smarter than they are. I don't think this is strictly correct, because intelligence is both an intensive and extensive quantity. The foregoing idea only appears to speak to the extensive aspect of intelligence, that a character cannot have larger thoughts (even constructively) or see a bigger picture than the author. This hides the fact that there's something special about the act of writing as a tool for thought.

A book is not a conversation, but an opus. It might masquerade as a conversation, but it is not. All books are very far from equal, but all are released from the deliberative constraints of the conversational form. An author can take much longer to sharpen their wits and say what they mean. They can converge on the limit of their cleverness. They can write and rewrite. Every author has a different rate of convergence, and a different final destination.

"One day I will find the right words, and they will be simple"

— Kerouac (paraphrased?)

Part II

Our deliberative power is constrained by time and energy. If you pause for long enough in a debate, you lose. If you die before you finish writing your book, you lose. Russell and Whitehead cited intellectual exhaustion as a reason for why the fourth volume of Principia Mathematica was never written. This seems reasonable for a series which (loosely speaking) takes several hundred pages to prove 1+1=2. Formalised tests of intelligence typically have a time-dependent aspect, and not only because the invigilator has to go home at some point. We intuit that the speed of thought and the ability to think sustainedly has something to do with intelligence.

Whilst some intellectual problems are stones which cannot be weathered down without a certain genius, others are eventually soluble in the acid of 'ordinary thought', the concentration of which varies. We actually do value fast thinkers, even where we suppose that we could have arrived at the same answers eventually. We can easily conceive of thinkers who have greater initiative and cognitive tempo than ourselves, and reasonably understand them to be smarter than us. We find that sometimes tokens per second is a quality all of its own.

Similarly, we can grant that superior powers of attention, recall, and the ability to "correlate the contents of one's mind" is an aspect of intelligence that—whilst lamented by Lovecraft—appears to be on balance a desirable thing. John von Neumann ostensibly had an eidetic memory, and was therefore able to memorise and recall textbooks full of mathematical identities rather than having to look them up. It seems like this may have been a contributing factor to him being one of the most consequentially intelligent people in history. I have been told that having an eidetic memory is actually a terrible weight because 'you would remember upsetting events in photographic detail'. However, this already appears to be my experience, without the benefit of—for instance—recalling every Laplace transform, or even where I parked the goddamned car. This would be a straightforward type of intelligence for an author to express, even if they do not possess it themselves.

This is all to say that we can break intelligence down into potentially orthogonal aspects, many of which are generally valuable because they act intensively on cognition as a whole. They are force multipliers. Deep Blue beat Kasparov in 1997 because it was faster than Kasparov at estimating the contingent optimality of chess moves. In the narrow sense of chess, Deep Blue was smarter than Kasparov, because determining the contingent optimality of chess moves is literally all chess is. Any other interpretation implies that intelligence is a subsidiary factor in chess skill. But it's also the case that chess is a soluble game, at least in principle. This means that there exists a memoisable function which takes the state of the board as an argument and returns the best move, thereby playing a perfect game. Some might baulk at the idea that precomputing and perfectly recalling optimal chess moves is a form of chess skill, however this would also suggest that the study of chess games more generally constitutes a form of cheating.

Multiple intelligences do exist, and they can kick your arse.

Part III

In mineralogy, the term 'ore' is interesting. What makes a mineral an ore? Simply, an ore is a mineral of economic significance, which is a fact of its economic situation. Ore is ore because it's worth mining. It might do us well to operationalise AI the same way.

It is a mistake to regard any AI which can compete for someone's job as not truly intelligent, not least because of the implications for those who are outcompeted. It's intelligent in a way that matters. It's economically significant right now. This is because speed, consistency, error rate, and knowledge base are aspects of intelligence that people are willing to pay for even if they are not the 'whole package'.

"Imagine a jet airliner with a pilot that never gets tired, never makes mistakes, never shows up with a hangover. Meet the pilot."