Human intelligence is just, well, one of those things man.
In the early days of toolmaking, we largely produced things that could do stuff that our nails and teeth accomplished. Rocks, then the improved obsidian adze, could cut stuff for us. Fire could make things more digestible, again, taking some responsibility away from our teeth. That's what the fossil records tell us; as we progressively grew more human, we found ourselves having more, better-shaped teeth well into our thirties. This was a tremendous boon for our species.
Over time, new technologies improved upon or simply asserted themselves over our environment and social life. I'll try and map them out as best I know their emergence, with a focus on significance. Agriculture helped supplant our need to hunt and gather, but appears to have worsened health outcomes by forcing us into monoculture diets. Pottery let us both preserve and (via microbial means) enrich these foods, and also gave us the ability to commit debt sheets and eventually dramas in the form of clay tablets. Shipwright disciplines helped open up new environments from Ireland to Indonesia: the boat saved our feet by letting us literally 'coast' via the frictionless seas and lakes to new places.
Each of these inventions were the byproduct of human ingenuity. What AI claims it can do is outright replace that human ingenuity the same way that ships saved our backs and feet and knives and fire saved our teeth. But I think that basically fails to understand what an invention does, so I'm going to skip ahead a few hundred years to consider the first significant worker replacement stories.
For thousands of years, in every culture where it was feasible, cotton has been a dominant economic force. It seems to have emerged as a cottage industry simultaneously in Africa, India, and Mesoamerica; not being indigenous to northerly climes, Europe found itself to the the odd man out on this organic wonder material. There had long been cottage industries across Northern Europe using wool, flax, hemp, and nettle to make clothing. As anyone who has been stung by the lattermost, it's easy to notice that it has great fibrous potential. But cotton produces the boll as a fruit, and you can probably imagine how much fancier the first merchants displaying their wares looked coming into England with their full-spun Egyptian cotton sheets. I'm not sure who did it where, but some Englander decided to follow up on the mechanization of coal-mining that was already going on in southern Wales and apply it to the process of launching a shuttle and controlling the weft and weave of a fabric. And so, over the course of a few decades, the cottage industries of fiber work were extinguished not only throughout the British Isles, but even India where it was violently suppressed by colonial leadership as a means of carving out a new market for the demand of this product.
Now, there's an assumption that AI is going to do something similar to the market of the mind. But in my limited judgment, it's going to fall short of what British-woven cotton fiber products could do. And I am going to do my best to explain those failures here.
- It doesn't think, or know, anything. LLM thought exists as a map of tokens differently weighted in terms of context. You're not asking it 'how many Rs are in the word 'strawberry,' so much as 'How many [1181] are in the [3141] [9876] [0901].' No wonder it has problems answering that. There's no reflection, and at its core, no knowledge. Now we can fuss all day about what intelligence is. But we know some level, it's the process of iterating over something until we feel we understand it better, if not completely. Cotton clothes could at least cover your body; LLMs just fool you into thinking they do.
- It's a bit too general. Mechanization replaced the crofter because you could go from raw cotton to a finished, dyed product under a single roof if you made the factory large enough. It cannot replace the plumber or carpenter because it doesn't work outside of a factory, unless you're building houses for delivery. While it can write (bad) fiction, it can't replace code with significant brownfield history or even write basic code requirements if it doesn't understand tribal knowledge. Which, by default, it won't bother to.
- It produces junk. It's great at searching for boilerplate code. Miles better than Yahoo or Google was for search. And if you want, it can go into a deep-dive about whatever topic you desire, but it's ultimately going to run into self-serving hypotheses that are hard to parse from conspiracy theories. This is not a good look for a product that claims to 'know.'
Now, it's not awful at everything. Visual recognition is an impressive tool. It can do that repetitively, and at high speed. Alarm systems are going to become much better, especially with the sheer amount of visual learning data that are being fed into this machine. So it is with faces, which may lead to its own intelligence and human rights nightmares. But that's not actually acting on human judgment, just the recognition of a few billion triangles mapped onto somebody's lips, nose, and eyes.
Much to the dismay of folks like Ray Kurzweil, this whole AI hullaballoo is designed for one thing: impressing investors and making gobsmackingly tall stacks of money. Jensen Huang isn't the first instance of someone who realized that the time was ripe for the taking of stacks of investor cash, we are currently in the throes of what may be termed an AI summer. It might be springtime or fall in relation to the agonizing winter that will soon come, but the basic tropes are as follows, and have been occurring on a similar level for about 60 years, or ever since we believed that AGI was an achievable output of bigger and faster computing systems:
- Hardware systems reach a critical juncture. Some advancement or another produces a state in which a theorized form of AI starts to take off. It might be Perl, which optimized pattern recognition. Or it might be modern LLM work.
- A cool demo goes public. Maybe it does voice recognition properly. Or draws us a cool picture. Or recognizes that a camel is different from an elephant or a horse. Either way, this wows investors.
- Investment spikes. More novel hardware is applied. Factories spring up in Japan or Taiwan or China or Arizona, depending on the year in which this happens. The technology is (currently) more valuable than extracting raw minerals.
- Advancement plateaus. Throwing more hardware at the problem doesn't produce more Very Cool Demos. We start to find bugs that are are features. ← YOU ARE HERE
- Investor money chases something else. Improving access to bauxite, oil, timber, gold, or concrete becomes more valuable than shares in these companies.
- Investment declines. Factories shut town or are refactored for building proven technologies that still have demand. Winter descends for a decade or two until novel hardware emerges and we go back to stage 1.
We are now navigating this same passage for the 5th, 7th, or 9th time, depending on how you count it and who you ask. More investment capital is in technology now than ever before, and we may be in for a beating once the excitement cools. Plenty of tech companies who foisted AI on their workers are coming to realize that it cannot replace skilled technical work at scale, and the biggest ones will be suffering for this choice for years after the leaders who made those decisions have bailed out with their golden parachutes packed and ready to rip. Perhaps this will have better macroeconomic outcomes than I hope for, but I've learned to be disappointed by the behaviors of private industry.
My disillusionment with where capital choses to invest itself is, however, not the point of this essay. It's more to point to what intelligence actually provides for us, and what might eventually reproduce it. There well may be some distant future in which higher cognitive loads can be impressed on a purely chip-based matrix, but they're not likely to be particularly novel, unless they're something that mechanized adoption does particularly well. Let's think back to the shuttle-weft-weave dynamic of cloth manufacturing I mentioned earlier. Sure, it fits nicely into something with universal demand that can be considerably cheapened. But certain skilled trades will resist being cheapened, some by their very nature.
I mentioned earlier that we were at step 4 of 6 of my rough map of the AI boom-bust cycle. That doesn't mean we're closing in on the winter phase anytime soon. Industrial growth phases can sometimes enter managed declines rather than apocalyptic collapse. If there's a lot of money at state, it makes more sense to share in a shrinking pie than to ruin the party for everybody. Failure will accumulate, companies will consolidate, talent will go elsewhere. There's no need for (or profit in) running around screaming about the end of the world. There's also no proof the singularity is anywhere near.
Note: This missive was written entirely without AI, save for using Google to look up the title of 'Empire of Cotton,' a fabulous book about the simultaneous rise and eventual smothering of indigenous crofting communities from various parts of Eurasia, Africa, and America. If I could've avoided it, I would have, but it's indistinguishable from Google Search now.