How Low-Dimensional Is Intelligence?

If we only trained students in math, and then assigned students best at math to learn our most important jobs, that might go pretty well. After all, humans who best learn math tend to be well able to learn many hard tasks, and we can judge math ability unusually well. However a nation would not get rich by mainly training its people in math, and trying to sell math consulting to the world. After all, total world demand for math consulting is quite low, far too small to support a nation. And productivity spillovers from math to other professions seem too low to justify this strategy.

This difference is related to a key long-debated AI question that I recall from when I first started to learn AI in 1984: if we train AIs to be “super”-intelligent in just a few key areas, are they then “super-intelligent” in general, finding it far easier to learn to do most any other tasks? Or must high levels of AI development efforts be put into a great many different topic areas in order to create AIs who can do the full range of tasks in our modern economies, and thus replace human workers wholesale?

Asked another way: how highly-dimensional is the concept of “intelligence”? Is an AGI (artificial general intelligence) one that has learned how to do many particular things, or one that has learned a few key ways to “learn to learn”? Long ago, when I was (one of?) the first to greatly engage Eliezer Yudkowsky’s AI fears, this was one of our key cruxes. And it is related to the question of what exactly makes smarter humans: do they use a few better general abilities to learn faster about everything, or do they just have better brain hardware to faster and more reliably apply the same human inheritance of a big standard library of ways to do many particular things?

This is now a key crux re fears that AI labs are now or might soon “foom” to “super-intelligence” via “recursive self-improvement” (RSI). Many have been quite impressed with recent frontier AI abilities at math and coding, and METR recently reported that Anthropic has seen a recent ~x2 jump in the rate at which it has improved on key metrics. Some who “get it” say that this shows that Earth will within a few years see AI gods completely remake our world, and most likely kill all humans. But others interpret this as just a one-time jump, or due to benchmarks saturating, or Goodhearting of metrics, or shifting training compute more to reinforcement learning. Or it actually shows diminishing returns to key inputs of cash or hardware.

But even if we took such accelerating gains at face value (which I don’t), we can reasonably doubt huge near-term social effects of an AI industry that made large and accelerating gains at mostly only math and coding. After all, AI just does not seem to be gaining much traction at improving most jobs, and AI firm customer revenue has plateaued at levels far below what can justify current AI spending levels, raising fears of a market crash and big cuts in AI development spending. Furthermore, though coding is where AI seems to gain the most traction, coder and software engineer wages have not noticeably risen, an effect that we economists clearly predict if AI substantially improves overall coder productivity.

Sure, if just one AI got far better at military and security aspects of coding, that could destabilize relative power in the world. But that’s much less of a problem if many AI firms offer similar super-coder abilities to most customers.

Many plausibly attribute the much faster gains in math and coding to the fact that results in such domains are more easily verified. AIs also seem great at essay writing, plausibly due to LLMs being first created from huge databases of human writing. Some hope that physics simulations will let physical interaction domains become more verifiable, but we have seen only limited gains there so far.

My bet is that smarter humans mostly have better brain hardware, not a few general tricks, and that even an AI RSI “foom” soon in AI math and coding abilities would still take decades to greatly penetrate the world economy, as math-foomed AIs could not actually learn how to do most other tasks that much faster. A pattern economists have seen re past general purpose technologies.

However, I’d prefer policy depended less on mine or others’ particular estimates. So I support robust strategies, such as requiring large (~$100B/event) liability insurance for AI firms re the possible large damage events AI execs have suggested as possible soon, perhaps even enhanced more closer to more dangerous scenarios. If insurance firms estimates high risks, the premiums they demand will force AI firms to halt or greatly slow down, appropriately for that scenario, but if they instead estimate low risks small premiums won’t much slow development, appropriately for that scenario.