I. The only remaining fundamental disagreement between people who think AI is a normal technology and those that don’t is the belief that the AI can do everything a human can do, but better. If you believe this, then it’s almost tautological to say that there will be no jobs left for humans. And if you don’t believe this, then it’s almost tautological that there will be jobs left for humans. In other words, the debate about AI can basically be entirely boiled down to whether or not it can continue to acquire capabilities until it covers the full range of human capacity. Smoothing out the jaggedness, so to speak.
This may be obvious, but I want to spend some time on this, because I think there are a lot of other things that get wrapped up in the AI discourse. Questions like ‘is AI conscious?’ or ‘does AI have aligned goals / motivations?’ or ‘is AI violating copyright?’ These are all maybe important questions, but they are not what makes AI potentially unique as a technology, different from every other technology that came before it.
I have yet to hear any explanation for why I should believe AI cannot do any given thing a human could, including manage other AIs. We even have a mechanism of action: just pay enough people to let you record them doing <thing you are trying to automate> and incorporate the data in the next training run. At one point the null hypothesis was ‘AI cannot do X’ and the burden of proof really was on the AI accelerationists. But we are well past the point where that null hypothesis should shift. Even the physical world is not safe, the world’s capital is funneling into robotics and major advances are happening constantly.
Sometimes people will argue that the rise of AI will lead to new kinds of jobs, much like how the rise of factory automation led to new jobs in robotics and maintenance and computer science. Let’s take that premise for a second — we’ll ignore that such automation also led to the hollowing out of towns and cities across the country, and is a significant driver for the economic discontent that led to Trump among other things, and that the new kinds of jobs were not 1:1 transferable both because of capability (”learn to code” implies the ability and time and money to learn, which not everyone has) and amount (there were not as many new jobs as old ones).
AI is different from previous technologies because any new jobs could in theory also be done by the AI.
Let’s do a little proof by induction. Start with three premises:
that there are some number of jobs;
If the AI does a job better than a human, it will end up doing that job;
AI can do everything better than a human.
There are two possibilities. Either the AI does not end up creating new jobs, in which case the AI does all the jobs. Or the AI does end up creating new jobs, in which case, just go back to step one and repeat.
Obviously, people take issue with the third premise, but notice that as time has gone on the people who are grasping for some indication that AI can’t do everything have become increasingly desperate. If you push hard enough, the answer is inevitably something like “the AI simply cannot [dream, love, laugh],” an almost religious appeal to some kind of soul-like quality of the human spirit, as if this matters to the finance team trying to do their accounting. But, again, even in this abstract sense, this is still just a question of capabilities. “The AI won’t take all the jobs because there are things that humans can do that the AI cannot.” And you’re back to the original tautology.
The reason AI is not a normal technology is because it is fully general. People compare AI to things like cars. If you invent the car, it may revolutionize transport, but it isn’t obviously going to impact how baristas make coffee. But if you invent a car that can do anything, well…
II. There is one thing AI cannot and will not ever be able to do, which is be human. For example, AI art will by definition not be 100% human made. There will always be demand for any kind of identifier, so there will always be demand for things that are authentic-human-made. But this is a very narrow slice of the economy, not unlike a couple buying hand-made wood carvings or hand-stitched hats from an “authentic” Balinese shop keeper in Ubud while on their honeymoon. You could justifiably round off economic demand for “human-made” to whatever the global trade is in such upper middle class kitsch, which is basically a rounding error on the global economy overall.
The reason is straightforward: you buy “authenticity” as a preference, not a necessity. Anything that people actually need is always going to be purchased without regard for provenance. And even among the long list of potential wants — of which the want for authenticity must compete — my strong sense is that authenticity trends pretty low on the list. Well below ‘quality’, at any rate, I’m always going to prefer a good pizza over an authentic one, I don’t care if the original Romans wouldn’t put nduja on it, that shit is delicious. Even Anthony Bourdain couldn’t claim to enjoy unwashed warthog rectum as a meal, though I’m sure it was certainly authentic!
Authenticity ends up being valuable only because it is rare relative to demand, not because there is objectively a lot of demand or because it is objectively valuable. Anything that AI can do can, in theory, become infinitely cheap, which becomes equivalent to not valuable.
This is part of what’s happening with AI writing. The AI is mimicking the writing style of prestigious award winners and Kenyan essay writers, i.e. things that used to be rare that people would pay for. If you told me 5 years ago that everyone could write like that, I would’ve thought that we would enter into a golden age of heightened literature even in our ad copy. But that’s the problem — ad copy is still ad copy. The moment that kind of writing became ubiquitous, it became synonymous with cheap vapid thinking, and people literally turn their brains off or react with violence when they discover they are reading AI slop.
Thus, the rise of poor punctuation and bad capitalization in your average substack post or tweet. People use AI all the time in places where they need to communicate a lot of information quickly and efficiently, or where the writing is primarily a chore/work and not a hobby. This is most obvious with code, a language that is not literally devoid of aesthetics but which is certainly not valued for how it elegantly it reads. But you can see the trend in other places too. What percent of papers submitted to prestigious scientific journals is written by AI? What percent of legal briefs? Do you think that number will go up or down next year?
If you assume that the AI can do anything a human can do, the syllogism writes itself. Things are valuable if they are rare; anything an AI can do won’t be rare; therefore, the value of “anything a human can do” drops to zero.
III. I’m writing this with the mathematics community in mind, which is currently in a state of shock over the industrialization of their field. From Field’s Medalist and math extraordinaire Terrance Tao:
Over the last few months, the mathematical capabilities of LLMs have improved dramatically, to the point that they can solve major outstanding problems in many fields of mathematics. However, the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics, and to the mathematical community. The goals of the AI companies and the goals of the mathematical community are severely misaligned. We see these as part of broader alignment issues impacting other scientific and creative professions, as well as the whole of society.
…
In recent months, the success of AI in solving major mathematical problems has made headlines even outside mathematical circles. But solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight. Forgetting this in the world of AI may turn the tool against the primary goal. Indeed, the mass production at faster and faster pace of “true/false” statements could destroy fertile ground instead of breathing life into new ideas.
…
We are witnessing a general threat to intellectual work, with misalignment between the outcome of the use of AI and its initial purpose. In many fields and activities, years of training have traditionally served not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions and ideas. However, building on a vast body of previous human work, AI systems are becoming increasingly capable of producing the results of such work directly, and these goals cease to align.
(One interesting note: Tao has been a huge supporter of using AI in mathematics. It is especially interesting to see him come out against it here)
I’m an engineer, I love making things more efficient, and even I can recognize that there is something intrinsically perverse in solving long standing mathematical problems by simply prompting the AI to “do a breakthrough.”
People who are not mathematicians and likely never intend to become mathematicians are crowing about how this will let the average person solve Millennium problems, as if that was the point of all of this. It’s a classic example of overfitting, of mistaking the measure (solving hard math problems) with the goal (making math more accessible). Destruction parading as democratization.
But at the same time, this must be what the original Luddites felt, right? Surely they also recognized that some ineffable quality of craftsmanship was being lost and worth persevering? The wound inflicted by the industrialization of mathematics may also fade with time, and future generations may shrug in their history classes and go back to prompting the models to “do a breakthrough” the same way kiddos today will scroll through shein while ignoring the assigned textbook reading on how people once destroyed textile mills. Assuming that there is still something that looks like a history class, of course.
IV. The title of this piece is derived from AI as Normal Technology, an essay written in April of last year. It’s an important essay, with an important goal. In their words:
We articulate a vision of artificial intelligence (AI) as normal technology. To view AI as normal is not to understate its impact—even transformative, general-purpose technologies such as electricity and the internet are “normal” in our conception. But it is in contrast to both utopian and dystopian visions of the future of AI which have a common tendency to treat it akin to a separate species, a highly autonomous, potentially superintelligent entity.
…
The normal technology frame is about the relationship between technology and society. It rejects technological determinism, especially the notion of AI itself as an agent in determining its future.
Many people framed this article and the people behind it as an answer to / in opposition to AI 2027, which is much more focused on the risks of superintelligent malice. For what it’s worth, even though I am concerned about our ability to control the bots…
…I think it is extremely important to also think about the risks of AI that fall outside of the terminator-skynet-extinction variety. This is not an either or thing. I’m concerned about both of these outcomes!
The main premise of AI as Normal Technology is that we (as a species) will have time and capacity to intervene because AI will not FOOM. Importantly, the authors do not contest that AI will continue to acquire new capabilities indefinitely; rather, they think that our institutions will have time to respond to those capabilities because the AI tooling will remain in our control as those capabilities increase.
My problem with this framing: AI is an abnormal technology on multiple axes. The whole ‘AI may have a mind of its own’ thing is just one of those axes. I don’t think anyone in the Valley is really grappling with what it means to have AI tools that are better than humans, what it will do to our global economy and how we value human labor. The authors imply that we may be able to put policy controls such that AI does not immediately take over all the jobs. But, bluntly, there is no obvious control mechanism that will prevent organizations and governments from using AI.
Tao’s article above purposely uses the language of alignment applied to the AI industry. That is because the AI tools we use are not the only optimizers out there. This is something I explored more fully in The Optimization Theory of Everything. Corporations and governments and individual people are also optimizers with their own incentives, all of which encourage defecting when it comes to using AI or not. Every boardroom in the world will scalp any CEO who states that their firm will not use AI for ethical reasons, and every governing body in the world will shoot down any proposal to gut their AI industries if other countries continue pushing forward. Short of a global agreement to avoid further development, I don’t really see the use of AI slowing down, and I do not want to rely on the largesse of business owners to not ruthlessly deploy AI in their companies.
So AI is a normal technology, except it can do everything, can continue growing its capabilities, and can eventually outperform human labor in most tasks. Which is to say, it is nothing like any technology we have ever seen before, ever.
IV. A year ago, I wrote my first Meditations on AI post. It’s worth revisiting here:
If you were a clarinetist back before the phonogram really took off, there were a lot of jobs and they paid well. Any bar, restaurant, theater, or dance hall needed live musicians to have any music at all. Now, it’s a desert. No one wants clarinetists, because no one needs live orchestra when they can ask the genie in their pocket for whatever music whenever they want. I’d wager that you have way more people able to make music today than you did at basically any point in the early 1900s. But the ability to make a career in that space has plummeted. The barriers to entry have dropped, taste is way more important, and there are only like 3 clarinetists who matter enough to make anything close to a middle class living on taste alone. Sure, every now and then you’ll get an Elton John, but, like, you can’t exactly plan for that! This is the worst case scenario for what AI will do to all jobs everywhere, except it’ll be even worse because it’ll happen over like 2 years instead of 100.
On a semi-related note: earlier we talked about what AI did to the graphic design sector. One thing that I didn’t mention is that the ‘true’ artists, the senior types who have been in the industry for a while, are thriving — those guys were primarily valued for their taste anyway, and now that is even more valuable. There’s an important pattern there: generally seniority and taste are highly correlated, so we should expect AI to disproportionately hit junior, freelance, and entry level positions. The more senior you are, the more insulated you are, the more you’ll benefit.
People on the pro-AI side of the debate argue that the existence of automated general intelligence will massively increase quality of life, and I think history suggests that they are right (again, caveat that it doesn’t kill us all). But in the short term, I think we should expect some pretty significant labor shocks and a possible increase in wage inequality.
As for the medium to long term, who knows? In my opinion, the nature of work, all work, will look fundamentally different, and it may look very unpredictably different. A musician a 100 years ago would never have been able to predict Spotify or DAWs; what makes you think anyone can predict what comes after AI?
We know what happens in industries where the fundamental skill of the thing becomes super cheap and accessible. The future of the global human economy is rockstar musicians alongside a vast sea of people who cannot make a living making music. Traditionally though, that does mean more access — more people have the opportunity to get into music today than ever before, they just need a day job. But does that still work with AI? Maybe you can turn math into a hobby and keep a day job as a software engineer, maybe that means more people can do math than ever before. But the AI also does all the software engineering, so then what?
Going back to the very top, what are the things that are uniquely human that no AI can do? Not now, not in twenty years, but ever? The best answer I can come up with is “be human.” But that on its own does not provide much comfort.