I am already permanently underclassed (and other thoughts on the impending AI apocalypse)

When all about you lose their heads, accumulate capital and take cheap Chinese peptides

Spare a thought for the permanent underclass. They’ve got it hard. Disempowered, disenfranchised, unloved - their very reason to exist has been obviated. They are vestigial, pointless bags of meat waiting to be made extinct by our robot overlords.

Oh wait. You don’t know who the permanent underclass are? Well, here’s the rule about the permanent underclass: if you don’t know who the permanent underclass are, you ARE the permanent underclass. Also, if you DO know who the permanent underclass are, you’re still the permanent underclass. We’re all the permanent underclass. Isn’t that a lovely message of unity and shared destiny for the human race? We can all be permanently underclassed together, forever.

Like these first two paragraphs? Give me your email, pls and thx

The permanent underclass meme is a big deal in silicon valley. You can read about it here and here. As AI gets so good it becomes super-intelligent and replaces all our jobs, humanity will slowly be cut out of the loop of the economy, and all but a small cadre of elite (if even them) will be left wealthless and pointless.

In this world, the AI gets so good, it starts making itself better without human input. Then it starts building robots and commanding those robots to build more and better robots. Before you know it, the AI is recursively self-improving its intelligence and physical capability through giant robotic mega-factories.

Remember when tech people were optimists about the future? Good times…

I’m not here to tell you these ideas are right or wrong cause, ya know, I don’t know. But man, I’m p***ed off. I’m p***ed off about all the people who take this meme too seriously and I’m p***ed off by all the AI skeptics who don’t take this meme seriously enough.

So I thought it’d be a good idea to dump all my stupidest hot takes into one document so our future AI overlords could laugh at my ignorant inability to predict their inevitable victory. And then I thought, well, why not share said document with you, my oft-tolerated readership?

So here it is: a large, baggy mess of an blog. I’m just going to lob some grenades out there for y’all and you can tell me if I blew anything up.

Before I do that, a bit of house-keeping. There are three themes I’m addressing here, and I appreciate they’re likely to get muddled throughout, so I’m going to tease them apart for you up front to help make the rest of the article clearer.

The first is: what happens to the economy if labour-replacing AI occurs? This is the most concrete theme, in that it has clearly stated assumptions and goes from there using established economic theory. On reflection, I’m extremely sympathetic to the people exploring these ideas.

The second is: is it actually feasible to develop broad-based labour-replacing AI? I’m in two minds on this one. It may happen, it may not, but either way, timing will be very important.

The third is: should we be immanently concerned about this permanent underclass business or is it mostly hot air?

With that said, I’m going to jump into a post which, I remind you, is mostly hot takes and reactions, not a deep settled opinion. Here goes.

1. I am a pessimistic meta-inductor

Pessimistic meta-induction is a fancy way of saying “hey dude, we were wrong before, why is this time different?” Here’s a great article that sums up all the patently absurd ways people in the past have worried about the role of machines in taking jobs.

Here’s Keynes talking about it during the great depression:

We are being afflicted with a new disease, technological unemployment.

And Einstein:

According to my conviction, it cannot be doubted that the severe economic depression is to be traced back for the most part to internal economic causes: the improvement in the apparatus of production through technical invention and organization has decreased the need for human labor, and thereby caused the elimination of a part of labor from the economic circuit and thereby caused a progressive decrease in the purchasing power of the consumer.

I don’t know about you but I don’t exactly feel a progressive decrease in the purchasing power of the consumer (ie me) relative to my grandparents. I can buy runner beans from Kenya and coffee from Brazil and phones from China, and fridge-freezers and microwaves and ovens and central heating and fans and non-stick pans and projectors that once would have been the preserve of large cinemas. And that’s just a Monday evening shopping spree from the comfort of my sofa.

The point to ask is, why is this time different? Everyone who ever invents a new technology seems to compare themselves to Prometheus, thinking they’ll be punished eternally by the Gods for their sins. This, erm, never really turns out to be the case, perhaps to the disappointment of the technologists who (somewhat egotistically) imagined themselves to be earth-shattering prometheuses (promethei?).

Technology is just that - technology. And like all technologies throughout all time, it creates some winners and some losers, it does some things better than others and it has profound transformative effects on our society - but so far, no end of the world.

The current crop of AI will no doubt do that, is doing that, will continue to do that. But LLMs are just another tool in a long line of tools, sometimes being built by a bunch of tools (okay, that was a cheap shot - sorry Sam and Dario!). Yes, we’ll see big changes, and it would be flippant to compare this technology too directly to, say, a typewriter.

But, yeah, basically it’s like a typewriter (or maybe a word processor). Okay, that was flippant of me (how dare I be flippant). What I mean by that is, it’s making some people more productive, and putting some people out of a job (remember when people hired other people to type things up for them? Old times…).

Maybe one day we’ll have machines smart enough to replace all of us. But reader, that day is not today. And since we’ve been warned that that day will be tomorrow many times before - maybe we should treat such warnings with a healthy dose of scepticism?

2. I am already permanently underclassed

In the permanent underclass world, human beings will no longer be able to control the systems because the only alternative to doing what superhuman AI suggests is to opt out, and those who opt out will lose, because superhuman AI is smarter than you, you dumb meat-bag loser.

One interesting question is, is this actually that different to the world we live in today? Hot take: maybe in degree, but not in kind.

If you’re lucky enough right now to have a healthy amount of money in the bank, you’ve got a choice to make. You can spend it, keep it as cash, or invest it in some kind of investment vehicle. Spending it is generally frivolous and keeping it as cash is a mental illness mostly localised to the British Isles.

So you’re investing. And unless you have a lot time on your hands, you’re going to struggle to actually make good investments yourself. Besides, passive funds perform better than active funds these days anyway.

Put simply, you’ll throw your money at daddy Vanguard and let them do the investing for you. This might as well be you giving your autonomy over to AI. Probably Vanguard are using AI somewhere along the way to help manage their portfolios (this is just a claim, no evidence - but come on, man!).

Sure, you can choose not to give your money to Vanguard or any of the other big fund managers. But you know what happens if you don’t? You end up poorer. Do you want to end up poorer? No? Okay, then give them your goddamn money!

Across our modern, always-on consumerist society, the idea that people live with total autonomy is already an absurdity. Our decisions every day are directed by tech companies who spam adverts at us through the hell-rectangles they sell to us, which we carry around at all times in our pockets, because the only thing worse than having a hell-rectangle in your pocket is not having a hell-rectangle in your pocket. You want a place to eat food? Google Maps will tell you best local places, and buddy - it ain’t giving you unbiased recommendations. You think that’s not AI at play there? Dude, come on!

Our lives are already uncontrollably managed by giant systems, both of an economic and a technological kind. Wake up and smell the coffee everybody: you’re not autonomous, and you haven’t been since the rise of the modern global economy. We’ve built a system of containment and everyone knows it. That’s why it’s so laughable when Andy Burnham suggests we shouldn’t be in hoc to the bond market.

Besides, all this hand-wringing about autonomy, control and ownership is silly anyway. What even is autonomy? What is free will? We are all of us always prisoners of the environment built up around us, and as the modern world becomes ever more optimised and spread out in ownership between many people, our scope for control over aspects of it diminishes.

Freedom is an illusion, except insofar as we can have freedom from systems which artificially constrain our actions. I’m not going to go into this fully here cause, you know, I could write a whole book on this topic and still not say much of any use. But the permanent underclass thesis isn’t necessarily claiming the robots will rob us of our free speech, or our freedom to have sex with whomever we want, or our right to eat and drink and travel as we see fit. Just that the robots will structure a choice architecture that will manipulate you in ways you can’t even understand, and plausibly to the advantage of someone other than yourself, so that you are slowly made poorer and less relevant as those who own capital take control.

But, hey - with rising inequality and stalling living standards in some of the developed economies, that’s the world we live in anyway! If you’re going to critique AI on these grounds, then you should also critique the iPhone, and Instagram, and the executives of JP Morgan.

The point here isn’t that you shouldn’t be worried about the permanent underclass scenarios. The point, rather, is that concerns about capital dominating the economy through mechanical means are well-founded. But they’re not very new.¹ So to treat this as some categorically different concern rather than, say, a continuation of the sorts of concerns people on the left of politics have always had about capital’s control over our lives, is in my view incorrect. I mean, some dude famously wrote a book on this topic…

3. What the heck is super-intelligence anyway?

I mean, seriously, what is it? Great at maths? Brilliant at memorising every poem ever published? The world’s finest draughtsman? Lots of humans meet these requirements - sometimes we call them autistic savants, and they’re often bad at grappling with human emotions.

It’s important to recognise that there is no one linear concept of intelligence. IQ does not measure everything. Though, as I’ve written before, try telling that to the people of silicon valley. It’s often a trade-off. Get good at one thing, and you end up getting worse at another. It’s not obvious (to me, would love to see some arguments on this!) that intelligence can tend to infinity. Electricity can only travel so far so fast. For real time processing, there’s a limit to how large a brain can get before latency becomes an issue. To be exceedingly flippant (the theme of this article, and the right tone for Substack): the sperm whale has a brain several times larger than a human’s, but we hunted those suckers almost to extinction (looosers!).²

This notion that there’ll ever be a vaguely God-like, omnipotent super-intelligence capable of everything is very engineer-coded, and fails to face up to a fundamental reality about the world: there are some things that don’t have answers, and there are some situations that can’t be “solved” in any meaningful sense of the word. What would it even look like to have a creature of inconceivably greater intelligence than our own? Well, it’s inconceivable.

This argument doesn’t have as much relevance to the more hard-edged economic arguments out there, which mostly focus on general labour replacement as opposed to more sci-fi AI overlordship scenarios (though I still think it’s relevant in that a non-negligible amount of human labour involves resolving, rather than solving, insoluble situations). But as to the doom-mongering, “AI will be so smart we won’t be able to cope” kind of argument, I think it is relevant.

If extreme intelligence were a driving factor of one’s ability to manage and control people, and to rise to the top to lord over others, we surely ought to have our best and brightest in charge. But Einstein ain’t running the US economy. Just saying…

And I know the arguments contra. Recursive self-improvement. More data. Better robotics. So on and so forth. But data can only take you so far. Yeah, the machines might know more maths than us. Yeah, they might solve nuclear fusion and build ten times more effective space ships. Heck, maybe they’ll even crack the light barrier. That doesn’t mean the machines will always be able to lead or control the humans around them - you ever seen a maths whizz captain a sports team?

Obviously, yes - a very smart, systems-enabled intelligence may well conceive of a systemic way to control human behaviour, and that’s for real scary. But that’s not the same as the quasi-religious belief in supreme intelligence. Any such theories need to be grounded in real, feasible engineering possibilities.

4. The relevance of existing economic models diminishes if the economy greatly changes

One of the arguments around economic disempowerment (as put out by Phil Trammell) goes like this: “Here’s my economic model for the economy as it currently exists, and here’s what it says will happen if I invent a technology that will so vastly change the economy it will no longer exist in anything like it’s current form.”

Okay, I’m obviously being very harsh, flippant, and unfair here (sorry, Phil!). But I do think it’s a little bit reasonable to critique economic arguments about world-destroying AI on something along these grounds.

People often forget that economic theory is not founded in natural law like physics (and btw, even “solid” physical theories turn out to be wrong eventually once you enter different regimes - cf Newtonian mechanics, Quantum, GR). It mostly doesn’t describe universally true concepts, because the economy is something created by people. The more theoretical economic theory becomes, the more detached it is from basic principles of how the world works, and the more it becomes contingent on very specific aspects of how society is currently structured.

You start relying on ideas like “capital” and “labour” which, sure, seem like really obvious concepts to us. Except what constitutes capital is defined by our concepts of ownership, and the laws that surround them. Those laws could change - arguably will change if world-altering autonomous AI is brought into creation. And sure, labour is labour is labour is labour, right? Except, you know, when you change your baby’s diaper and no one pays you to do that work. So is that labour as a share of the economy?

Okay, I’m being a little silly, obviously. But I think this is a directionally reasonable point to make. Economic theories about the fungibility of capital and labour and its impact on the economy are laden with assumptions about capital, labour and their role in the economy that won’t maybe hold up if all of human society collapses under the weight of labour-eradicating machines.

And if you’re thinking this all sounds like pessimistic meta-induction all over again - well yeah, it basically is.

Theorists are often wrong. Past behaviour is no guarantee of future performance. Thomas Malthus theorised the inescapability of the population trap, just as human society managed to escape it.

Of course, in this particular case, it’s more like optimistic meta-induction. What I’m saying is: hey folks, I know this theory sucks, but we’ve been wrong about the future before!

5. Smart people aren’t right forever

This meta-inductive point is also relevant concerning the people who are making these predictions. I hear a lot of friends say things like, “well, the smart people in AI who have predicted the progress of AI so far, even though other people were saying they were mad, are now saying this, so you should take it seriously.”

And I do take it seriously, which is why I’m writing a taxingly large number of words about it. But “dude was right about things before” is not a robust argument. There are lots of dudes throughout history who were right about things before, who get ahead of themselves and end up being wrong about things to come.

Leopold Aschenbrenner compares these early visionaries in AI to Oppenheimer with the A-bomb. After prophetically predicting the possibility of the A-bomb, and then building it, Oppenheimer spent the rest of his life arguing that the same technology meant that humanity’s only hope for survival was the obsolescence of war. Which, er, hasn’t happened.

Nerds who invent new technologies are often very good at predicting the trajectory of those particular technologies, but very bad at understanding their likely social and political consequences. Probably because science nerds are bad at understanding social and political systems in toto (which is why they rarely win public office, or even succeed in getting their candidates elected to public office).

Want to find out just how wrong (questionably) smart people can be? Subscribe to my Substack!

6. Pascal’s wager is still a weak argument

Some people’s brains have been totally cooked by these models, to the point of religious belief. These are people who legitimately argue that the only hope for you (or for anyone) is to join one of the big AI labs, accumulate capital and hope that you just about scrape into the not-permanent-underclass group in the future, where you can live forever on cheap Chinese peptides and electrolyte-infused kombucha. The thinking goes that, even if there’s only a modest chance this turns out to be true, it’s better to be wrong about it and live a life in the AI industry, than be right, choose to become an artisanal carpenter and be completely ruined by the future economy.

Nice argument. Pascal made a similar one about believing in God. Thing is, it doesn’t really work. Just as you could respond to Pascal with the question “which God?” so too you can respond to these people with the question: “which future?”

Even taking it as read that these economic models are correct, they don’t predict what future will actually arise - only what future would arise under specific circumstances. Yet for all we know, the world could blow up tomorrow because Mr President decides he likes firing off nukes. Climate change could turn out to be even worse than we think and swamp our major cities under the ocean. An asteroid could collide with the planet.

There are a million-billion possibilities for the future of humanity out there. I think it’s essentially impossible for us to assign accurate probabilities to most long time-horizon events. It’s not just the known unknows, it’s the unknown unknowns too (this sentence was not written by AI fwiw).

Preparing for retirement? A sensible idea, for sure. You’ll either get old or you’ll die, and either way retirement saving seems sensible. But preparing for our robot overlords? “Even if it has a 1% chance of being true…” they say. Well sure, but what if there’s a revolution that slaughters anyone who actively pursued the creation of our AI overlords? And what if that has a 2% chance of being true? Which approach looks better now, eh?

You can’t play dice with the future in this way. Not always, at least. That’s why we have non-utilitarian moral theories. Not every life choice can become a calculation. Sometimes you have to be guided by what you think and feel is right, numbers notwithstanding.

7. When is as important as what

A lot of the most fevered concerns around this are phrased like this: imagine if we suddenly had a technology that replaced all human labour…

To which I say: well I never! Imagine if pigs could fly! Imagine if we instantly cured all disease! Imagine… imagine… imagine if 1 trillion lions fought the sun, who would win?

The point is, thought experiments might be fun to play around with, but they also avoid obvious realities. Pigs can’t fly, it’s impossible to instantly cure all disease, and obviously a trillion lions can’t fight the sun. I mean, how would they even get there? Lions don’t have wings. Think about it, you idiot.

To be clear, I think it’s within the realms of physical possibility for us to, and therefore quite plausible that we could, develop superhuman machines that would replace most human labour as currently conceived.

But, you know, there’s a fair amount of time between now and then, and that development won’t happen as a nice linear process. We won’t wake up tomorrow and find our local barista replaced by a robot (well, a small fraction of us might).

Since this technology will take time, and since at each stage of development, the economy and the law will have to adapt to the latest changes, it’s really very hard to say what changes in trajectory might occur. Just because things are physically possible, does not mean they have to come to pass.

Some obvious examples: governments could implement hard restrictions on what gets done. There could be mass uprisings or societal breakdown beforehand which causes the technology to stagnate.

To be clear, many of the more serious people take these possibilities into account. But less often taken into account is this idea that, even within the concept of “super-human robots taking our jobs and running most of our economy” the shape of this world to come is under-defined. There are many different versions of how this plays out. It might be easy to get worked up about the most disruptive thought experiment possible - instant replacement. But the less disruptive version - that it happens piecemeal over time - gives us the theoretical ability to design towards success.

People are always going to turn things into simple binary, a la Pascal’s wager. But the world doesn’t work like that. There isn’t a binary moment in time when any one of us will be asked to usher in the AI overlords or throw away human technological progress. The reality that unfolds will exist somewhere between those two poles. Time may present us with the opportunity to land in a sensible middle ground.

And I want to say at this point - yes, some of the people theorising the worst case scenarios are doing so precisely because, if the AI labour replacement train starts accelerating, they want to motivate the case for designing towards success. They want policy-makers to take these ideas seriously and to start preparing for this possibility. That’s great and I support it.

I just don’t support the people who take these thought experiments, wrap them up as fait accompli statements of the how the world will actually be, and use it to foment doom-mongering and hopelessness about the future of humanity. Or even worse, use it to shill their technology products and pull the wool over regulators’ and retail investors’ eyes.

Of course it’s generally true that we live in a world without conceivable physical barriers to labour-replacing technology. If nature can make entities of human-level capability, then given enough time, so can humanity.

I’d even agree that, charting the path of human development since the end of the Malthusian epoch, we should logically assume that new technologies will over time arise that will continuously rewire our economy further away from Labour output. This does sensibly imply that capital will grow as a share of the economy, and those who are rich in capital rather than labour are advantaged by the change.

Some of the folks working on this are raising it as a warning, to highlight the importance of good redistributive approaches to capital. I think that’s great. I’m not here to deny the possibility of this thesis. It’s a compelling thesis. A lot of smart people have varying degrees of faith in it, the same smart people who had a lot of faith LLMs would get as good as they got when others dismissed them as stochastic parrots. But nevertheless, it’s just a thesis! And so I want to pour some cold water on the more excessive opinions that people are starting to develop in this area.

Sure, in the long run, nothing in physical law stops us developing this technology. But as Keynes acerbically said: “In the long run, we’re all dead.” Timing is important here, the same way timing is important in markets. Which is to say, being right at the wrong time is the same as being wrong.

We do not yet have AGI. And whilst we’re making great leaps forward in robotics, we don’t have machines that are yet fully autonomous in a way that could broadly replace human labour in the economy.

It’s fine to think we might do some day. Our rate of progress towards that may even accelerate over time. We ought to study this, understand it, and prepare robust policy responses to any such developments. Great stuff. Applause, applause.

But it’s less fine to messianically believe that we will do in short order, or that we ought to radically alter our personal behaviours or attitudes right now in anticipation of that future, abandoning any hope that humanity will, you know, muddle through like it always has done.

One transformer does not an AI overlord make. It’s not rational or sensible to whip up fear about that possibility, and then to stake trillions of dollars of the global economy on AI capex, on the thesis that anyone who doesn’t invest right now will end up part of the permanent underclass. That kind of binary thinking is dangerous, and leads to irrational responses.

Again, this isn’t to say I don’t think Anthropic (to take the trending example) is a good business, because I think in many ways it is. Their coding products are fantastic, and Claude Cowork is, in my experience, modestly good at automating grunt work.

I just want to say, gently, for the record, that if you’re investing your time and/or money in Anthropic because you think otherwise our robot overlords will make you part of the permanent underclass, rather than - say - because it’s a good business with strong product-market fit and growing revenues, then you might want to seek out a new financial advisor.