AI sucks. Hating it is not enough.

I do not want to be writing this essay. I, like you, am sick of think pieces and discourse about AI. And yet I have seen my communities, my friends, tearing each other apart for months: bitterly divided on both the bare facts and the correct response to AI. And I, I feel I must write on it¹: in my own voice, in my own name, in public.

I, and others like me, have been accused of ignoring the moral questions around AI; placing them to the side in the same fashion as those who want “politics” out of open source, out of science, out of games. That, unfortunately, is not true. I cannot get those questions out of my head.

Dear AI hater, we are fighting for the same cause. I know you may not believe me. I know I will likely fail to convince you. But please, please listen². I cannot do this alone.

In brief: AI risks terrible harm to the world. And yet we cannot uninvent it, we cannot boycott it out of existence, we can barely regulate it in any way that matters. The urge to do so is understandable, laudable and one that, on some level, I share. The problem is that while these things might help, they will not work. We must navigate the facts as they stand, not as we wish they were, and set our sights much higher, even when it feels impossible.

I know this sounds like that refrain you have heard so many boosters sneer³ in the past: “AI is inevitable; get on board or get left behind”. Your skepticism is warranted, and I shared it. But the things I have seen and heard, with my own eyes and ears, tell me that something very real is coming. I must warn you, even if I fear you will not believe me.

These tidings should not ring triumphant; instead, I wish to impart upon you a bitter hope. There is a dark shape on the horizon that we must not ignore. We must fight to shape the future in this time of monsters, even if we cannot uninvent AI. We must adapt, even if we do not need to adopt.

When I have made these arguments in the past, those who do not know me presume that I am a booster, a shill, a fascist, a fool. It is true: I am a technologist. A programmer, a scientist: fascinated by statistics and machine learning, experimental design and automation. A transhumanist⁴, who sees glimmers of personhood in animals and machines, who wishes everyone could freely shape their own bodies, who wants to expand our moral circle in ways that seem heretical.

And yet I am an ecologist, an environmentalist. I gave my whole life, my wealth, my health, my future to saving the earth. To fighting climate change and promoting cycling, to understanding the beautiful world around us. And when I could no longer walk that path, when I became so sick that I thought I would never again walk through the forests, marvelling at every fallen log and mushroom, I tried again. I found a new path — of machines and code; of desperate survival and brutal compromises; of hard-fought independence through careful reliance⁵. And when I could, I gave again: sacrificing years of time and embarrassing amounts of income to give back to the commons, in the form of open source.

I have only one life to live. I’ve always known it would be given in service⁶.

Do not tell me that I do not fight for a brighter future. The world is not so simple.

Footnotes

1. It would not be my own voice if it did not have far too many words. There are too many branching lines, too many holes to plug, and I will, this once, commit the cardinal sin of persuasive writing: arguing defensively. ↩

2. Grant us your charity, as we grant charity to those who trespass against us. ↩

3. That’s negative polarization for you. Surprisingly powerful stuff. Be careful with what you say, and with what you consume. You never know how it might change you. ↩

4. Eclipse Phase was formative in my vision of transhumanism. Technology as transformative liberation; not a technofascist project to create a master race. It, like “liberal”, is a charged and complex term, but I refuse to let others take those identities — or the em dash — away from me. ↩

5. The em-dashes in this essay, and every word you read here, are certified human-produced — copy-pasted from Google the good-old-fashioned way, and typed with my little meat fingers on a mechanical keyboard. ↩

6. “I’ve sacrificed everything. What have you given?”. Illidan was kinda spittin’… ↩

AI sucks

AI, as it is wrought, is an unfolding disaster. The worst people in the world are riding it to power. It empowers terrifyingly accurate surveillance on an unprecedented scale¹. The tools that AI companies glibly sell are destroying entire industries, threatening to displace billions of workers with a willful disregard for their well-being or future. Built on the backs of exploited labor and creators that they will never credit, the machines serve as perfect twin engines for propaganda and censorship. As we use them, our skills degrade, fostering an ever-so-monetizable reliance. And it all consumes massive quantities of power, emitting carbon we cannot afford.

These harms are real². They are tremendous. And if you are reading this essay, you almost certainly already agree with me on that³. Go read something else if you want to dispute them.

If AI works, it threatens to disrupt the world in terrible ways. If it does not, if it’s all a scam, a bubble — we are in for one hell of a crash. Both, to put it mildly, seem Bad.

In light of that, it makes sense that many people are refusing to be complicit — choosing to become AI vegans. Is that enough?

Footnotes

1. I would like the personal AI companions and biohacking parts of cyberpunk, not the crushing surveillance state and eye-watering wealth inequality please! ↩

2. Some purported harms are mostly not real, or very speculative. Heat island effects and infrasound harm are effectively pseudoscience. Water, noise and land usage concerns are relatively low-consequence and fixable, relative to other industries. AI safety X-risk concerns and sentient machine life slavery are somewhere between worrying and delusional. But these quibbles don’t matter: the very real harms are enough for me to say that this technology is going to cause serious problems for billions of people. ↩

3. Being generous to your opponents is important. Almost no one thinks these harms are fake, even if they say they do. They’re not stupid. People who are pro-AI, even in the strong boosterism sense, just think that those harms are worthwhile or unavoidable. ↩

Boycotts, data centers and NFTs

Everyone, to a first approximation, can see that AI sucks. These harms are not esoteric: they are plainly evident, widely publicized and boldly advertised! Gleefully crowing about how your product will destroy lives does, in fact, make the public hate you!

And so opposition has risen: across the left and right, across lines of gender and class and race, across the world¹. The primary weapons of the self-proclaimed AI haters are three-fold:

1. Encourage boycotts of AI technology and its products
2. Delay and block the construction of data centers
3. Make the case that AI does not, will not, and cannot work

The argument for boycotting is simple. AI is doing tremendous harm: as consumers, we should avoid giving these companies money, and turn our noses up at the free AI features shoved down our throats. We should boycott the games and movies and products and open source libraries that use it, and support human artists.

We can make lists of slopware, review bomb games on Steam, complain loudly in the social media comments about new AI features, and disparage the technology to our friends. Capitalism needs customers, and with sustained pressure and a critical mass, these campaigns can kill “promising”, harmful or worthless technologies.

3D televisions and AR glasses, NFTs and New Coke, GM food and apartheid. All were pushed heavily by greedy corporations or powerful states. All were defeated, in large part, by concentrated consumer contempt.

Why would AI be any different? I have been told, convincingly, that AI is not inevitable².

To block data centers, AI haters bring one of the West’s most powerful weapons, NIMBYism, to bear on the problem. With a mix of environmental and social concerns, local groups have rallied broad coalitions: winning delays and stalling out the construction of data centers across America, Europe and the West more broadly.

There are real victories, which sometimes prevent real harm. Colossus is a true monstrosity, spewing pollutants and noise in a dense urban environment, in blatant defiance of local governance. The Stratos project threatens to increase carbon emissions for the entire state of Utah by 50%, by building 9 GW of new gas power plants! These projects are terrible, and need to be stopped.

But, if you listen to the most prominent AI haters, to people like Timnit Gebru and Emily Bender, the problems run much deeper. AI, they claim, is fundamentally a scam. They argue, with papers and PhDs and years of industry experience, that LLMs simply regurgitate their inputs³; mere stochastic parrots. No true understanding, ergo no true value. It is, after all, mathematically proven.

Ed Zitron carries this argument into the world of business, using detailed financial analysis to make the case that CEOs, in their collective AI psychosis, are pushing this technology on us, inflating valuations and sucking up capital before offloading it all on the stock market suckers. Their services are sold at a loss, propped up by VC dollars, and there is no real market for them — only a tangled web of hype-driven contracts and self-dealing.

And indeed: like any new technology, AI has attracted morons and scammers like flies. CEOs are clamoring (“we have to pivot to AI!”) even as they salivate (“we can cut payroll by 80%!”). Most of the new AI “features” are worse than useless, and most of the AI startups will fail, including some shockingly large players.

Footnotes

1. This, of course, is not particularly true. Consumer opposition to AI, and especially to data centers, is a mostly WEIRD phenomenon. Americans, Canadians (me!), and Europeans are dramatically more negative about AI than developing nations. Still, I am WEIRD, and you, the curious hater reading this article, probably are too. The filter bubble you’re in really affects the information you see! ↩

2. The strongest arguments I have found on this front come from David Krueger, who argues against inevitabilism on the basis of real physical inputs that can be restricted (like nuclear proliferation), and Mar Hicks, who makes a compelling case that inevitabilism is a rhetorical and marketing strategy. ↩

3. We must have missed the solution to the unit distance Erdos problem during literature review. ↩

Unfortunately AI does work

If you have only ever interacted with free-tier models — ChatGPT in 2023 and Google’s god-awful search helper and random crap that companies are throwing in to be hip — it’s very easy to believe the stochastic parrot narrative. That’s an understandable position to be in: boycotting AI tools is both rationally and morally justifiable, especially if that’s all you’ve seen.

If that was all there was to it: you would be right. AI would be a horrifying bubble, a technological dead-end, hyped by suckered sycophants hopped up on synthetic sycophancy.

But that’s not the end of the story, and the next winter is nowhere in sight. As a programmer, as a scientist, as a writer: I have seen what these tools can do first hand.

Effortless generation of working (if mediocre) code, nearly-free code review that spots real problems (and many imagined ones¹), troubleshooting of horrific Linux driver problems that would waste days (70% of the time, it works every time).

Instant imperfect literature review, superpowered semantic search, statistical expertise that is better than every single science grad student I have ever met².

Copyediting is nearly free now, and universally accessible. The machines can follow your arguments better than your average social media user, pushing back against you or strengthening your points, but only if you think to ask. You have access to endless, inaccurate facsimiles of test audiences, limited only by the creativity of your chosen lens. Every author can have a lore master of their very own, cross-checking every fact and character quirk.

Imperfect tools that will change the world.

Don’t get me wrong: simply asking the machine to write for you, to work for you, does not work well³. The promised robotics revolution⁴ is not here yet, even as the videos from China grow more unnerving by the month. Labor broadly, the kind that makes your brow sweat, may be safe for now. And yet AI, in a weak but very real sense, works.

Laziness never pays off: this is not the singularity, the end of thought and craft and toil. AI cannot replace me⁵. Not yet.

My friends and family have tried these tools and, overwhelmingly, said the same. They, like you, hated AI. They didn’t think it worked, not for real. But as they tried it themselves, they changed their mind⁶. Many of them still hate AI⁷. But their calculations of its trajectory and the best ways to fight it have seriously shifted.

I know you don’t believe me, that this is just anecdata, written by someone who must be lying or who lacks real expertise. Someone who wants to be fooled.

But you can spend $20⁸ to test it yourself: one month with Claude Code or ChatGPT or Kagi Assistant, using it for something that you genuinely know well⁹. You will see failures, you will see flaws. Carry that skeptical attitude into this trial, but do not blind yourself to the strength of your enemy¹⁰. Ships sink and crops fail, and yet both changed the world.

I know this is the grifter’s move: to say that you have to try these things. But I do not know how else to show you. Anecdotes can be countered by anecdotes, studies by studies¹¹, mathematical proofs (of uselessness) by mathematical proofs (of breakthroughs discovered by LLMs). I cannot ask you to trust my evidence. Please trust me enough to gather your own, the good-old-fashioned way.

There is a wave on the horizon and we are standing on the shore. Don’t take my word for it: borrow a spyglass, and check for yourself.

Footnotes

1. A popular argument is that the presence of hallucinations renders these tools useless. Alas, humans (and books and peer-reviewed articles) suffer from errors too, in surprising and nonsensical and infuriating ways. You must build quality and resilience into the process; demanding perfection simply doesn’t work. ↩

2. Admittedly, a low bar! ↩

3. I love the craft of these things; they are part of who I am. I will not stop writing and programming and doing science, even when I am obsoleted. ↩

4. AI, in the form of LLMs and modern computer vision, is primarily a threat to knowledge work. It will accelerate other forms, but without embodiment the impact is limited to “merely on par with the internet”. ↩

5. The machines are much more effective with a skilled operator (currently). Both taste and domain expertise matter a lot. You have to be able to carefully evaluate the output, and decide what’s worth building in what order. Firing me tomorrow and replacing me with an agentic LLM would lead to a huge mess, and cost a lot more money hiring me back! ↩

6. While writing this article, I have spoken with dozens of colleagues and friends about AI. Exactly one of them has serious hands-on experience with AI and thinks it works poorly, disputing that it will change the world. Maybe you’ll agree with them. Maybe they’ll be right. ↩

7. Maybe I hate AI too. I can’t tell. ↩

8. If $20 is a real hardship for you (and it has been for me in the past), shoot me an email. If we’ve interacted before I’ll spot you it. I mean that! ↩

9. I know, there is real harm even in a month of personal use. But knowing your enemy is important: if you are to fight back you need a clear-eyed view of the genuine strengths and weaknesses of these tools. ↩

10. In Emily Bender’s own words: “I make a point of not reading synthetic text”. The stochastic parrot argument was put forward in 2021, and continues to be used in the face of dramatic changes in capability. I would ask if the parrot is dead, but apparently that’s considered quite rude! Must be merely resting. ↩

11. Alas, the scientific process is not infallible, and evidence is not the only factor in changing minds. Teenage Alice would be very annoyed. ↩

Hating AI that works

Suppose we concede¹ that AI works, that it genuinely, honest-to-god functions in a June 2026 way. What becomes of the weapons of AI hatred?

Perhaps things won’t be so bad. Technologies form S-curves: they all hit diminishing returns and level off. Despite science fiction concerns of unbounded recursive self improvement (the vaunted Singularity), there are real mathematical, physical and sociological challenges to that hypothesis. Perhaps we’re already there, and this is just about as good as it’s going to get.

Unfortunately, widespread disruption, widespread harm, does not require further technological breakthroughs. The future is already here, unevenly distributed. That future will diffuse: through the programming industry, into low-skill white collar work, into the daily fabric of everything anyone does using a computer.

For all of their talk of the urgency of rapid reimagining, corporations and institutions can only change so fast — the whip is still coming down. We may not all be laid off, but things will change radically, and it’s not clear that Jevons Paradox will save my industry from a dramatic contraction.

But that’s all theory and rhetoric; what retort would a bright and well-read hater offer? Let’s begin with the most popular argument among my peers: that AI is a bubble. That might still be true. Ed Zitron’s critics could be wrong: the dodgy accounting might be too damning, and it could still all come tumbling down.

But the problem here is that such a bubble, if and when it happens, will look a lot more like dot-com than NFTs. Real value, fake valuations.

The speculation stops, the technology stays around. Reshaping our lives in drastic ways, for better or worse, just like the internet.

Cherished hater, hold your tongue a moment longer. I can hear your objection ringing in my head even as I write. Any value is outstripped by costs! These companies are wildly unprofitable! AI, you say, will fail completely if inference (the thing which actually performs the work of the models) is being subsidized by VCs and desperate corporations, just like Uber was.

And yes! Yes inference (and to an even larger extent training) is being subsidized², as companies race for monopoly control over market share and capabilities in the world’s largest penny auction. The Uber analogy is more-or-less correct: the companies shoving AI down your throat have bought into the hype, and are trying to solidify a durable, profitable lead.

Like any penny auction, the costs will ruin many of the participants: leaving them gasping for air as their adversaries run ahead and claim the prize. Ed Zitron recently released a report on OpenAI’s financials using leaked internal accounting. In classical Silicon Valley fashion, they are lighting money on fire at a truly staggering rate. Massive expenditures on hardware, on inflated salaries, on marketing that you are surely sick of. Each bid costs a staggering amount, and the competition for GPUs and labor and market share is brutal. The costs far outstrip their revenues, and they need to keep raising money to keep this up.

Unfortunately, I don’t think the numbers support the broader thesis. The thing that needs to happen for AI to die as a sector and as a technology, is for these services (not these companies) to be fundamentally unprofitable. Companies need to be selling their services (on the whole) at a loss, and unable to raise prices to the point where they are unit-profitable.

Right now, that doesn’t appear to be the case. Our best estimates of gross margins on inference (for prices so low that half of my programmer friends are preparing for a rug-pull) are somewhere between 30 and 60%. By examining Ed’s own data, we can see that OpenAI is sitting at about 43%³.

While I have my quibbles, I do think Ed Zitron is basically right: there is a bubble in AI, and it will, eventually, pop, likely to explosive effect. OpenAI will probably crash and burn. I hope they do⁴! While serving models is profitable, the companies behind those models are mostly not, and correlated failures are a terrible thing.

But when I see people approvingly quote Ed, there is a common logical leap: “when the bubble pops, AI will (basically) go away”. We’ll all be able to stop worrying about this. Sanity will be restored, and things will be okay again. In fact, because that pop is inevitable, we don’t even really need to worry now.

If we see a crash, the race will slow, model training will become an annual affair, or shift to dramatically cheaper fine-tuning. Salaries might even drop from “pro sports” to “educated professional levels”. The current era of irrational exuberance and frenzied, frantic scrabbling might draw to a close.

But capitalism’s calculus is coldly consistent: if a technology provides value above its cost, corporations will pay for it. As funding dries up, frontier models may be priced astronomically, at some double digit percentage of a skilled worker’s wage⁵. But much of the work, both good and bad, will be done by much, much cheaper open-weight models. You can, today, run a model on your phone that would have left people convinced there was a man hidden under your table.

In the end, inference is cheap⁶, and a crash only makes that worse. When the dot-com crash happened, the overbuilt servers and fiber lines did not somehow despawn. Instead they were sold, for pennies on the dollar. The survivors bought them⁷, consolidated power, and made bank: ushering in the broadband era of cloud computing and internet video that we love or hate.

A crash may come, the bubble may pop, but in the end, it doesn’t even matter⁸. If AI works, it will stumble, get up and march on all the same. Even after 99% of the companies in the race-of-the-21st-century drop out, you still cannot uninvent the technology, cannot wish away the harm.

Footnotes

1. Welcome back dutiful reader, one month later. I hope the hands-on experiment was informative! I’m sorry that I could not convince you otherwise; such a gap does terrible things to the flow of my essay. ↩

2. The inference subsidies are most apparent in the “personal” subscription plans that the various companies offer. They’re betting on a) underutilization, on average and b) these plans acting as loss-leaders, getting people hooked on cheap AI and then pitching their corporate leadership on extremely profitable enterprise deals. Same model that SaaS has used for decades now. ↩

3. That includes, by the way, the hundreds of millions of free users. The many open weight Chinese models are something like 5x cheaper than the American services for roughly the same quality, although they lag behind by a couple months in capability. ↩

4. OpenAI is not as bad as Elon Musk’s forays into AI, but they are still very very bad. You should not demonstrate willingness to collaborate with the Trump administration to make autonomous killbots and spy on Americans! You should not completely betray your founding principles and make a mockery of your non-profit status! ↩

5. In the West, that means premium AI subscriptions could rise to tens of thousands of dollars a year! In other words, near the lower end of estimates for the commercial value they actually provide. ↩

6. You can run competent-but-not-cutting-edge open models on a box in your closet, today and forever-more, for something like $10k of capital and $100 a month in electricity. That’s a very real ceiling on price and floor on capability. Oh and any safety controls are trivially stripped. GL;HF! ↩

7. The dynamics are not identical. GPUs, which make up much of the cost of modern machine learning operations, depreciate rapidly (2-4 years, vs 20 something for fiber), driven in large part by advances in efficiency and performance. The racks, buildings and infrasctructure around them is still useful though, and very cheap compute would be around for a bit before the GPUs hit end of life. ↩

8. I apologize, I am a product of my times. ↩

A budding Butlerian Jihad

If we cannot uninvent the LLMs, perhaps we can stop them. They are not some abstract mystical thing, running in an aetherial cloud. They are run on real hardware, in the real world, managed and approved by real people.

We can gum up their town halls, we can sue them in courts. We can pressure our representatives, and vote out the bozos who approve data center construction.

We can lay down our bodies in front of the bulldozers, chain ourselves to the gates of the secured data center facilities. We can slow them, we can stall them, we can stop them¹.

We can take their machines, we can smash them: gut their corpses and sell them for scrap. Claim the glorious tradition of the Luddites, and resist the machines that capture our craft.

And yes. Yes you can. Maybe you even should. It might buy us time. It might make things better, locally. You will win on data centers, some of the time.

But those victories won’t matter. The Luddites may have been righteous, but they failed all the same. The Industrial Revolution advanced, unmoved by the martyrs.

The problem with data centers is that they don’t particularly care where they are built. They need electricity, security and a warehouse worth of space. The really expensive ones, the ones that matter, are doing model training and those don’t even need internet.

They can build these anywhere in the country, anywhere in the world. Far from opposition, far from resistance². They only need a few defectors, a few counties or countries that fold, and they have piles and piles of money for bribes.

As with any maligned industry, local opposition will simply shift where data centers are built — or, at best, make them a bit more expensive. They will move, like they always do, to remote areas, to marginalized communities. Concentrating suffering and externalizing harm.

That is simply the nature of NIMBYism. The harms are not stopped, merely pushed out of sight.

Footnotes

1. The strongest piece I have read on this argument is Stopping AI is easier than Regulating It, which focuses on this problem from an AI safety lens. ↩

2. No, you will not manage to storm all of the data centers across the country or the world by force. If you could, you would be overthrowing capitalism and ushering in an entirely different sort of revolution. ↩

The limits of boycotts

Boycotts (and consumer pressure more broadly) do work, sometimes. Genuinely harmful, pointless technologies like NFTs have been strangled in their cribs by a consumer base that simply rolled their eyes and dragged their feet and moved on.

And yet they do not always work, and the differences are informative. The modern world is full of terrible things which, on the net, probably make society worse. Cars and plastics, algorithms and advertisements, printer ink and pesticides, online gambling and mobile microtransactions¹. Widely used, widely reviled — technologies that, some days, I wish we could uninvent.

And yet they persist. Why?

As always, incentives determine equilibria. Their benefits are concentrated, and their harms are diffuse. They spark griping, but griping is not enough. Structural problems require structural solutions.

And for many of these cases: their cost-benefit balance is mixed enough that you cannot muster the political consensus required to ban them. Cars and plastics and pesticides, terrible as they are, are genuinely useful. If AI is the same, we should expect the same sort of outcomes: widely used, widely reviled, harmful in devastating, diffuse ways.

But AI is even more insidious. Humans are social creatures. Social pressure and social consequences² can be used to enforce norms. We can signal our values, make lists of those who betray us and cut ties with those who sell out. Demand disclosure of AI use and punish those who defect from our cause.

And that, in social circles which are sympathetic to your argument, will have a very strong effect. On disclosure³.

But disclosure is not use. Use, the thing which actually does harm, is effectively undetectable. Sure, you can spot the obvious, lazy uses. But you cannot catch the author who turns to the machines for an editing pass. The mother who asks for help with meal planning. The artist who uses generated reference images.

Unlike plastics and pesticides and cars, AI can be used without a trace. If the benefits of use are real, but disclosure is punished, you’re just incentivizing dishonesty (often, of omission). Over time the coalition will rot from the inside, through a billion tiny betrayals, each justified, in the moment, as different, warranted, or not really so bad.

Even as the coalition crumbles, some will resist. You see that today: in vegans⁴ and Mennonites⁵; people who buy all their food from the organic market or build earthships in the desert. Your footprint, your complicity, can be meaningfully reduced. You can opt-out and you can be genuinely happy⁶.

Those personal choices matter to the people who choose them, even if globally, they are not the first note in an inevitable crescendo. For all of the beauty and hope⁷ of these movements, for all the better ways they illuminate and minds they change and concessions they win, I fear that opting-out is ultimately self-indulgent. Again and again: indulgences are sold⁸ to absolve the privileged of the weight of sin, while the true evils go unaddressed.

If you want me to believe that rejecting AI will be different, you need to clearly show me why it will succeed when so many movements before it have failed.

Footnotes

1. Ah, but I repeat myself. ↩

2. Not, of course, to be confused with harassment. ↩

3. As I write this essay, I doubt myself, again and again. Wondering if I should publish it; if the personal costs are worth it when it would be so easy to just stay quiet. ↩

4. Forgive me: I am not a vegan in my personal life, despite knowing, on some level, that they are right. The temptations are too much to bear. I try my best to atone for it in other ways. ↩

5. I grew up in Waterloo, near a large Mennonite community. They were, from the ones I met, ethical and community-oriented people; genuinely happy and trying to carve out a better way of living in the face of a society that accepted a Faustian bargain. ↩

6. While beta-testing this post, one of my stringently anti-AI test readers told me that they were primarily concerned with societal degradation of skills and knowledge, which is why they are such a strong proponent of boycotting. That was a new and fascinating perspective for me, and strikes me as a parallel to existing craft preservationists or tradition bearers. Deliberately abstaining from AI is a pretty good solution for that particular problem! ↩

7. I have these daydreams too, you know. A quiet life; a chosen family; tending an endless food forest, free of the strife and stress of modernity. Perhaps one day I will sail across the Great Sea. ↩

8. Would you like a carbon offset with that flight ma’am? Don’t worry, we promise that it works. ↩

Regulation will save us!

Thankfully, humanity has invented a powerful social technology for exactly these sorts of problems. It’s called “regulation”.

While personal or local action is both accessible and mollifying, it is ultimately ineffective in the face of a technology that works¹. Defection, on a personal and local scale, means that opposition stumbles, and the equilibrium is reached all the same.

We must strike at the root of the problem, and demand that these companies do better. Remind them, for all their brilliance, that the state holds the monopoly on violence, and the state, in functioning democracies², reflects the will of the people. Us. Those who, overwhelmingly, hate AI.

We can pass regulation that chips away at the harms: solving the power problem by requiring renewable energy production and effectively eliminating water demand by mandating closed loop cooling. Those harms are chosen by the companies who build and run these data centers. We can simply not permit them.

We can ban the publication of deepfake pornography, impose restrictions on the tools to protect the mental health of users and make it harder to generate propaganda.

Hell, with the power of the state, we could even declare (and enforce) a data center ban. Reimagine copyright and force these companies to pay artists for their training data³. Negotiate a global pause on development to prevent the creation of runaway misaligned superintelligences. Even, if we dared to dream, ban the development and use of AI outright, like nuclear weapons. Technology is not, in general, inevitable.

The state is bigger than any of us, and structural problems require structural solutions. Regulations are the true prize we should be fighting for.

Footnotes

1. Please, I am begging you: try it, and trust the evidence of your eyes and ears. Everything hinges on this fact, and nothing I can say will convince you. ↩

2. Let me know if you find one of those please, I’m sure it would be very educational! ↩

3. Adobe Firefly has taken this path. Not coincidentally, they are pushing for legislation that would make this the only legal approach in America. ↩

Regulation will not save us?

All of that is, in theory, within the power of the state. Yet, when we look globally, even in the most hostile, anti-AI jurisdictions, we see legislation of the first kind, incremental and ineffective, not of the second kind, sweeping and bold. Can’t they see the harms?

To the credit of legislators, laws have already been passed! California is legislating transparency. Minnesota is imposing severe environmental restrictions on data centers. Even Tennessee has banned deepfakes.

And yet, no one is trying to regulate AI out of existence, stopping its harms once and for all. Even the EU AI Act, the boldest and most sweeping of proposals, only regulates high-risk applications, mitigating real harm without blunting job replacement.

These victories ring hollow, because they stop before they actually make a difference. That is not a coincidence, not a signal that we simply need to take the next step.

Unfortunately, defection is not merely a personal or local dynamic. Decision makers across the world are starting to internalize that AI works. Taking sweeping unilateral action risks being left behind¹, consigned to irrelevance by rival states that can out-compete you economically, field superior weaponry, and sabotage your critical digital infrastructure without ever firing a shot. You cannot ban AI, not in any way that matters. Hell, you might even need to race to keep up, lest you be denied access to critical tools by your frenemies.

It does not, on any level, from the single individual to the international bloc, matter how anyone feels about the net benefit of AI. The only way out of this trap is to cooperate: to come together and act at a level that denies defection.

Perhaps, in a different timeline, that would work. Build an international consensus, create binding treaties, enforce a ban for the good of humanity and regulate the blatant externalities that threaten to destroy society as we know it.

Climate change shows us that this will not happen. China is ruled by prideful technocrats, insulated from public backlash. The US is ruled by lunatics, and democracy crumbles.

If Europe, if Canada, if Japan and all of the other middle powers and developing countries ban AI together, all they will accomplish is unilateral disarmament.

But defection echoes further, and the picture grows bleaker. Even within a restrictive regime, bans are leaky. You can block the mega data centers, stymie corporate adoption of AI, push AI as a whole to the shadows.

But local models are already at the point of dramatic societal disruption. The technology and prerequisites are diffuse and dual-use: they are everywhere, and a full ban is ultimately an eco-primitivist stance².

AI, at the level which threatens to overturn society, is not like nuclear weapons, which require specialized expertise and large, trackable machinery. AI is like alcohol and cryptography and pornography. Arguably harmful, easy to create once you know how, trivially smuggled. And, crucially, highly desirable.

Prohibition has failed again and again. It would fail again here, for entirely unsurprising reasons. All it would yield is systems of oppression and control, while those in charge use AI in secret to enforce their grasp on power.

No one can stop most of the worst harms of AI: propaganda and cognitive offloading and labor displacement and surveillance. Because structurally, there is no path to do so.

Technology is not, in the abstract, inevitable. But that does not matter: stopping AI requires a constructive proof, a real path forward for the very real problem as it stands today. And none have ever been proffered.

Footnotes

1. Ah, the dreaded marketing phrase! Unfortunately, its accuracy turns entirely on whether or not AI “works”. That seems like it might be important to determine. ↩

2. Much like a communist revolution, reasonable people can disagree about whether or not this would be good. It doesn’t matter though, because we’re talking about paths we can use to get out of this mess. ↩

A future worth fighting for

Enough, enough! When will it be enough? That is the question I cannot escape.

In case it is not clear, no: I am not flinching from the ethical questions posed by AI. Instead, I am trying, very hard, to look them clearly in the eye. To read the hand we have been dealt, to think more than a single move ahead, to chart the least bad path that we, as progressives, have out of this mess.

I am not opposed to sweeping AI boycotts because I think AI is good. I am opposed to them because I think AI is bad. You only get so many moves, and you should be careful to spend them wisely. Do you remember the bans on plastic straws?

The harms are real, but they are not homogenous. As with climate change, surviving AI requires a mixed strategy, even if a pure one, in the abstract, would be better. Even targeted boycotts and NIMBYism, for all that I have derided them, have a role to play.

Do not, for the love of god, give your money to MechaHitler. Hell, think carefully about giving it to any American AI company. Learn from the campaign against apartheid, and coordinate on divestment to punish companies that betray the social contract. Fight, in every way you see fit, the data centers that spew poison into the air. Those are real harms. Those are places where you, as an individual, can make a difference.

But please learn the lessons of environmentalism: individual action is not enough. It is good to cut out meat, to stop flying, to install solar panels on your home, to bring reusable bags to the store. But you cannot let it become a pressure-relief valve. There is more work that must be done, and we need your help.

Learn the lessons of the left: every second spent infighting¹ distracts from the battle against our true adversaries. Every call to glorious revolution frightens those you need to convince most, pushing them into comforting complacency. And yet we should not, in equal measure, peddle that complacency: telling people that we need only vote, and our elected representatives will handle things.

These are, as I keep saying, structural problems which require structural solutions. Urgent, serious action is required, even though I don’t think revolution will succeed. So what do we do?

The problem is not, ultimately, the technology. Working less, with the collected art and science and knowledge of humanity at your fingertips: that is the promise of the future I have always dreamed of. How are we to reach fully automated luxury gay space communism without automation?

The problem is the social structure that surrounds it. The technology may be inevitable, but its consequences are not.

We must deny the concentration of benefits, and mitigate the diffusion of harms, in three important ways.

1. Regulation must be passed to stop the pointless avoidable harms and ease the transition.
2. We must decentralize control over this technology.
3. We must redistribute the dividend of prosperity, lifting all of us up.

Shattering false hope without replacement is a false kindness. I hope you will accept the bitter hope I have to offer instead. Please, honored hater, join me. We are only just getting started.

I want to teach you how to travel with the current, instead of fighting against it.

Footnotes

1. Yes yes, I see the irony of complaining about infighting in this post targeted at allies I hope to win. I hope that, in the eyes of God⁵², there is a difference between bitter sniping and attempting to build bridges. ↩

Regulation that works

Regulation is one of our most important tools, but we must be mindful of its limitations. We need to focus on legislation that will a) genuinely address harms and b) have a chance of actually passing.

We’ve already discussed legislation that might address harms but will never pass. Proposals like “ban all AI” live in this quadrant. For an example of legislation that might pass but will not meaningfully address harms, look to Adobe’s CREATOR Act. It trades one monopolist for another, offering token payment to the artists while selling the tools that obsolete their livelihood.

We must be careful, when legislating, to avoid splash damage. The CREATOR Act will kill remix culture and stifle creative expression, strengthening copyright in a way that only ever helps those powerful enough to enforce it. Full-on bans on AI use and creation would be much worse: stripping privacy and control over our own devices in our own homes, with limited effect on genuinely malicious actors.

Legislation that targets specific harms of AI (deepfakes, propaganda, AI psychosis, and so on) is moderately more effective. Its impact will be slow and partial, as with any new technology, but we must take what we can get gladly.

We can mitigate almost all of the direct harms of data centers with straightforward environmental regulation. Force them to pay for grid upgrades, to use strictly renewable energy, to limit noise and heat and use only closed-loop cooling systems. If needed, zone them away from houses, like any light industrial land use. Do not worry about burdening these companies: they can afford it.

Do not be content to nibble at scraps. Why strip the data centers of their copper when you can privateer their profits? Why stall until they relocate when you can instead extract concessions? Why should we think narrowly about regulating power, but accept that regulating power lies forever out of reach? Their wealth, their windfall, their surplus, belongs to us, because it comes from us. Not as consumers, but as citizens. We can use it to make our systems more resilient, even after they’re gone.

When regulating AI, there is one essential heuristic:

1. Controlling the visible, public inputs (land, water, labor, power) and outputs (lies, wealth, power, in the other sense) of AI companies will work well, within your jurisdiction.
2. Controlling the public use of AI will work less well, as generative output quickly becomes indistinguishable from genuine human creations.
3. Controlling the private use of AI doesn’t work at all: it is simultaneously ineffective and intrusive.

Cooperation is essential to make this work, but cooperation is only possible locally, in both space and time. You must consider defection at every turn, even as you choose, again and again, to reject it.

Decentralization that matters

One of the most appealing aspects of boycotts, is that they do not require us to sit on our hands, waiting for our captured, ineffectual governments to save us. I share that cynicism: voting harder and writing strongly-worded letters is not enough. We need to find paths that route around those in power, be they democratic, corporate or authoritarian.

In case you have not noticed, there are genuinely bad people in the world. Those people, as best as I can tell, are the ones poised to control the most powerful AI models.

Unlike the safetyists, I am not particularly worried about alignment. It is not the rogue machine god that worries me, it is the institutional actor. Alignment to who is the critical question. “Only I can be trusted with this great and terrible power” is an argument you should be deeply skeptical of.

If the capabilities gap grows, and persists, AI becomes a powerful tool of oppression and conquest.

We must fight it: through the creation of sovereign AI, and open weight models, particularly ones that can be run locally on consumer hardware. Power and agency must be distributed across the countries of the world, and across the citizens of those countries. Restrictions best serve oppressors: be careful what you ask for¹.

These efforts will not be cheap. They certainly will not be morally pure, and they are, assuredly, not enough. Nevertheless: they are essential, and more importantly, they are robust to defection.

Those in power cannot stop us from copying their vaunted models², twisting their “training data” logic against them. They cannot stop the distribution of weights, even across borders. They cannot stop you from running a model, weak but uncensored, on a phone you bought with cash. We can defect from their game — for good this time.

We must seize the means of generation³, stealing fire from those who would keep them from us for our own good and their own profit. The costs have already been diffused; the benefits must follow.

We need to find ways to use these tools for good, and approaches that reduce the harm that using them does. Again and again: harm reduction is the least bad, most effective strategy. I suspect AI will be no different.

I don’t know what “harm reduction for AI” means in full, even if I do know that not all harms can be reduced. I don’t know, precisely, how to preserve our craft and train our replacements in a new age with new tools. I want to try anyways.

I hope we can find out together.

Footnotes

1. A lesson most recently learned by Anthropic, who argued for control and restriction of powerful AI models, and then promptly had their flagship product (Fable) restricted by the US government. ↩

2. In a funny quirk of language, this process is known as model distillation. And just like a century ago: those who would impose restrictions on the public for their own good will be powerless to stop distillation from undermining them. ↩

3. Marx, contrary to popular internet belief, felt that the Luddites were primitive and misguided. He argued for seizing the means of production, not smashing it. Unfortunately, the Communist Manifesto, much like this article, is so long that most people never finish it. ↩

A future I believe in

Utopianism may be a non-constructive proof, but it is not a dead end. Polaris can still guide us, even if we cannot see a path to walk the stars.

While I don’t believe in the AI-vegan vision, that we can simply turn back the clock and uninvent AI, I still cling to hope that a world with AI can, in the long run, be genuinely good.

While every breakthrough from agriculture to the automation to the internet has brought its share of terrible problems, they have each raised the standard of living (on average) in a way that I would never want to reverse. Prosperity is, in isolation, great! The challenge is managing its side effects: rampant inequality, reckless pollution, capacity for oppression.

My solarpunk dreams center around ecological integration, blending careful regulation with radical personal freedom, and a carefree, meaningful existence for every single person on the planet. Technological advancements are a welcome, necessary part of that vision. But when it comes to managing “AI that works”, what we need is a redistribution of the productivity dividend.

By that, I mean redistribution in a very real way; in a way that makes the wealthy¹ flinch and gasp and clutch their wallets and vacation homes close to their chest. Not the lip service of Sam Altman and his ilk; bread tossed down to peasants to hold the riots at bay. Genuine, nearly-flat wealth curves, capturing the gains of another industrial revolution, and distributing them, for once, across humanity writ large.

The rich can hide their books, but they cannot hide their wealth: their property, their companies, their yachts and their land. What good is money if you cannot spend it? They need both suppliers and customers: you can sanction bad actors; you can deny them your market. Visible inputs, visible outputs: all within reach of the state. Capital flight is a tale old as time, but let the bastards try and defect; I suspect that they will find the rules of the game have suddenly changed for them too.

But in the end, I want something more than to tear down the tyrants. I believe in a society built on hope, anchored in equality, and powered by abundance.

That, finally, would be enough.

Footnotes

1. This, in any sensible redistribution scheme, includes me. That’s okay. Having been brutally poor, let me tell you: the diminishing returns start to hit pretty damn quick. ↩