So, Where Does Next-Token Prediction Leave Us?

Solved/Cooked

AI maximalists in some corners of the internet hate it when people refer to LLMs as just “next-token predictors” or “stochastic parrots”. It is instinctively taken as a pejorative. They use the words “solved” or “cooked” to signal the end of industries or classes of work that take real human creativity, expertise or effort. “Animation is solved”, “Hollywood is cooked”, “coding is solved”, “postgrad students are cooked” and so forth. It is far from a neutral description of progress, there is a certain glee to it. They celebrate the obsolescence. There is a belligerence in discussions and it is starkly reminiscent of people taking political sides online. I cannot think of any other piece of technology that comes close to this level of tribalism. Hmm, maybe cryptocurrency? Arch Linux users? Not even close.

These are extreme examples.¹ I had to put this on top, because from where I stand, I feel like the same people cheering now are the same people being economically priced out of this.

Why?

Machines that can think have been an important trope in our collective and literary fiction for so long. It was always a question of when and not if and the graph of when and what is going parabolic. I feel there is something primal underneath it: the hubris of creation, playing God, intelligence squeezed out of sand.

So, what makes a fanatical proponent? Why do they seemingly have contempt towards human ingenuity and labour? Do they have an overly optimistic view of living off of universal basic income, spending all their time in leisure while the machines subsonically² hum away at work that could end up being considered beneath humanity? We cannot and obviously should not generalise these things. But lately I’ve been thinking if it is just a class issue?

This cohort of people likely have a cushion that softens the concussive blows they are doling out right now. They perhaps have the luxury of a somewhat functioning government and a social safety net that they are witness to in all walks of life. Over half the world does not. Science and technology, I feel, has always had a certain apathy towards the plight the people at the bottom rungs. And it is by design, I fear. To break in, or bear the fruits of, you at least had to have been in a position to get an education.

The cushion and the safety net is largely transitory. It cannot be sustained forever, unless we could do something like tax the corporations for cutting out labour or something crazy like that. We cannot even fairly tax garden variety billionaires right now.

In the long run, there is no winning team here. There is no Basilisk to appease and no side to be on, the real world implications are coming for you too.

The Meta-Contract

The founding tenet of AI, “saving us” from all sorts of difficult things: climate, disease, poverty, conflict is falling, fast. The frontier labs have shifted to signaling at a more mundane, perverse motive: that of simply cutting labour.

Even the most exploited classes of people, always had - at varying degrees - a bargaining chip. A chip that corporations spend billions trying to snatch away. The chip of labour, of being needed. CEOs and proponents of AI flippantly announce and proclaim this chip will not be legal tender anymore. Anthropic’s CEO Dario Amodei has quite the reputation in this department. OpenAI’s CEO Sam Altman says, college degrees are worthless now because AI can teach better, in a more personalized way. Downstream to these people, VCs say in the next 5 years this and this class of work will be “solved” - only with the finesse of coming from money.

What happens when you strip a human of their economic utility? The promise of capitalism has always been - you will have a spin at the roulette table. If you work hard enough, you can make it. We chugged along and ushered in this ultra capitalistic world where if you work at an Amazon warehouse, a bottle is your bathroom, a dollar for your dignity. What happens to billions of people whose only shot at upward economic mobility is a kid that got a degree and goes on to find a job?

AI democratises the very things that the CEOs want obsolete. Sure, it may teach you stuff and build you dashboards, but you will not be able to sell them. It concentrates, it is the literal concentration of (if the dream is realized in full) the means of production into the hands of a wealthy few. A capitalist’s dream. It is behind a $200 subscription or behind a beefy GPU and the goodwill of labs to release open weight models. AI raises the ceiling, of human output, at the same time, it also raises the barrier to entry.

I see a lot of knowledge workers online adopting the mantra “with AI, not by AI” and this largely reads as cope. The corporations spending $250k + token costs will want to instead spend $30k + token costs hiring people in SEA. Again, this labour arbitrage is transitory too, they ultimately want to spend token costs alone.

God forbid anyone find an ounce of joy or contentment in their craft - it is now squarely gone. You are now a node. Your only job is to maximize for throughput. You take an input, produce an output with your AI. You are to keep pushing a stream of work at rates you cannot reliably review or verify. Even if you wanted to do it, there will be people that won’t and for the corporation, you are now an under-performer. You can task another agent to do it though, at an additional charge.

Recently Fields medalist Tim Gowers wrote about his experience with ChatGPT 5.5 Pro ($30 per M input, $180 per M output, mind you!). He says:

So if your aim in doing mathematics is to achieve some kind of immortality, so to speak, then you should understand that that won’t necessarily be possible for much longer — not just for you, but for anybody.

With AI, not by AI? The labs and corporations will have agents babysit other agents bruteforcing solutions and discoveries. You cannot do the same because you do not have the unlimited compute or specialized models. It is the same pattern as the structural hurdles that prevent an average person becoming a billionaire. You will be gatekept. You will license your ability to work. The chip is now for rent. The bargaining chip is now manufactured by TSMC and sold by Nvidia.

Non-technical middle managers who have not written a line of code in their lives, now feel that the biggest obstacle between them and greatness has lifted. They do not have to deal with pesky programmers anymore. They do not need to ask a programmer to change colour, sizing or the style of a breadcrumb on a webpage anymore. No more protests about how it is bad UX or the code complexity is not justifiable enough for some useless flashy feature. The AI does not complain, the AI does not unionize and it does not protest. It will listen to you. It will say something you said in passing was truly impressive and that it has not seen many people think that way.

How Did We Get Here?

Every website, every book, everything ever written, produced, photographed, videotaped became an opt-out by default into the training corpus. Opt-out only if you run a website, add lines to your robots.txt and respectful scrapers will stop. There are still troves of scrapers that do not identify themselves and I would imagine there is a black market for scraped content. Anything other than a website, you are out of luck. Thousands of people paid peanuts to label, refine and optimise the datasets. So many people conscripted into paying higher for their utilities because of datacenters being built at breakneck speeds.

Grotesque amounts of money are being put into the AI machinery. Nobody does it without a promise of returns. We would be reversing climate change and saving the turtles if that was not the case. Instead we are told AI will solve these very hard problems, with the additional bonus of a return. In many ways, the AI rush is the poster child of capitalism. This would never have happened any other way.

World leaders were lied to, persuaded and coerced into thinking if we are not ahead in the AI race, it is literally doom and gloom. The labs made it a national security issue to cut oversight on datacenter buildouts. It is also put forth as an indispensable tool in defense. Reading “You are absolutely right” repeatedly is probably not a good thing for world peace. Once it is “national security”, again there is no opting out, you fund it by default.

Anthropic is the master of anthropomorphizing LLMs - the “soul” document, philosophers on payroll, meetings with clergy and what not, yet their models are highly sought after by the American Dept of War. The compartmentalisation that must be required by the scientists and engineers to reconcile with the fact that their work being used to bomb and kill people must be crazy.

What’s Left?

We spent years educating people on online hygiene and what cookies were and why they need to consent to websites storing small strings of text in their browsers. People now willingly and unwittingly give up their and other people’s private information and stories into unassuming text boxes.

OpenAI, even in their API offerings (at least to the average person) do not offer a zero-retention policy. They store your conversations for an unstated period of time. I’ve been seeing this infographic floating around about how we are running out of training data. Do you really think they will not use your conversations to edge out the other labs in this cutthroat race, to be ahead in LLM Arena’s tables?

We pay with money, with data, give up our bargaining chip, give up simple pleasure and joy of craft one may have stumbled upon early in life or accidentally taken up as an adaptive pattern. The loop is self enclosing and we are locked into the garden, but out of the harvest.

So, where does next token prediction leave us? In a perpetual loop of rent-seeking for something made with humanity’s collective output of centuries. It is not a good place for an individual to be in, regardless of class.

- 0x5FC3