Large Language Models and the Overproduction of Text

consolidation, not collapse

The idea of a "crisis of overproduction" has been around since Marx made it up. The idea is that capitalist economies are good at producing stuff, and that the technologies of production under capitalism are driven by competition to continuously improve, so eventually any industry will be producing more stuff than its customers can possibly consume.

These ideas are not original to Marx: John Stewart Mill and David Ricardo were both aware of the tendency for increased productivity to result in decreasing profits, which is one way "overproduction" manifests itself in the real world.

These "crises" do happen, but Marx's idea of what would happen next was completely wrong, to the extent that it's kind of embarrassing that anyone still thinks it's correct.

Because what Marx predicted--almost 200 years ago now--was that crises of overproduction would lead to the collapse of the capitalist order.

As good Bayesians, we should want to test this idea, and fortunately testing it proves to be not difficult.

Look around you. If you see capitalism, it means that Marx was utterly, absolutely, and completely wrong in his prediction of how the complex dynamical system that is the capitalist economy would react to making too much stuff.

This should not be a surprise: Marx was a catastrophist, given to the cognitive error of thinking that the only change possible was revolutionary, sudden, and--preferably--violent change. So the fantasy of the capitalist order collapsing under its internal contradictions was deeply appealing to him, and his prediction is best understood as motivated reasoning.

Since we still see capitalism, and yet Mill, Ricardo, and Marx were not wrong in their predictions about the tendency of profits to go to zero as productivity increased, we need to ask: If they don't collapse, how do capitalist economies react to crises of over-production?

High tech in general and software in particular gives us some nice examples over the past few decades.

For instance, commodity (desktop) software ran into the issue as feature sets matured in the '90s. In the early days of personal computers nobody knew what a good word processor or image editor or even OS should do, and people bought new releases of products like WordPerfect and CorelDraw and Lotus123 to get new features. After a decade of experimentation there was a pretty clear idea of what made a "good" application in a variety of areas, and software companies started dropping dead.

They had reached a crisis because once they stopped adding new features the cost of making the next copy of the mature software was essentially zero, and the equilibrium market price of a commodity whose marginal cost of production is zero, is zero.

This is as severe a "crisis of overproduction" as one could possibly have. Productivity became as close to infinite as possible. The effort required to copy a floppy disk was the entire "production" cost of a new instance of Word or PhotoShop.

This did not lead to the collapse of commercial software development. No one was arrested, tortured, and sent to the gulag as the Dictatorship of the Proletariat claimed its rightful place atop the heap of bodies after having managed "to wade through slaughter to a throne,/And shut the gates of mercy on mankind".

Instead, there was consolidation.

Dramatically lower profit margins in the "overproduced" goods created conditions where only enormous enterprises were still viable. By taking the lion's share of the market they could use quasi-monopoly powers to keep prices above zero.

Microsoft, for example, used their dominant position in the OS market to bully hardware manufacturers to include a copy of Windows on every machine they shipped. The alternative was for Microsoft to not support the hardware configuration of that particular machine, so the majority of users who bought Windows separately would experience poor, buggy, performance, if it ran at all. This was not an abstract threat: it's important to understand that Microsoft execs were willing to look hardware execs in the eye and say, "We can't promise a good experience for your customers on Windows if you don't agree to ship with Windows, because if we aren't ensured of the sale of a Windows licence with every one of your machines, it isn't worth us putting in the engineering hours to support it with drivers and testing."

Large companies were also able to create subscription models for what used to be commodity software. Adobe, AutoDesk, and the like were big enough that selling subscriptions could make them profitable. Smaller companies could not survive on the margins available.

All of these were done in aid of making software look like a scarce commodity, and the result was its production and support was still incentivized under the capitalist mode of production.

Incentives matter.

Socialism fails for many reasons, but one of them is that nobody knows how to incentivize socialist managers to make the same decisions that capitalist managers would make, and we want them to do that to a surprisingly large degree. But we want to capture the benefits of those decisions for the public, not just the management and executive class.

Lacking appropriate incentives, socialist managers tend to make decisions that will benefit their own position within the amoral hierarchy of the state, not ones that will benefit the public. Thus far no democratic system has proven itself able to reign in this tendency.

Capitalism provides monetary incentives to build nice things, but those incentives often also promote the creation of considerably less nice things, and the creation of nice things in entirely not nice ways.

Open source software is increasingly competitive with capitalist software, but problems of organization, governance, and incentives persist. I'll likely come back to this issue later.

For now, I want to focus on two more crises of over-production, one in recent history the other just bursting upon us.

The first is over-production of advertising space, which happened during the initial dot-com boom. Advertising space used to an extremely limited commodity. If you've seen "Mad Men" you may recall the ruse Don Draper's team used to limit the reach of Richard Nixon's 1960 presidential campaign against John F. Kennedy by buying up all the available airtime for ads in the southern U.S.

Ad space was a scarce enough commodity that newspapers up until the dot-com boom made a fair fraction of their revenue from classified ads. The overproduction of ad space online practically eliminated that revenue stream, setting the traditional news business on a slide to desperate mediocrity it is mired in today.

The Web is more-or-less unlimited, which meant there was a more-or-less unlimited volume of ad space suddenly available.

And again: contrary to what every good Marxist continues to believe, this did not lead to the collapse of the capitalist system, even in the limited scope of the advertising market.

As in desktop software, it led to consolidation.

In both these cases, a good or service that had been in short supply and therefore commanded a high price and supported multiple profitable enterprises with their own employees and shareholders suddenly becomes practically unlimited, causing its price to fall precipitously, such that only a small number of very large, very efficient enterprises can survive on the profit margin available. Part of the "efficiency" of such enterprises often involves quasi-monopoly powers that allow it to make suppliers or partners "an offer they can't refuse".

This process is as old as capitalism. If pottery is made manually by expert craftspeople, many thousands of experts can make a living at it while everyone pays for expensive crockery. If pottery is made by machines tended by people who take a week to train, a small number of companies can sustain themselves, and no expert potter can create enough volume to make a living at the margins set by the machine-made products. Sure, the machine-made stuff may be of "inferior quality", but the price is a factor of a hundred lower and the quality is down by ten percent. Only a tiny number of canny marketers can survive as individual crafters in such a market.

Scale matters in this transition away from individuals and small enterprises.

Because individual humans beings have a scale there is a lower limit to the economic profitability of any activity that individuals can undertake. Below that limit individuals can no longer make enough to keep body and soul together. This is why as profitability drops due to "overproduction", the natural capitalist response is consolidation.

Capitalist corporations, like socialist bureaucracies, are amoral hierarchies (this is why so many "socialist" nations end up with what socialist ideologues sometimes describe as "state capitalism": the underlying organizational system is the same.) Amoral hierarchies scale: they can grow without any known bound.

The recurring "capitalist crises of over-production" don't result in the collapse of capitalism, they result in more capitalism, with bells on. And giant quasi-monopolies dominating a devastated landscape where once a rich ecosystem of small producers thrived.

The crisis of over-production that is now looming menacingly on the horizon is the over-production of coherent, readable, text, brought to us by the metastasis of the Large Language Model (LLM) in recent months.

My fundamental prediction remains that the capabilities of these models will be limited by their lack of consciousness, because consciousness is an evolved capacity for regulating intelligent behaviour that nature could not find any alternative to. Given the well-known downsides of consciousness (hello, anxiety!) there is every reason to believe nature would have settled on such an alternative were such a thing possible. Nature didn't, so it probably isn't.

This does not mean that LLMs won't be hugely disruptive, but the nature of that disruption is best understood as a crisis of overproduction, which we know capitalism has no difficulty at all accommodating via consolidation.

This leads me to predict that if LLMs continue to improve to the point of being able to generate adequate simulacra of commercial fiction, the days of the individual writer are numbered. It won't be indie authors cranking out Chuck Tingle clones (if you don't already know who Chuck Tingle is, you may be happier not knowing).

It'll be mega-corps generating new thrillers on demand. Ever want to read a novel about a wandering ex-soldier named, say, Rack Jeacher who's <<fill in race/gender/orientation/kink here>>? That may very well be possible in a few years. Or sooner.

If LLM-generated commercial fiction becomes possible, the history and theory of overproduction tells us it won't benefit individual authors. It can't. Large corporations will cut the author out of the equation, hire low-paid prompters, and generate books 'in the style of' anyone you care to name.

Individuals won't be able to compete because an over-produced good is only profitable at scale, and only amoral hierarchies scale.

The overproduction of text--or anything else--also won't lead to the collapse of the capitalist order. For that we need an organizational technology that out-competes the amoral hierarchy.