LLM 不是不可避免的技术

About as common as the statement that AI technology is just a tool is the statement that AI is "inevitable", so we should all just give up and get with the program because the technology is definitely going to persist and we're definitely going to give up our thinking to the machines. This is obviously a thought-terminating cliche, but it's work digging into a little further: after all, given how much it's being said, it would be worth knowing exactly how and why it's wrong.

The idea of a technology being "inevitable" is a curious one: after all, nobody knows ahead of time what technologies are going to be developed or adopted by society, so you certainly can't say, for example, that workable nuclear fusion generation is inevitable: that would be foolish. Equally, in hindsight any technology that is widely adopted and popular could be said to be inevitable: you get no points for saying that the development of the iPhone was "inevitable", not the least because we don't actually know if that's true for the shape the technology took. It's probably true for something with the vague shape of the technology in question, but it could have taken any number of forms, including a bunch that we'd probably think are quite strange.

So what is meant by inevitable when AI boosters say it? And how does AI measure up by their implicit standards?

What does it mean for a technology to be "inevitable"?

The glib response here would be to say that no technology is ever inevitable and that everything is contingent. This has the advantages of being technically true and sounding wise, but unfortunately technologies very much exist that a reasonable person might well label that way. Firearms are a good example of the latter case: once they've been developed, you quite quickly hit the situation where if you want to maintain your independence and safety in a society, you have to either use one or have one used on your behalf: no other options really work.

The inevitability of a technology is not, however, a factor of the technology alone. After all, mesoamerican civilisations very much understood the wheel as a technology. However, in dense jungle and without a road network already existing, wheels aren't actually hugely useful: it's thus no surprise that mesoamerican civilisations primarily relied on water transport and pack animals and relegated wheels largely to toys for children. More saliently, the value of cannon as a military technology was significantly greater in Europe, where political fragmentation led to fortifications being built with tall, thin walls designed to resist escalade, than they were in China, where stone-faced walls built largely of rammed earth were the norm (a cannon can blast a breach in the first much faster than it can in the second, though it will eventually breach both).

The inevitability of a technology is thus, in itself, a thing that is contingent on social and cultural structures: the same technology might prove to be "inevitable" in one cultural sphere, completely useless in another and useful but far from being "inevitable" in a third. Our criteria for "inevitability" relative to a given culture might thus look roughly as follows.

A technology is inevitable if:

• The technology can perform a task to a standard equal or not markedly inferior to that of an alternative technology doing the same task.
• The task in question is socially valuable: we need a lot of it done, and it's currently done by labour that is, in some sense, scarce
• The technology can, in volume of adequate output or in quality of output, overwhelmingly outclass what alternative technologies can do. A moderate improvement isn't really enough to force inevitability: it has to be a massive difference.
• There's no way in which the people not using the technology can change their behaviour to neutralise the impact of the technology. Firearms were overwhelming in this sense while the counterweight trebuchet wasn't because, while it's essentially impossible to mitigate the impact of cannon without radical changes to how fortifications are built, a counterweight trebuchet can have its impact blunted with incremental changes.

Unless all of these conditions hold, the adoption of a given technology can be fairly held to be contingent and not inevitable: it's not a choice that a social grouping was forced into, but one that they chose relatively freely. In short, inevitable technologies are ones where, if you don't adopt them, you will swiftly be overwhelmed by the cultures and societies that do. This seems to me to be largely in keeping with the way that LLM advocates use the term: LLM technology is inevitable because people who don't use it will fall so far behind those that do in power that they will end up under the control of LLM users.

Looking at this, we can quite clearly see the pattern of inevitability when it comes to gunpowder. A cannon can breach a curtain wall much faster than a counterweight trebuchet, and more efficiently too. Breaching curtain walls is something that, in a fragmented mediaeval Europe, is very socially valuable: if you can breach a wall, you can take a city or a castle, and the faster you do that, the less time you have to spend laying siege. Cannon is overwhelmingly more powerful than other kinds of siege equipment, and even early small arms were able to deliver overwhelmingly more energy in a shot than an equivalent crossbow or longbow could. Finally, there wasn't any particularly easy way to neutralise firearms: they're very difficult to effectively armour against and while you can design fortifications that are effective against cannon, the labour inputs required for that demand a massive restructuring of society. Practically speaking, then, the moment cannon arrived in Europe, militaries found themselves in a situation where they had to adopt firearms technology as quickly as possible or else be absorbed into the state of someone who had. That, I think, is what we colloquially mean by a technology being "inevitable".

We might compare this, for purposes of discussion, against the magnetic compass. The magnetic compass meets the first two criteria: it works a hell of a lot better for finding north than taking sightings of the sun and calculating does, and it works even absent known landmarks. Moreover, knowing the direction of north is rather useful when you're a maritime power and want to be able to navigate effectively. However, it's not overwhelmingly efficient: you can still navigate effectively without one, and it's easy enough to neutralise simply by virtue of not taking an interest in maritime activities and becoming a land power. Thus, while the magnetic compass allowed the colonial powers of Europe to massively extend their reach and exploit resources that would otherwise have been unavailable, Russia, for example, would likely have become and remained a major power on the continent whether they had the magnetic compass or not. The compass, then, while it was a very powerful and useful technology, was not inevitable in the way that firearms were.

Are modern LLMs inevitable?

So how does modern AI (in the LLM sense) measure up against these standards?

The first point is the question of whether an LLM can do work not markedly inferior to that of a human doing the same task or not. This one is debatable. An LLM absolutely cannot produce work to the highest human standards, or even a good human standard at the moment, and it's unlikely that it ever will. Most human work, however, is not done at a good standard. At best it's mediocre, and at worst it's downright horrible. The bulk of human knowledge work done in an awful of settings, then, is unfortunately more or less interchangeable with that of an LLM (seriously, so much human work is just awful: I wish I could say otherwise). On this point, then, LLMs probably pass the criterion.

The next question is that of social value. A lot of the knowledge work that LLMs purport to replace is scarce and valued: the writing of software, research work, graphic design and so forth. Generally, however, the LLM can only replace the least valuable parts of that work, to a greater or lesser degree (an LLM can generate a meaningful volume of tolerably mediocre code, but only the very worst kinds of graphic design and fairly superficial research documents). This means that, when any level of real quality is required, human expertise will have to be involved, whether an LLM is being used or not. On this question, then, the results are mixed: LLMs can do socially meaningful work, but only to an extent, and the more socially important the work is, the less useful LLMs are for it.

Now we come to the question of whether or not the difference that LLMs bring to the table is sufficiently overwhelming to make their adoption inevitable, and on this I think the answer is a clear no. As established above, LLMs just can't do top-flight work at all: at the heights of basically any discipline, we're still very much relying on unaugmented human work. Even at lower levels of accomplishment, LLMs still aren't overwhelmingly powerful: even when an LLM is being used, the amount of effort that goes into making sure that the LLM is directed correctly and that the output is adequate to the task means that the volume of output is sharply restricted. This means that while you might get some more output from using an LLM than you might without it in some cases, the actual speed-up isn't drastic. You might be getting between 20% and 50% more PowerBI dashboards out of an LLM-enhanced workflow than you would without LLMs, but the dashboards are likely not a core bottleneck in your business in any case, and as of now, at least, it seems that the closer you get to those core bottlenecks the fewer efficiency gains you see. Either way, I cannot help but feel that if there were a real overmatch in capability, we would very much have seen it so far. There would be at least one tech company producing five to ten times as many new tech products as anyone else and slowly forcing everyone else out of the market. We have not seen this: in fact, what we have seen is that the companies that have been going all in on AI have been struggling more than the others. The overwhelming levels of real impact that we would have had to see (think tanks rolling through the Ardennes Forest, crewed by methed-up Wehrmacht) to conclude that the technology must inevitably adopted... it simply isn't there.

Finally there's the question of whether or not there are things that other people can do to neutralise the impact of the technology relatively cheaply. There's not much evidence for this one way or another at the moment because social trends take a while to shake out, but in the forms of social competition that our societies tend to engage with, I'm inclined to say that there absolutely are. The level of social opprobrium directed at LLM use and at people attempting to pass off generated slop as work is immense, and that cannot but have, in the long run, negative impacts on the technology. In general, new technologies need some level of social acceptance in order to be adopted, and while many technologies have been adopted in the face of indifference or mild dislike, very few have been adopted in the face of the kind of white-hot hatred that's directed at LLMs at present. Moreover, we aren't talking about John Henry and the steel-driving machine here: in terms of actual useful output, humans can still match an LLM quite comfortably even if an LLM outmatches us in terms of sheer volume.

On these criteria of inevitability, then, LLMs don't come out very well. They only really pass on one criterion, and that incompletely. Moreover, on the criteria that push a technology from being merely useful to truly inevitable (criteria 3 and 4), the LLMs absolutely fail to get anywhere. Whatever the utility of LLM tools is, then, it seems clear that LLMs as such fall entirely in the category of contingence rather than inevitability.

This isn't an end to it, however, because we can clearly see that LLMs are having a massive impact over large swathes of our current social situation. Even if the technology, by the standpoints given above, isn't inevitable, it certainly feels, to a lot of people, as though it is. Why is that?

LLMs as chemical weapons

In his work Why don't we use chemical weapons any more?, Bret Devereaux discusses the question of why, since World War I, top-flight militaries have almost never used chemical weapons in battlefield applications. What's immediately striking is that this is basically true: despite the development of increasingly potent nerve agents such as Sarin, VX and Novichok, these have almost never seen battlefield use by the great powers. Neither Soviet Russia or Nazi Germany used them during World War Two, despite both of these powers being pushed to levels of desparation that would likely have resulted in their use having happened had they been at all effective. They weren't used in Vietnam or Korea (despite superficially similar defoliant agents being used), and while there have been some reports of Russia using chemical weapons in the Russo-Ukrainian war (mostly, as it seems, as a kind of pro-forma atrocity), reports of them being used in any serious way that enabled Russian battlefield gains are almost entirely absent. Given the sheer lethality of these agents, then, one might ask the question of why they've not become inevitable technologies that we cannot avoid using in war, however immoral their use might be.

Devereaux's answer to this is that for top-flight armies operating on modern manuever warfare principles, the great lethality of chemical weapons does not, in fact, translate to great utility. Modern manuever warfare, after all, relies on a great deal of tactical mobility on the attack, aiming to move fast enough to isolate and disorient resistance, preventing a co-ordinated defence and causing defending forces to collapse. In such a situation, the last thing you want is to use munitions that make large areas impassable to both you and your enemy, which chemical weapons inevitably do. Adding to this issue is the fact that more or less complete protection against chemical agents (gas mask plus NBC suit) is, by the standards of a modern economy, remarkably cheap: as Devereaux phrases it, there's a good chance that complete protection against this kind of weapon is cheaper than a soldier's rifle, and a wealthy country can, if it chooses, basically give its entire population protection against chemical attacks. And besides that, modern explosives are arguably more lethal in a battlefield situation, weight-for-weight, than almost any chemical agent (in principle, the lethal dose of a chemical agent is very low, but actually getting someone to be exposed to a sufficient dose of a chemical agent on a battlefield is a crapshoot at best: five litres of sarin in the Tokyo Subway attacks only actually killed twelve people, despite the situation being about as favourable for a chemical attack as you can get). This means that if you have an army capable of modern-system manuever warfare, whether you're fighting another modern-system army or a static-system army (to use Devereaux's terminology), chemical weapons are basically useless to you.

If you're a weak power fighting another weak power, though, chemical weapons can be quite useful. After all, if none of these points hold, using the technology might well make sense, as we saw during the Iran-Iraq war. Similarly, chemical weapons might be used in terror attacks or to send a message (Sergei Skripal's poisoning in London seems to have fallen into this category). Notably, however, you're only going to use this technology if you are, in one sense or another, a weaker power: you don't have the budget, the will or the skills to do modern manuever warfare. In general, chemical weapons are an excellent object lesson in why a supposedly very capable, effective technology is far from being inevitable and is in fact highly culturally contingent in terms of whether or not it's useful. And, to be clear, it's obviously wrong to say that the widespread adoption of chemical weapons is inevitable in our sense: one might perhaps say that it's inevitable that the technology will persist (it's impossible, after all, to un-invent a technology), but it's clearly not inevitable that advanced militaries will inevitably make and use the devices.

Now, to the point of this article: it's my contention that LLMs more or less fall into the same technological category as LLMs when it comes to their inevitability. Chemical weapons, after all, fall out roughly the same way as LLMs in our earlier categorisation: they can do some things in a military context, sometimes better than equivalent conventional weapons, but they're not overwhelmingly better than conventional weapons, they're often way worse and they're trivial to neutralise.

The inevitability of LLMs, therefore, is of a piece with the inevitability of chemical weapons. Weak players who are fighting other weak players might well find some use for them. I suspect that we're going to be seeing second or third rate software engineering shops using these tools (whether they're the hyperscaler models or local models) for a very long time. For them, the primary concern is volume of crap put out, and the functionality of what's produced is at best a secondary concern (Shopify and Mailchimp are barely-functional pieces of shit and were so before LLMs became widespread, but they've always been quite competitive at the lower levels of their respective applications). People and companies who are at all serious, however, are likely to use them only a little or not at all (the pattern I've noted among people who use the technology whom I'd consider serious is very different from other use , and also generally much lower in magnitude even for the heavy users), and if the pattern we see with chemical weapons holds, those companies will massively outclass what the bulk of the competition is able to do. Similarly, LLMs might have meaningful applications in hybrid warfare, as in the use of chemical weapons as a terror weapon (the recent rise of LLM-backed cyberattacks and LLM astroturfing bots genuinely concerns me), but remain uncompetitive when it comes to actually doing work of substance. I think we see both of these patterns come up frequently enough that I strongly suspect this is going to be the pattern going forwards, at least in the near future.

Certainly, given that situation you might say that LLMs are inevitable in the weak sense: inevitable in the sense that we can't uninvent the technology and that there are going to be weak actors that will persist in using the tools. In the strong sense, however? Certainly not. Any actor that's at all serious in their field is, sooner or later, going to hit a point where LLMs cease being useful to them and actively hinder their further progress, and we absolutely do not see the massively overwhelming advantages that LLMs would need to have in order to become inevitable.

The question of how our future looks going forward, I think, depends on our last point: how well can we neutralise the effects that LLMs are having on our situation? If we cannot find a way to effectively make it difficult for LLMs to hurt us, the situation is likely to be one where LLM adoption isn't inevitable, but it'll continue for largely contingent reasons. After all, most combats in our society (so to speak) are conflicts of weak parties against weak parties, and if the LLM cannot be effectively neutralised LLM technology can still act as an effective deterrent against stronger powers. People doing things for whom LLMs are useless will still be able to hold their own, but they won't have enough of an overwhelming advantage over LLM users that they can make LLM usage prohibitive. This is, unfortunately, quite a bleak world: it's a world where there is still space for real skill and care, but also one where a significant minority of the population is stuck pointing LLMs at each other and rotting their brains away in the process, the consequences of which we'll all have to deal with: a well-poisoning situation in which we all have to dedicate significant effort to filtering the LLM crud out that could be spent on something else.

If, however, we can neutralise the LLM threat somehow (perhaps by some combination of regulation, economics and social opprobrium), there is considerable hope. If we can mitigate the effects that LLMs have on us and reduce the filtering burden, we can, more likely than not, simply outcompete LLM users. After all, LLMs, as we've previously established, are mostly useful to people who are not good at what they do: weak players in our socioeconomic game. LLMs are, at present, mostly allowing them to partially keep their heads above water relative to relatively skilled practitioners. And I think there is hope that we can neutralise that effectively. A large part of why LLM users are able to hold their own against strong players, for the moment, is that there's considerable belief in the power and potency of LLMs: a belief that is convenient for a lot of people but also fundamentally irrational. The idea that LLMs can do everything a human can only better, or even that it can lead to massive productivity gains, is already eroding. We can speed that erosion, and the moment that the belief erodes enough, the people who don't use LLMs because they already outclass them suddenly develop a crushing competitive advantage. I don't know when this will happen: it might take a few years, or even as long as a decade. But eventually I think this belief is quite likely to collapse.

Looking at all of this, the idea that LLMs are in some sense "inevitable" is more than a little foolish. Certainly, the technology we will always have with us. Certainly, the astroturfing bots are not going to go away. But is this technology overwhelmingly powerful enough that we all need to be using it now lest we fall under the power of people who do?

Nope.

Arca is my project to help maintain the skills and knowledge that we need to outclass the LLM users. We're in very early days yet, but the platform is available in early access, so if you're interested, please sign up and help me realise the project in full and keep me fed and housed while I do so.