LLM Inevitabilism

The early internet connected personal computing together. It built on technology that was democratizing.

LLMs appear to be democratizing, but it is not. The enshittification is proceeding much more rapidly. No one wants to be left behind on the land grab. Many of us remember the rise of the world wide web, and perhaps even personal computing that made the internet mainstream.

I am excited to hear the effort of the Swiss models being trained, though it is a step behind. I remember people talking about how fine tuning will accelerate advances out in the open, and that large companies such as Google can’t keep up with that. Perhaps.

I’ve been diving into history. The Industrial Revolution was a time of rapid progress when engines accelerated the development of cheaper access to fuels, more powerful engines. We were able to afford abundance for a middle class, but we also had enshittification then too.

While there is a _propensity_ for enshittification, I for one don’t see it as inevitable, and neither do I think an AI future is inevitable.

I do. The web was the largest and most widespread enshittification process to date, and it started with the first sale made online, with the first ad shown on a web page - this quickly went into full-blown land grab in the late 90s, and then dotcom and smartphones and social media and SaaS and IoT and here we are today.

The "propensity for enshittification" is just called business, or entrepreneurship. It is orthogonal to AI.

I think comparing rise of LLMs to the web taking off is quite accurate, both with the good and bad sides.

The process of creating the AIs require mobilizing vast amount of energy, capital, and time. It is a product of capital with the expectation of locking down future markets. It is not orthogonal to enshittification.

Small web was still a thing through the 90s and early ‘00s. Web servers were not so concentrated as they are with hardware capable of running AI, let alone training them.

I am optimistic that local models will catch up and hit the same pareto-optimal point. At some point your OS will ship with a local model, your system will have access to some Intelligence APIs, and that's that. Linux and BSDs will probably ship with an open-weights model. I may be wrong, but this is my hope.

If you're interested in a taste of that future try the Gemma3 class of models. While I haven't tried agentic coding with them yet, I find them more than good enough for day-to-day use.

Exception that proves some markets are still inefficient enough to allow people of good conscience to thrive. Doesn't change the overall trajectory.

> The process of creating the AIs require mobilizing vast amount of energy, capital, and time. It is a product of capital with the expectation of locking down future markets.

So are computers themselves. However free and open the web once was, or could've been, hardware was always capital-heavy, and it only got heavier with time. Cheap, ubiquitous computers and TSMC are two sides of the same coin.

> It is not orthogonal to enshittification.

That's, again, because business begets enshittification; it's one of those failure modes that are hard to avoid.

> Small web was still a thing through the 90s and early ‘00s. Web servers were not so concentrated as they are with hardware capable of running AI, let alone training them.

You can "run AI" on your own computer if you like. I hear Apple Silicon is good for LLMs this time of year. A consumer-grade GPU is more than enough to satisfy your amateur and professional image generation needs too; grab ComfyUI from GitHub and a Stable Diffusion checkpoint from HuggingFace, and you're in business; hell, you're actually close to bleeding edge and have a shot at contributing to SOTA if you're so inclined.

Of course, your local quantized Llama is not going to be as good as ChatGPT o3 - but that's just economies at scale at play. Much like with the web - most of it is concentrated, but some still find reasons to run servers themselves.

However, fine-tuning may be run locally -- what are you thinking about in terms of training?

"At some point your OS will ship with a local model, your system will have access to some Intelligence APIs, and that's that."

There's a secondary effect that I had not even discussed in detail here. I don't know how to explain it concisely because it requires reframing a lot of things just to be able to see it, let alone to understand it as a problem.

Let me see how concise I can be:

1. There are non-financial capital such as social capital, knowledge capital, political capital, natural capital, etc.

2. The propensity is to convert non-financial capital into financial capital at the expense of the other forms of capital. I think this is the core dynamic driving enshittification (beyond how Cory Doctrow described it when he coined it).

3. While LLMs and AIs can be designed to enhance the human experience, right now, the propensity is to deploy them in a way that does not develop social and knowledge capital for the next generation.

Ok, I can see that is true.

"Exception that proves some markets are still inefficient enough to allow people of good conscience to thrive. Doesn't change the overall trajectory."

That depends on what you are measuring to determine market efficiency. Social, political, knowledge, and natural capital are excluded from consideration, so of course we optimize towards financial efficiency at the expense of everything else.

Which comes back to: business does not have beget enshittification, and it isn't because of market inefficiencies.

I think we're going to have to agree to disagree on some of these points.

1. Expanding in more detail, my feeling (absent correct truth obviously) is that we'll find a Pareto Optimal point of intelligence. At that point my feeling, again unproven, is that with available hardware fine tuning open weight models would not be too difficult.

Training from scratch? I'm really not sure that's something an individual can do. But at some pareto optimal point I do think small organizations can train models.

I'm okay with that threshold. This world I envision is like today's open software which runs on closed hardware or interacts with closed systems like networks that use proprietary hardware.

2. I'm not sure the tendency to convert everything into financial capital is as all subsuming as folks like Doctorow make it out to be. The sentiment certainly drives engagement among a certain demographic, usually a college educated progressive one, but while I see evidence of its truth in some spaces I don't in many others. PCs are a good example. PCs continue to offer more functionality to the end user. Despite endless avenue to lock down and monetize the idea of a PC, it doesn't happen. Firms like Apple and Lenovo remain fairly committed to offering strong consumer experiences.

I suspect this sort of financialization is more prevalent for services or goods that are difficult to meter. Search is a great example.

> That depends on what you are measuring to determine market efficiency. Social, political, knowledge, and natural capital are excluded from consideration, so of course we optimize towards financial efficiency at the expense of everything else.

I'm taking a loose but classical definition of it, so obviously in financial terms. But there isn't really much choice to be made here: the market, as a system borne of aggregate human behavior at scale, is optimizing along a specific dimension for structural reasons. "Social, political, knowledge and natural capital" aren't excluded at all - on the contrary, they're converted into dollars and become part of the optimization, competing with other things that are also converted into monetary units.

It's just that, it turns out, those other forms of capital you mention tend to not have that much value, so they get optimized away, especially under strong competitive pressure.