If you liked this piece, you should subscribe to my premium newsletter. It’s $70 a year, or $7 a month, and in return you get a weekly newsletter that’s usually anywhere from 5,000 to 18,000 words, including vast, detailed analyses of NVIDIA, Anthropic and OpenAI’s
I think you’re correct that it’s what they’re hoping to do. The premise with AI is that companies could retain a handful of workers who will drive these tools to do the work of entire departments. In this sense, it would be like moving from artisanal mode of production to factory work. Instead of having each piece of code written by a human, you leverage automation with the human only having to know how to operate the tool that does the bulk of the work.
The problem with this model is precisely what @Philosoraptor@hexbear.net pointed out, which is that unlike machines in a factory, an LLM is not a deterministic tool. On top of that, the nature of work it automates is such that it requires active decision making, hence why it hasn’t been automated so far. And here’s where the whole scheme falls apart because the LLM is unable to actually determine whether something is correct in a business sense. All they do is produce stochastically probable outputs, which means human is the bottleneck in the whole process. You can get an LLM to generate code orders of magnitude faster than a human could, but somebody has to figure out that it’s actually doing what’s intended.
Some productivity gain is possible here. For example, you can focus on making specs, and tests, treating the code produced as a black box and testing the product functionally. But, all of these things still require a non trivial amount of work. So, the gains are marginal in the end, and that’s why we see every report coming back from companies that went all in on this tech that they’re not seeing the results they were hoping for.
While I think that this tech has value, and it can do a lot of useful things, it simply doesn’t do what people marketing it claim it does. So, the analogy with factories works in a sense that the product is produced through automation rather than being crafted by hand, but it doesn’t translate into the same type of productivity gains due to the human still being the main bottleneck in the whole process.
That’s definitely an important issue regarding making things with LLMs. Deterministic tools like compilers and LSPs didn’t cause as much of disruption to the industry to my knowledge, even if they did have their naysayers. The human bottleneck issue isn’t what I was trying to get at, though.
I’m trying to look at it from production of software all the way to its consumption. My suspicion is that the differences between software and physical products – both in how they are produced and how they are consumed – might have an effect on how surplus value is extracted and how that effects the organization of society at large that is, in some ways, qualitatively different than what happened during the industrial revolution.
If my hypothesis sounds vague, that’s because it is. I’ll definitely need to finally read Capital in full at least.
I’m not sure that software fundamentally changes the dynamic between fixed capital and variable capital. Workers do often own their own computers and using open source tools. In that sense, there is a difference from a factory where the capitalist owns the press. But I think we need to ask what really counts as the means of production in software.
The real bottleneck is distribution, hosting, and access to users. A developer can write code on their own machine using free tools but they cannot deploy it at scale without cloud services like AWS. They cannot reach users without app stores or corporate owned repositories. They cannot monetize their work without payment processors owned by capital. The open source code itself often depends on infrastructure maintained by capital backed foundations or corporate employees.
So the means of production in software are the servers, the networks, the data centers, the authentication systems, the advertising platforms, the app stores. All of those are owned by capital. The worker’s personal computer becomes like a craftsman’s tool in the putting out system. It looks independent but it is actually subordinated to capital’s control over the marketplace and the essential infrastructure.
That means the underlying relation has not changed. Capital still exploits labor for surplus value by controlling access to the means of producing and distributing value. Software just makes that control more abstract and more centralized. Software just makes the means of production cheaper to reproduce and more abstract. That actually strengthens capital’s hand because it can amass enormous fixed capital in the form of code and data with near zero marginal cost. And the bottleneck is still labor since someone has to build and maintain the software. I’d argue that the extraction of surplus value just becomes more opaque, and the user becomes a source of data and training material. That is just an evolution of industrial capitalism rather than a qualitative break from it. The real difference might be that software allows capital to bypass traditional labor relations altogether by turning users into unpaid workers. That is a new twist but the underlying logic of accumulation remains intact. Read Capital for sure but also consider how the platform economy updates Marx’s categories rather than overturns them.
I didn’t mean to do otherwise. What you say is true though. This definitely isn’t a qualitative break with industrial capitalism in the same way the financialization that happened/is happening for more than a century isn’t a qualitative break with it.
yup