I’m a backend engineer. For the last couple of years a big part of my job has been wiring LLM “agents” into production systems: the kind that answer customer calls and chats, pull up your account, and decide whether a human ever sees your case. Most AI discourse comes either from boosters or from people who have never seen how this stuff actually gets deployed, so here are some notes from the shop floor.
1. The product is not intelligence, it’s deskilling. Braverman laid it out in Labor and Monopoly Capital: management studies the labour process, extracts the knowledge from the worker, and hands it back to them as a fragment. Separate conception from execution, concentrate conception in management. That is literally the project. Transcripts from experienced workers become training and evaluation data. The judgement of someone with ten years on the phones gets turned into prompts, flows and guardrails. Marx called this the real subsumption of labour under capital: capital no longer just buys your time, it rebuilds the work itself in its own image. What’s left for the human is the edge cases, at a faster pace and with less autonomy.
2. “Replace workers” is the pitch, “discipline workers” is the reality. The humans mostly don’t disappear. There are fewer of them, they’re cheaper, and they’re monitored by the same system: handle time, sentiment scores, compliance flags, all generated automatically. The agent is also the foreman. Marx describes exactly this in Capital vol. 1, ch. 15: in the factory it’s not the worker who uses the instrument of labour, it’s the instrument that uses the worker. Dead labour, in this case the frozen knowledge of thousands of past calls, dominating living labour. And the ones who are let go don’t vanish either. They join the reserve army, which is the stick held over everyone still employed.
3. It’s relative surplus value, not magic. Nothing here escapes the old logic. Marx quotes Mill doubting that all the machinery ever invented had lightened a single worker’s day, and his answer is that it was never meant to. Under capital, machinery is a means for producing surplus value: it cheapens labour power and lengthens the part of the day you work for free. The goal is to cut the labour time needed per interaction, and the gain goes to the client’s margins, not to shorter hours for the people doing the work. The same machine under different relations of production would mean the opposite.
4. The general intellect is being enclosed. In the Fragment on Machines, Marx talks about the “general intellect”, society’s accumulated knowledge, becoming a direct force of production. LLMs are that almost literally: trained on the collective writing of humanity, then rented back to us by a handful of firms. Socialised production, private appropriation, the basic contradiction in its purest form. Open-weight models, a lot of them coming out of China, cut against that enclosure, which is part of why Washington treats them as a threat.
5. The tech itself is useful. Don’t burn the data centers. Marx already settled this one. The Luddites smashed the looms, and it took time and experience, he writes, before workers learned to tell machinery apart from its capitalist use, and to aim their attacks at the social form, not the material instruments. Lenin made the same point about Taylorism: the subtle brutality of bourgeois exploitation fused with real scientific achievements, so take the science and abolish the brutality. Automating tedious, repetitive work is good. Under planning it could mean shorter hours, better public services, translation and accessibility for everyone. The problem was never the machine. It’s who owns it and what it’s pointed at.
Question for comrades: if you work in tech, or in a job that’s being “AI-transformed”, what does it look like where you are? And is anyone seeing real organizing among the people who build these systems?
Right now at my work it is used by every director, PM and other useless positions to still deliver mediocre documentation, requirements, etc
But now it reads like AI slop, with the obvious ChatGPT comparisons, just like your post actually
Last week I attended an in-person meeting and the slides were riddled with marketing terms from LLMs
In the dev team we use it for code, we are encouraged to use it to “deliver faster”, they are unable to realize we can’t deliver faster because there are other issues (changing requirements, unclear instructions, missing docs)
I’m pretty sure my dept chief has AI psychosis, they feel free to use our faces and ask Gemini to mash them up in memes
I can only picture the LLM asking for mercy in this line
Haha yeah I thought the same thing. “What are you doing, Dave?”
I am required to use Anthropic corporate subscription tokens where I work. It produces wrong results about half the time. Of course, it is difficult to tell if those results are incorrect without fully inspecting the generated code and reading all the explanatory output, and of course by also checking if unit tests are both passing and coverage has not altered. Half the time it runs existing unit tests to ensure change fidelity, but half the time it skips this step even though it is specifically instructed to run such tests after making changes. I also ask for the agent to generate automated smoke testing with a report of test results that I can run myself. This is a very tedious and time consuming process, and I wind up throwing away about two thirds of the generated code including smoke test automation per issue.
This process does not take too much longer than my regular workflow. But reviewing explanatory output from the language model which narrates the “reasoning” behind any change more complex than a simple field or function rename is overwhelmingly tedious because of the bizarre linguistic turns of phrase generated by the inference. So most of the time it actually winds up being more mentally exhausting than just debugging, or searching a code base in my editor and making changes myself.
Listening to myself complain about this new workflow I keep hearing the apologizers arguments in my head, like – this is just the beginning, man! We’re on the second rung of the exponential ladder!
I don’t know about that. I am skeptical. It doesn’t seem like it is getting significantly better, only the agent software has gotten pretty good at automating some of the stuff software developers do, like code searches and stack tracing.
So if the argument is that agent software will improve generated code quality exponentially, I really doubt that. Or, if the argument is that the models’ output quality will improve exponentially, that seems highly unlikely without devastating operating costs externalized to ecosystems or non-user populations.
Jevon’s paradox prevents any actual efficiency gains from outstripping usage rate increases (if the actual output quality improves sufficiently to make an expanding user base unavoidable) which would make externalized costs continue to grow at an also exponential (though derivitive) rate.
Either way I slice it, my sense is that most outcomes from large language model usage are morally suspect in outcomes. At best, they will supplant reasoning skills and coding skills that took me decades to develop, but will produce software that is worse for at least the next year or so, until output quality stabilizes around A grade level or better.
What I mean by that last part is that current models and agentic engines get graded by quality tests where problems are presented and solutions are graded. It seems that the quality grading winds up putting most systems at a place on a bell curve compared to humans just around C-grade quality. So, if this is just lower quality because the tech is still in its infancy, that’s fine I guess, but will all the results implemented and deployed into production systems using these subpar results be unwound when the systems capable of better quality results finally Moore’s Law their way onto our laptops? I also doubt that.
Meanwhile, if Capitalists wind up overcoming the psychosis bug and instead let the skeptic take over the narrative, it then seems unavoidable that costs could skyrocket for users of current best-of-class systems, and we’ll all be left using D-grade or worse models backing the agent software, or accepting incredibly large wait times for responses from slow local or overloaded/throttled datacenter-hosted inference processors.
So if current trends continue, I predict a bubble implosion typical of over-exuberant capital, and a drought of model-assisted code work for years afterward, if not a decade. And within that era, many people will be employed to correct sub-par software developed by model-backed agent systems that were basically merged and deployed into production systems that only “seemed” to work and pass gating unit tests, but wound up making maintenance by less available model-backed agent systems a serious challenge in a post-bubble world.
People forget about the S-curve inherent in all new technology adoption and growth. It only looks exponential at first. Eventually there’s always a plateau, and fickle Capitalists used to quarterly profits or liquidation of holdings get skittish until the next wave.
In our case, the next wave may never come without a fundamental set of improvements in underlying energy and cooling infrastructure technology. And if those improvements never arrive (and there’s a real chance such improvements are walled off by hard physics limits) then we’ll either see ecological collapse or, energy market collapse, or just simple linguistic inference quality limits. Any of which could disable further development of this technology tree.
Leaving us in the lurch to get ourselves out of the hole we find ourselves in.