Martin Hähnel

How To Moderate LLM Enthusiasm At Work

Published:
Tags:

"Are we in the middle of jumping the shark?" is what I am thinking while this Video plays in the background at work: Claude Code Explained in 39 Minutes - YouTube.

But I have found a new way to talk about this in context of my company's size and resources: "We simply can't just run a "Gas Town" equivalent for every developer. We are still living in the "free money" era of LLMs as coding tools, but it ain't staying that way. Of course, taking the human out of the loop will almost certainly lead to code churn and in the worst of the cases disaster down the road. Therefore the plan should be to build something that can work without the use of LLM agents (like Codex) and with a human in the loop guiding the system. This tends to be more token efficient and is of course (more) deterministic than just simply letting a group of agents with different roles working on the same code at the same time with only light supervision."

This seems to work because it doesn't talk about things like the environment or social problems it talks about token amounts (money) and code quality (money). But less tokens certainly also means less strain on the planet. Doing it indirectly is less of a red cloth.

As with all tools there is a learning curve and a price. Certain approaches - like using MCP-servers are falling out of style, because they are less efficient than just giving agents full permissions and let them use cli tools directly (like I did yesterday with Playwright). But MCP-servers tend to be more secure - if certainly not fail-proof, but still - then just pointing an agent at a cli tool. I think that looking at the state of things right now, that having a human in the loop and working towards a mixture of planning/clarifying/exploring solutions (tracer bullet or prototype style) and iterate on the solution and incorporate refactoring and tests.[1] This tends to make us quicker in some things that get measured and ensures at least some contact with the code which is absolutely vital, especially when the free money ends, which it inevitably will.

What I don't see, is any kind of resistance about this "from above". And I honestly also don't see why they would resist this stuff. It is a survival fight on all fronts it seems. Since it's not only workers losing their jobs but also companies having to survive in an increasingly competitive market. And if that is the case, I think it is less surprising to see people, good people even, wanting to learn using this tech in an attempt to stay employable. The fact that it actually does do something, opposed to what is still often claimed, makes it likely that adoption will just continue to rise.

I linked the other day to an angry post blaming people "in front of them" for having to use LLMs and I said back then:

I suspect we'll see more things like this, because people need to eat, people need to feed their kids and people with years and decades of experience in the industry can't simply get a job in other industries (and in all likelihood wouldn't get the salaries that they are accustomed to that pay for the houses these people live in and so forth).

Here's another one (in two parts):

A quote from the overview that frames this pretty nicely (and these are good articles, too!):

Overall, I still believe that LLMs are a net negative on humanity, that the destruction of our infosphere is going to have generational consequences, and that if the whole thing disappeared from the face of the earth tomorrow, I wouldn’t be sad. The damage is would still be out there, but the cheapness of bullshit pervading everything would at least resume being human content mill scale. Not to say that that was good before LLMs came along and made it this bad, but it was better. That said, that’s not going to happen, and the amount of effort required to make it happen would be much better spent on organizing labor and climate action.

In other news:

  1. Not necessarily in that order or all of these steps only once. ↩︎