halfbyte

(This headline is meant as a pun. I have written this by my own, no LLMs involved, as will become obvious within a few sentences, I hope.

This is probably my tenth try to write a consistent take about LLMs and their impact. So this time, I decided to wing it, not caring for consistency, not caring for well-thought-out structure and such. Think of this as a stream-of-consciousness kinda text.

A few weeks back, for the first time in ages, I lost a client (For readers missing context: I am, when I'm not working on Depfu.com a solo freelance software developer) and not in a particularly fun way.

I had my issues with the owner of the company (No, I won't go into details, it's too fresh and too sensitive for that) from the very start but one of the things that lead to my very spontaneous and specifically abrupt ousting was my unwillingness to let LLM tech creep into my team's workflows.

And so it begins, it seems. What's funny about this particular case is that I know my sentiments are shared by large parts of the company's developers and that a preliminary “AI policy” had just been established that very clearly laid out that nobody will be forced to use LLM tooling. Well, consistency is obviously for losers, so whatevs.

But it did make me think a couple of thoughts again I have been revisiting over these past couple of years every now and then:

1. Do I miss out on something important if I keep ignoring this technology? Will I be left behind?
2. Will this technology persist in the long run in our industry? If so, in what capacity?
3. Will there be another AI winter?
4. What does all of this mean for my place in this industry and what does it mean for my ethical, social and ecological concerns?

I don't have answers to all of these questions, of course. Nobody is able to predict what's going to happen when the financing bubble pops and LLM provider suddenly need to deal with the difficulty of getting their customers to pay real prices. I have ideas, of course.

First of all: While there may be some specific knowledge required to set up LLMs and agentic systems around them efficiently and effectively, the basic idea behind all of that is that it should be easier than software development itself. So I don't really think I need to be afraid of being left behind in that sense. I may miss out on opportunities, of course, but that is priced in into my decisions.

Will LLMs be around forever in software development? I surely hope they will be replaced by something that delivers similar value but with more consistency and less ... randomness. For example: to have a technology operate entirely on textual representations when much better forms of representation in the context of machine processing exist (AST's for example) makes no goddamn sense to me. But right now, text extrusion is all we have so that's what we're doing.

My (faint) hope is that in the long run, the LLM craze will lead to a reckoning within our industry on questions of quality, but also how we want to treat the humans in our processes and companies and how we have maybe gotten off track on all of these accounts long before the LLM rush started.

I think, just by virtue of finding out the correct pricing for LLM tech, we will see that this tech will probably find places where it continues to be used, but I am also relatively sure that usage will be considerably lower than it is now, just because people will need some time to find out about all the places where it doesn't make sense to use it, or, rather, where it produces actual harm. I do believe we will see, at least in some form, a cultural, social backlash on LLM tech, pushing it back in some areas, because of that. My hope is that this will include things like security screening, application processes, etc. but only time will tell, if we, as a species, are clever enough to make these decisions. My hope is that this will come with a general backlash on our current species of unregulated, wild running capitalism, but it is also only, at this point, a faint hope.

Last but not least, what does this mean for myself. I don't think my stance, which I've been very vocal about, regarding the social, ethical and ecological impact will change unless massive change will happen in the way we construct and use these systems. This also means that while I do not rule out occasional (sometimes involuntary) usage of these technologies for myself, I cannot base my whole existence in the form of my professional career on such technologies.

If this means that, in the long run, there is no longer a place for me here where I am right now, working as a freelancer on web applications (for the most part), then so be it. I believe myself to be resourceful and skillful enough to find niches for me for the time being. It's also important for me to point out that part of this stems from the fact that with programming, I have made one of my hobbies as a kid by profession. This isn't always the greatest idea, but in this case it very clearly means for me that a job where I would replace actual programming with instructing and herding a group of agents would mean that I would have to abandon this thing that I always found immensely fun and satisfying.

What it does also mean is that I will double down on what I think are good principles for my own work. I have not always been taking these super seriously, but :

1. Focusing on quality over speed of delivery. Be deliberate about changes. Understand the systems you're trying to build and shape them actively instead of “letting things happen”
2. Focusing on the humans. Make software for humans, not for KPIs or someone's bottom line. If this impossible in a system you're trying to work in, leave that system. Remember, every system works as it is supposed to.
3. Focus on consistency vs. dogmatism. For me, it is much more important to be overall consistent in my values and my belief systems than to be dogmatic about them but then being inconsistent where the dogmatism reaches its natural limits. For example: I'm okay with implementing an MCP server for a service I run, when I know this helps other people doing their job, but I'll write that implementation by hand.