Conflicts & Agents
Being a person that's in the IT space both professionally and as a hobbyist. I've been feeling increasingly pressed about the ever-forward marching of our capitalist overlords. Their influence changed not only my boss's mindset towards using LLM's. But also the mindset of most (Clout seeking) developers online. And while I do acknowledge that the people I'm seeing are just the latest installment of "Crypto-bros". Something here is different.
Because the use of LLM's is being framed as a force multiplier for people who write code. Shifting the purpose of our roles from designing and building, to designing and delegating. And for the vocal clout-chasing group of developers this isn't an issue. Since they're in tech strictly to make money. So more output directly correlates to more value for their bosses/customers.
Because a large part of (at least my own) perceived value is in building cool stuff while enjoying the process. Having an LLM do part of the work for me is simply not enjoyable. I love toiling with design patterns. And building stuff that looks and feels satisfying. Which in of itself does not scale well with the use of these LLMs.
Then there's the issue of ownership and responsibility. I'm responsible for the code I write and whatever I build. And so are the LLM tech-bros. But I'm writing small projects. Things that only a handfull of people use. So the impact remains small whenever I make a mistake. These tech-bros are generating lines of code like never before. Barely reviewing new features. But they're still responsible for the products they create.
In the words of Cory Doctorow: Code is a liability (not an asset)
Of course I'm not immune to the AI-agent craze. What makes them tick? And how is such a tool constructed. So I started working on one myself, hand-written of course. Because the interesting part for me was how people steer these token-predictors into performing a task.
What I found was a whole lot of 'vibes'. You query an LLM in a loop with your request, appending its previous output. And if the output satisfies your initial request. You make it stop looping. Its performance is solely dependent on how well you can capture the users initial request and optimise what the llm receives as input for each iteration. Antropic, OpenAI and open-source parties like pi.dev and Opencode make it seem like some magic process. But really its just repeatedly pulling a slot machine. And correcting if you don't get what you want.
While I love diving into one of these projects. I feel conflicted about continueing. On the one hand, I enjoy the process of figuring out effective patterns and building a product. But I'm also contributing to the AI craze. It might not be AI-slop. But it is still AI adjacent.
