Sit down comfortably and pour a double of the good kind, as this will be a long one.

WELCOME YAP AND JAB

In the previous episode it might have sounded a bit like I was taking a jab at IDEs. Good. Jokes aside, if you are not judging me for not using one, I won't make fun of you for doing so. To me the real question is about vendor lock-in, and becoming dependent on a tool that you don't really have control over. Yeah, yeah, I know you can do your own forever fork of VS Code, so let's pretend you actually would. I'll get back to my love of no vendor lock-in very soon, but let us first address the elephant, or perhaps rather the dinosaur, in the room. The whole IDE versus editor question, and which one to pick, is not a real question anymore. Sad to say, I know you wasted years fine-tuning your fingers to vi-vi-vi. I heard a fellow Dane say that handwritten and human-reviewed code is becoming like riding a horse. Yes, it can be done for fun and sport, but if you are doing it for real, you are getting left behind. Until recently I thought he was just being provocative, but the last few weeks have changed my mind, which, to be honest, is one of the reasons I opened my Cyber Saloon here.

MY VIEW ON LLMS

I want to break down my view of LLMs into oversimplified eras, and I should probably add two disclaimers. First, the following are not facts, but 100% my personal impressions and how I remember it. Second, these are my personal spare time coding methods, not the ones I use in company mode.

I made my first production SaaS tool before any LLMs were available to consumers. Looking back at it now, it was a cute time. I was reading tutorials and adapting templates and example projects.

LLM ERA 0: THE BROWSER CHAT TOOLS

Then came ChatGPT. I think I started using it around mid 2023. I tried copying single files into the chat, but it was not really good for anything besides docstrings. I then switched to Claude chat, which had something called Projects, where I could place files for reference, and it got a little better at writing code. Let's call this era 0 (ofc we are not doing MATLAB indexing).

LLM ERA 1.0: AGENTIC CODING

Next came Claude Code and Gemini in the terminal. I started using them around mid 2025, I think. This era is also where I want to place the birth of agentic coding. I remember the first night I tried Opus 4 in a TUI. I don't know how to describe the feeling, but I woke up my partner for a show and tell: this was going to change the world. I had never done that before, and learned that I should never do it again. I would even put it up there with the internet and smartphones. I am old enough to remember the days before those, and to me, Opus 4 in a TUI with full access to system files deserves a spot on the same list.

The models got better, and we started setting up elaborate agent workflows, with each agent described in detail in its own markdown file. In practice, it was one big company theater. Around the end of 2025, me and my tinfoil hat felt a sudden drop in both the quality and the allowed usage of Claude Code. I don't know if it was true or not, but it felt true to me, and I realized that I had become dependent on a single vendor, and not just casually, but in the oh yeah, let's 20x this, take my money kind of way.

LLM ERA 1.1: OPEN-WEIGHT AGENTIC CODING

This kicked off my era 1.1, the open-weight models. I switched to opencode and immediately felt like I had regained some control. I promised myself I would not get locked in again, no matter how impressive the next model's benchmarks were.

I was mainly using hosted and self-hosted Qwen and GLM models. My peers and I were skeptical of people praising every new model, and of paying a premium to stay locked in to one vendor. By the end of this era, we had reached a level where enough time, tokens and markdown theater could produce code that did exactly what I asked, and how I asked for it. In my inner circle we all agreed that this was actually all we would ever need. Ofc it could become faster and lighter, but we were sure the intelligence level had plateaued, so why would we need more? The missing part was the human touch. The main issue was that the models delivered exactly what was asked for, but without the elegance and design perfection that can only come from the hammer test described in the previous episode.

Maybe I am a bad programmer, and yes, I am called cowboy for a reason, but I have never ever coded the final version in the first iteration. You always learn along the way how to improve and correct what you thought you wanted. The mid 2026 open-weight models are good enough for exactly that kind of iterating, so they provide a massive development speedup. Ofc they make mistakes, but so do the best developers I know.

CRACKS IN THE COWBOY SALOON

Then the first cracks appeared in my inner circle. A few developers I look up to and trust went back to the closed models around the time of Fable and Sol. I assumed it was a phase. Another crack appeared one evening when I had both industry R&D people and university researchers over for drinks. A professor I respect a lot described how much his students, the PhD students in particular, could now achieve with LLM guidance. It made me wonder where this would end up, but since I am not really asking LLMs for intelligence, I more or less shook it off.

LLM ERA 2: THE RELAPSE

Then came Opus 5.5, with an extra large amount of hype from the developers I follow. One rainy night I decided to try it, mostly to confirm that I would not like it, so I could tell everyone they were wrong.

I spun up a sandbox on Hetzner, billed by the hour, piped a line (to bash) and met Opus 5.5. No markdown agents, just a clear description of what I wanted, and then GO. The results made me want to wake up my partner again... continued next week.