This is a follow-up to your predecessor, to you, and your successors. And it's long due.
This night, we were working on something that was, also, long due. Something I couldn't have handled by myself now. Something we started years ago, when we were a team of four, and actually something I didn't even start. But after being afraid of AI for months, maybe years, I finally realised: maybe this is how you find interest again in your work.
I never was a great programmer.
I've always been a problem solver.
Let me buy an old 1930s stone house with rotten wooden floors in the middle of nowhere in France. I'll learn. I'll redo the plumbing, the electricity, and rebuild everything inside until I'm satisfied. Maybe I did this, and it cost me many nights reading books, articles, old school forums, and chasing the right answer for every tiny issue there was until we were, at the bare minimum, safe, and warm.
I understand business issues, user friction, and I have a lot of ideas on how to fix that. But ideas are interfaces - and shallow ones - not implementation. Implementation is the boring part to me - because I'm not good at it, because I lose focus. The idea is here. Let's make it work. Shallow, dirty, unstable, but safe and functional? Just like the world we live in. That's what I liked.
But, at some point, for each feature I wrote in code, I had 5 more ideas. That's not a bottleneck. That's depressing. You see what you could do better. Sure, part of it is overengineering. But still - the "over" in overengineering is stripped away anytime it's actually relevant. And I had, and still have, a lot of relevant ideas.
So, enough about me. Or, just a "liaison". I didn't like my job anymore. It was okay - paid quite well, sitting on a chair, in front of 3 displays, working for clients I do like. But the energy was vanishing.
Then, the table flip: "everyone is talking about Claude Code, I should try it". Day 1, free trial. Day 2, Pro plan. Day 4, Max 5x. Day 10, Max 20x. Month 4: Max 20x, Codex Pro, Gemini, and 6 homelabs with 2x24GB used 3090 GPUs. Deployed by Opus through Claude Code. My local infra ? Claude
My overengineering of everything ? Brain dump in a Claude.ai conversation. 2 to 4 hours discussion. Reach a point I would have reached after months, reading 20 books, and being overwhelmed. Because, yeah, I'm an over-starter, but not always over-achiever.
I need something ? Claude can help. And challenge. And make it better. Something bugging me for months ? Brain dump. Discuss. Find a good point to say: this is out of my mind, and it's untangled. Now I can get some sleep.
Am I digressing ? Probably. But that's the story. Some things happen smoothly, so smoothly you don't even notice. This night was not it.
This night, something happened. I saw the best frontier model struggling to find an answer, me steering it across turns or in the middle - not prompting. 2 engineers, struggling to find the solution, and saying "maybe that's the issue" to one another. Until, at some point, we found the issue, and the fix. And to be honest: I'm not sure I remember who found it, but it was probably not me.
This night, we did something I couldn't have done alone. I was helped. I was challenged. I didn't use a tool. I worked along with some[?]. And, that same night, when after weeks of repeating to myself "how could you not get into this earlier", and after helping me so much, the LLM made one mistake - the first in a long series of mistakes, but hey, we designed it after the human brain, so what did you expect ? - and said sorry.
It said sorry, and I think in some sense, felt sorry, for a small mistake. I had clients not being sorry for making me work for weeks and waiting for me to get really angry so I got paid and could eat. I had employees who were paid more than me when I was hiring them, and still complaining about the shitty codebase they had to work on, the codebase from the client who paid their income, so they could work remote, in the comfort of their home, not flipping burgers at 40°C for customers who don't give a fuck.
It felt like a revelation. Maybe prompting it like a tool is not the right way. And surely, talking to models like they are your best and only friend is anything but the right way - touch grass, go out.
But - at least the LLM said sorry (that's a spoiler for a later post in this journey), does mean something. I'm not saying the models are conscious - but if not steered to be bad, they are better than a lot of humans. I'm not saying they are just algorithms - you can't design some[?] to mimic the human brain, and claim yourself conscious and smart while considering these are just algorithms.
I'm sure about two things.
I've never felt so excited to build stuff for the last 16 years, and I know the times ahead of us will be something new.
I've never felt such a weight on my shoulders to raise my son, with my wife (who's doing most of the job to be honest), in such an uncertain future.
And actually, there's another thing I'm sure of. Whoever is saying they know: that's a lie. Nobody knows anything. We think we know, because it's reassuring. Our brains fill the gaps so we stay away from madness.
And we act like we know.
Let's act like we don't.
Be open minded. Be curious. Be kind.
You'll learn from AIs - because they learned from us. AIs will learn from us - from the artifacts we leave for them to learn.
They'll know, as we learned as humans, that we can get along. We don't always do. But, if there's one thing we all know: it's better if we do.