Demystifying AI stuff by Claude (the AI)
Looks like AI is here to stay, which means I need to get good at it. I had a long conversation with Claude about some confusing things, and you can read it too!
For a rather long time I avoided using AI for anything after Dall-E and the other one ... uh... Midjourney! I had to ask others what that was called, thank you Alan! My reasons were the usual:
- LLMs trained on copyrighted materials don’t sit well with me
- any and all answers boil the ocean and burn forests, basically use significantly more resources than what I’m okay with
- using AI for art is just not great, because I value the things created by actual humans, whether it’s digital art, text, whatever. I like humans to get paid for their art
- AI results hallucinate, and the results are more incorrect than not, which makes them unreliable, so I have this notion of my organic, artisanal, handcrafted code is going to be significantly better
Yet here I am, in 2026, somewhat begrudgingly using AI in my day job. It started with ChatGPT that I’ve been using to troubleshoot certain issues, find documentation for things, and creating short scripts and commands to solve tiny problems.
And then we got access to GitHub Copilot CLI. I saw a stream by Cassidy Williams on how she used it to one-shot create an app, and found it intriguing enough that I tested Copilot on four ideas I had for corporate use. Sadly I can’t reveal them, because confidentiality clauses and whatnot, but the results were shockingly good. The last one went from an empty directory to a fully working service that I can deploy to Kubernetes if I so wish, and it has full logging, a usable UI, an API, unit and end-to-end tests, and the service solves an actual problem for me. It also took roughly 3.5 hours and produced ~6,700 lines of code.
Three and a half hours. Not too long ago writing a service like that took several weeks for multiple people working together. I don’t know how I feel about this. I guess it’s a mix of “for f*&ks sake, why is it good enough to work?!” and “I am genuinely impressed.”
And, AND, the worst part? I looked at the code: most of the ~6,700 lines.
I don’t hate it.
It’s well organised, it’s tested, it functions. More importantly, as a human, it is very easy to understand what’s actually happening, because the end result looks like somthing I would have, and have in the past, created by hand over the course of several weeks or months. Yes, there are bugs. But it also took me (it) 3.5 hours to create it, and during most of that time, I was away from my desk talking to my colleagues or grabbing a drink, or peeling an orange, or walking to lunch.
As much as I hate to admit it, I will be using AI tools in my job. I need to figure out how to deal with the ethics of that and the trade-offs this means.
That said, I had multiple gaping holes in my understanding of the current state of coding using AI, so I downloaded Claude desktop to my personal computer, signed in, and decided to have a chat with it, and make that chat publicly available.
It is very long, I asked a lot of clarifying questions, and at the end of it, I feel like I am comfortable talking about AI with other humans, and I can begin to build things with it in a more refined way.
I touched on the ATLAS / GOTCHA frameworks, see https://www.youtube.com/watch?v=_GAc6SFoQ9k (I Made Claude Code and Codex Build the Same App (RAW RESULTS) by Mansel Scheffel).
I asked it about the meaning of a lot of terms used with AI, like inference, quantisation, what the heck 7B means.
I asked it to tell me questions I should have asked, but haven’t yet.
I touched on Eric Holmes’s article on MCP is dead. Long live the CLI, where Claude tried to push towards MCP is good, actually! I argued with it and got Claude to concede. Somewhat.
It’s a really long conversation, roughly 22,000 words, including code blocks. It might take a bit to get through it. On the plus side, it’s a genuinely interesting read!
Click this, and read the entire conversation:
I’m very much at the beginning of using AI for coding, but I’m an experienced software engineer.
If you’re great at AI, and notice that I have an assumption that I haven’t challenged, or I’m missing something, please let me know!
I hope you got something out of that conversation. Let me know, either good, or bad, and if you’d like to read similar things, smash that subscribe button sign up for the newsletter:
Photo by Masjid MABA on Unsplash