I figured I’d take some time to jot what I think about AI. I’ve never formulated a cohesive opinion on it and this should help me whenever I have to discuss it or evaluate its part in a project. Let’s break this whole thing down. I will likely add more in time, but this is a good start for now.
My AI journey
I started using CoPilot and ChatGPT very early when I first heard about them (2022-3ish) and realised they could be a very powerful programming tool. I initially found them to be like an extremely powerful autocomplete and very helpful for generating smaller chunks of code. Things I still consider them to be good at. I experimented with using ChatGPT to create a Dungeons and Dragons DM after my groups moved away. It couldn’t even come close. I used ChatGPT in a hackathon to communicate Grasshopper scripts to it ^[1], it did very well. This project has been continuously cited by various academic papers which has been quite exciting. Eventually I realised I might be leaning on AI more than I’d like and decided to stop using it as an experiment. I’ve still never shipped code AI generated that I haven’t read, and I’m glad in retrospect I at least had that pride in my work. I’ve largely stopped using AI for programming (occasionally using it for debugging). I’m at an impasse where I cannot decide on the correct point in a programmers career to hand them AI. As an Architect my first year I was told to never use CAD and only learn using pens and pencils to learn the importance and intent of every line. I think AI should be treated the same way, but I don’t know how many years or which concepts I should have mastered first.
UPDATE : 16th May 2026
I am now using Claude Code extensively and very much enjoying it doing lots of awful, boring grunt tasks and code research for me. Prototyping and iterating has become very quick and fun. I’m also using, and now creating, MCP Servers.
UPDATE : 17th July 2026
I am still using Claude Code, but the honeymoon period is wearing off and there’s something not quite right about it. Whilst I find Claude very useful for certain tasks, and do enjoy some of these, it does not spark joy. Writing code and developing applications I care about with it is absolutely joyless, and I’ve only just started to realise it. I’ve heard it said users don’t care if you vibe code, and I think this is largely true. I find making something good, something polished, crafted with love, is not possible (for me) without being in the weeds and doing this myself. I spent my day today iterating on menus. Just menus. I spent all day trying things, learning things, bashing my head against the wall and discovering things I did not know. The end result was something that a Callum from yesterday could not have achieved without this process. And as a lover of the journey, I can see what I have been missing whilst experimenting with AI. If not for the MCP work I have been doing, yesterday would have uninstalled all AI from my machine.
Ultimately my goal in this life is to seek and bring joy, this tool does not bring me joy.
AI Revolution
I’ve heard people compare AI to looms and the industrial revolution where we saw a huge shift from manual human labour to automated machines in factories. I don’t think this is a good comparison. I’m from Manchester originally, I’ve seen (and used) manual looms and I’ve seen automated looms, they’re an incredible thing to behold. But there’s something that seems to be ignored when comparing AI to other parts of human history. AI is not determinate. A loom is very much f(x) = y. A loom set to a given encoded program will produce the exact same pattern every time. LLMs are non-determinate ^[2], they do not produce the same output for the same input.
I do think AI is likely to be a revolution, largely because of the gargantuan amount of human power and money that is being invested into pushing it along. Industries are clearly betting on this technology and there are areas that have found AI to be absolute game changer (e.g OCR, building renderings, girlfriends for losers, debugging programming, and other academic areas). That being said, I think LLMs are the start of this, not the end of this. They are WAY too unprofitable ^[3] to run at present.
UPDATE : 16th May 2026
It seems this determinative thing does not really matter at all. I’m able to iterate and get claude to create me things I need. Even if it does a slightly different job each time, this feels negligible. That being said, I am not sure how long this will last, or if this is a good comparison, the loom had a very determinative ROI, whereas AI has so much proprietary dependence.
What has spurred this update and change of page? In a word, Harnesses. The way Claude Code functions is absolutely excellent. I now use it every day and enjoy doing so. I read a long time ago that the winner of the AI Revolution will be the one who makes the best box to put AI in. I think they are right.
UPDATE 17th July 2026
After using Claude in the area it feels like it was designed for and has the most training for. I have to say that it is untrustworthy and and the solutions it comes up with are uninspiring and very often not good. It lies to me constantly and then admits it after the fact. I have found that when I want to do something I care about, this quickly becomes unacceptable.
Ethical Concerns
It’s of course worth noting that AI data centers in the US are built unethically and degrade the communities around them. Consuming ungodly amounts of electricity and wasting ungodly amounts of water. I wonder how many tokens is equivalent to a mile of air flight for energy used?
My next concern is for artistry. Image generation is trained on stolen work, text is trained on stolen stories, all of which was created by talented people who have as yet never been compensated for this.
If I create and expect to be compensated for my creations, how can I use AI un-hypocritically?
Trust
The longer I code, and the more I try to write programs to last forever, the more I evaluate and consider the trustworthiness before incorporating a technology, tool, library or language into my stack. Is it open source? Can I understand its code? What dependencies does it have?
I don’t trust OpenAI, I don’t trust Microsoft, I don’t trust X/Grok, I don’t trust Claude, and I don’t trust any for-profit (AI) company. They are unprofitable, and if they can make it past the ponzi scheme phase into profitability, that will, as usual, become their only concern. The AIs will start cramming adverts in, running at lower, dumber levels, and prices will be raised until morale improves amongst the shareholders.
So any company offering AI I can’t trust that it’ll be around in a year, two years, that the price will be reasonable or even how it works.
Local models largely resolve this problem. I could download a model from huggingface with some python code and stuff it onto a Raspberry pi that I never connect to the internet ever again.
Responsibility
ADDED : 4th June 2026
I vibe-coded a small project in typescript to automate tasks locally with some llm-y-ness inside it. I have not seen any of the code, and I’m quite astounded how well the first version worked. After some more hours of asking for improvements etc I went for a walk and asked claude to whirr for an hour and just improve things. It did a fantastic job.
Why would anyone ever hire a programmer again? Why involve humans at all? Why pay for a service?
I think ultimately where the rubber hits the tarmac is human to human responsibility and trust. If a service is worth paying for, let’s use Figma as an example. Figma is worth paying for because it handles all of the connectivity and hosting. Even if they 100% vibe code it, as a consumer of the app you expect it to work, and if it doesn’t you’ll stop using it and sales drop as a result.
If a human receives your service, they expect it to work, and if it doesn’t, the seller is held responsible, either by losing a customer + money, or violating an SLA or agreement the two parties have together. (This only gets more and more important when you’re dealing with sensitive data etc). In which case, someone at canva, who again, may 100% vibe code this has to be willing to take responsibility when the axe falls (e.g job loss). And the only way I’d be willing to take that responsibility is if I understood the code.
ADDENDUM : 17th July 2026 I think too that basing your business on something that could drastically change price, lie, fail is irresponsible, that said, capitalism is system that promotes responsibility, so maybe more unregulated countries will see many business that do. I think that if countries want to promote human skill, then increasing the severity and ease of punishment for things like user data leaks and SLA contract downtime would ensure companies have a human they can trust managing those parts of the business.
Consciousness
Imagine you are in a burning building. In front of you are two things. A cat, and ChatGPT on a thumb drive, the destruction of which will mean ChatGPT is gone and can never be restored. In effect it will “die”.
Which do you chose?
If your answer is not the cat, then I think you’re using your consciousness unconsciously. You’d save the cat, a living creature, with thoughts, memories, dreams and the ability to love. Nothing is worth more than that.
LLM AIs are emulating these things, they are not doing these things, they cannot surprise us, upset us and then make amends, or ever care about us.
^[1] https://github.com/enmerk4r/GHPT/
^[2] https://152334h.github.io/blog/non-determinism-in-gpt-4/
NOTE : It does seem that there are and in theory LLMs could be run more deterministically because they are a function, just with a billion parameters. To expand on this point, if I change a small thing in my code, the output will also change slightly. If I change a prompt slightly, the output could change quite drastically. I think this is an important difference when comparing a loom and AI.
QUOTE : Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns.