Thoughts & Articles
What I’ve learned about AI (and why I’m not using it)


Dave from 2001 a space odyssey
In my very early working life, in 1990, I spent some time doing graphic design at a company called Cambridge Technology Solutions (later Cambridge Technology Partners). I created slides and presentations, mostly. I was barely paying attention, focusing primarily on the next open mic or gig. Music was my main focus back then.
I remember the titles of the slides, which included “TCPIP - Tim Berners-Lee,” “Mosaic Browser - Marc Andreessen,” and “Networked Systems.” It all sounded like cultish nonsense to me. Everyone was very excited about these things. I wasn't and wouldn't be for another seven years. One however, stands out in my memory, and that was a slide titled “Machine Learning - Expert Systems.”
My point is, that machine intelligence has been a focus of the tech-forward people for decades now, even within my own lifetime. Machine Learning was a term for computer training models coined in the 1950s. It's not a new idea. It is a computer program that strengthens its internal connections based on input data through repetitive training. This is the basis of the modern LLM, which is marketed as Artificial Intelligence.
These systems are giant autocompletion programs. You may have heard this before, and it's hard to believe that all the models are doing is statistically computing the most likely next word, pixel, or sound based on their training data, but that’s pretty much it. The idea that they can or will evolve into something that thinks is simply marketing hype. The reason it seems like these models (ChatGPT, Claude, Gemini, Grok) seem smart is due to the incredible size of the computers behind these models, not because they have a neurological structure akin to our own.
But so what? If they work to help design things and write code, what difference does it make that they are mere imitations of experts? Why not use them to write more code, design more websites, and do it all for little to no cost?
There are several problems with this.
- The code these models create is derivative, buggy, and insecure. I fed Gemini a series of functions for it to rewrite into an new format, and because none of the code I provided existed on open-source libraries, it rewrote a small fraction of what I gave it and ignored the rest. When I asked it why it skipped so much, it told me that it had completed the job. I won’t say it lied, because the answer it gave me was not based on the results it provided. It was simply writing the most plausible sentence to answer my question.
- The designs it comes up with are also derivative. This doesn't mean that a model might not create a good looking design, but those designs are better researched by searching out competitive websites within the client’s field, and creating a more appropriate look for the client’s brand with those sites in mind, rather than taking a randomly generated derivation.
- AI services are cheap now, but they will not be cheap in the future if the major companies providing the models do not start becoming profitable. As I write this, Open AI and Anthropic have already tried charging a realistic rate to cover their costs, and companies are pushing back hard. In an attempt to capture market share, both afforementioned AI companies have slashed their prices to almost nothing, hoping they can raise them when their models become indespensible. This is not realistic. Uber created genuine demand before raising their prices.
- AI code is opaque. It may feel like a miracle to get a bunch of code dumped on your doorstep while you sip your espresso, but to ensure it doesn't expose sensitive contents from the database or die when actual people start using it, the engineer (me) will have to go through it line by line to find the security holes. AI code is generated with zero accountability. Add to that the problem that all this code will need updating as needs change, and AI models don't update. They simply rewrite the entire application from scratch. These models don't understand, they just iterate. This creates major issues down the road that I, for one, don’t want to encounter.
- Writing code is only part of the job of building a working application or generating a website. The other part is planning for human interaction, information architecture, personalization, and user experience. The models will generate something that looks good at first glance, but the devil is in the details, and the details involve how it is experienced by a human being, not another robotic model. There will always be the need for a specialist who undertands all the parameters of a task. This requires revision, not re-iteration. It's a difference that seems subtle until someone tries to build something that works reliably.
There are other problems, but I feel like I've bored you enough. Even though this is all very exciting, and in many years time, we may indeed have Expert Systems that are actually experts, I'm going to stick with my slow, hand coded, hand designed craft for now. The promises that have been made of late are all too good to be true.
Do I need to say that I wrote this article without a chatbot? I wrote this article without a chatbot. The errors are all my own.
–Jim
P.S. This is a song I wrote about AI psychosis. There is some hyperbole in this perhaps, but also some real warnings.
If you want to read more from the point of view of another skeptic, I recommend BetterOffline by Ed Zitron.
This video follows much of my thinking above but with references and statistics:





