AI and This Guy

· 9 min read tech AI

Where it all started

In March 2026 my stance on Artificial Intelligence / Large Language Models (AI/LLMs) was fairly clear:

These things are bad in every way and nobody should use them

I don't think my reaction to this technology was unfair, I work in tech (and specifically, supercomputing) and am an advocate of open source software and open knowledge. My dislike of AI was for many reasons:

  1. Source material: It's increasingly unclear where training data comes from (the whole of the Internet?) and to what degree it is fair to use different source materials in this manner
  2. Environmental impact: The compute resource cost for running models (especially large ones) is massive (and a big driver behind the massive increase in DCs being built around the world)
  3. AI maximalism: The technology being pushed as a "cure all" for, well, everything without any gap for nuance or considering whether the application of the technology in certain ways was even worth it
  4. Shallow adoption: Most public AI usage stories highlight poor use cases and even poorer execution leading to creation of literature, artwork, code with little meaning or understanding from the creator

It's all well and good having come to that conclusion, however:

I work in tech (and specifically, supercomputing)

Is it really right for me to blanket dismiss an entire technology (or, arguably, an entire sector) when it's something that I am increasingly needing to be able to directly work with and understand?

So I challenged myself. One month. One month of using "the best that's out there" to see if engaging in the use of AI/LLMs could cure all my problems improve my work or personal technology usage.

The month

It was tough to decide where to begin, there were so many questions to answer; what model should I use? What cloud platform is best to run on? Should I purchase a subscription or PAYG?

In the end I settled on using the GLM-5.1 model and running it through Ollama's cloud subscription offering (~£18 for one month). This seemed like a reasonable ground to start on, the GLM model series is well reviewed, high performance, good for logical workloads and Ollama is one of the few providers of compute resource that offers a monthly subscription instead of PAYG.

By trade I'm not a programmer, I can program but I'm neither fast nor efficient. My focus is usually better spent on specifications, user experience and overall product management. Through this means I decided that one particular use case I could attempt some AI usage is in creating an application for myself. I also figured that, if I'm testing it, I should go "all in". That is to say that I should try and use AI everywhere I'd potentially use "traditional" technologies.

My month of testing saw me use AI to:

  • Create an "office games" application: A full stack web application for tracking Elo-style skill ratings across foosball, table tennis and pool.
    • What AI did well: Initial scaffolding and boilerplate of the application, from nothing to a simple working prototype in minutes.
    • What I wouldn't trust AI for: Anything to do with security of a full stack application, that's why this project is deliberately an internally hosted, insecure webapp.
  • Prototype a number of websites: As part of my work I find myself needing to translate product ideas into digestible formats, in some cases I create websites to deliver the message of products. I use AI to build various prototype websites based off of the information, structure and styling that I already had in mind.
    • What it did well: Rapidly came up with multiple projections and views to have visually pleasing results in minutes (I suck at CSS).
  • Structure technical discussions: As I sit in a technical product role I am able to discuss and distinguish technologies however the deeper technical knowledge is sometimes beyond my immediate grasp. I discussed sticking points in my understanding of how feasible an idea is with an AI chatbot.
    • What it did well: Helped me reach a clear enough understanding of the deeper technical considerations. Doing so in what I would consider a much faster time than if I had researched myself.
    • What I could have done instead: Talked with a technical expert.
  • Reshape and structure meetings/documentation: There are times where I have a clear goal of what I want to communicate but can struggle to frame or structure things in a way that feels most productive, especially if there are points of contention to be dealt with. I had thorough notes on the things I wanted to document/discuss and rubber-ducked with a chatbot to reach a satisfying structure/delivery.
    • What it did well: Helped me reflect on my position and purpose and follow different possibilities rapidly, provided structure for meetings to best deliver my thoughts and targets.
  • Creative writing: I enjoy writing comedy and jokes and decided to see what it would do if I tried to write some jokes with it
    • What it did well: Structured a joke with a setup and punchline.
    • What I didn't like: It felt... empty. I think this is a wider reflection of AI for creative purposes. In my opinion the beauty and weight of art comes from the human behind it crafting it for a reason.

How to use it

As mentioned initially I had a strong dislike for the shallow adoption of AI so sought to do anything other than "throw in, ship out". How I approached using AI at all times was to:

  1. Have a clear vision: Have a thorough idea of what I know and what I want to achieve going into the initial engagement.
  2. Prepare modular goals: Ensuring I had small, achievable targets made it easier to follow along, review and understand each step of using the model for whichever purpose.
  3. Manually control changes: AI never did git commits for me, AI never directly edited my notes/documents. Everything that came from it (either directly as in code or indirectly as writing assistance) was reviewed, tweaked and approved by me.
  4. Manually test changes: This one is more specific to the application use case but I'm a pretty thorough acceptance tester so I would extensively test and attempt to break and understand all software changes in order to ensure functionality.
  5. Tweak and tinker: For the most part I find that there are things I want to adapt with tone, structure and delivery of writing-related AI work, copying+pasting is one of the things I try to avoid unless I feel like the words match what I would say myself.

I don't just believe this a best practice but the reason I got the most out of my experience - this is the reason my results were good and perhaps the reason why shallow adopters have low quality output. (A factor really can be who is driving the car)

The result

I went into this experiment expecting very little yet I was really impressed with the capability of the model and the structure of the tools (I used Claude Code on the CLI for programming and chat activities, I've since branched out but that's for a Future Blog Post which will probably be written within the next 1-100 months).

I'll admit that my low expectations came from my preconceived notions and AI was by no means a fix all to any problem. There's certainly a gap that exists still between some of the hype around the technologies and what can be achieved when pointing it at a task you understand.

It's important, I think, to only be using AI agents to enhance your own abilities and not to do things you can't inherently do. This is because oversight is key so using it in contexts of things you understand can minimise unexpected outcomes. I didn't use AI for anything that I couldn't reasonably see myself being able to do given enough time (e.g. digital art, writing stories) and focussed on aiding me in improving what I can do (writing documentation, iterating on product ideas, evaluating business logic, creating websites, programming).

Where AI didn't work

As mentioned I tried to use AI in every facet of my day-to-day where I could (because it's the solution to every problem, right?) so there are certainly areas where it's useless, usually because it's the wrong tool for the job. For example:

  • As a basic search engine: Short, simple web searches are not something AI will replace. Hitting ctrl-T, typing a few general key words and getting the technical guidance or wikipedia article, etc is miles quicker than awaiting a response from an LLM. Unless the query is the start of a research thread (something I'd expect to open a few search result links for and cross-reference) then I'd keep well away from AI as a search engine. This is also partially due to models tending to lean into being verbose and nobody wants an essay (except maybe you reading this?) to answer a simple question.
  • Creative writing: While it may do well with structuring logical and technical documentation, anything that involved creative writing felt vapid and it was difficult to pin down why. It gave me the same feeling I get when I see an image online that's clearly generated by AI, there's just no meaning to it.

Where my mind didn't change

Despite having quite an eye-opening experience with AI which shifted some of my attitude towards it, I still feel concerns about

  1. Source material: Despite positive results for some of my use cases there is still a big question mark over the source ethics. In my opinion, if you're digesting the Internet for a model then the model should be fully and openly returned to the Internet.
  2. Environmental impact: Having used some frontier (at the time) models I certainly did contribute to larger impact. Going forward my goal is to seek local-first model use cases which I think is fairly achievable for general chat stuff but only really limited when it comes to large scale software projects.
  3. AI maximalism: Yep, anyone who still thinks it's the solution to every problem is annoying.
  4. Shallow adoption: As I covered in "how to use it", shallow adoption does give usage a bad name, that's a user problem (not a technology one!)

Resolution: unresolved

So, where do I stand now?

I think that AI is impressive and that there are a number of ways in which it can help streamline and optimise some workflows but equally I am concerned at the general landscape of the technology and it's resources. I think it's healthy to stay mindful of the concerns and issues related to such a technology while it's not contradictory to still use it (or, at least, there's no more dissonance with this than many things we engage with in the modern world).

DISCLAIMER: This post was entirely written by a human, the views are entirely that of a human, the structure was aided (not decided) by AI