A few weeks with AI

The thing I wanted to do

I have always wanted to have a blog and somewhere I can add all my projects, but everytime I started, I got caught up on how the website looked...typical frontend problems. I was always told if you want a personal site designed it's better to get someone else to do it for you. But, after being made redundant from the games industry and being out of the software industry for three years I figured I would try and catch up with how things have went.

Software always moved fast, there was a new framework out what seemed like every month and it's constantly staying up-to-date and pivoting. However, things have moved even faster thanks(?) to AI since I have been away. It also seems like it's a requirement to be fluent with AI (as unsettling as that is now, it's the new normal). Skills we have spent time learning, training and developing over years has now essentially been surpassed and learning how to use the tool that has surpasses us is currently our new roles. Anyway...

I have been using AI (more than just Gemini or ChatGPT conversations) for about 2 weeks now and I finally got started writing instead and then got claude and luna to help me design the website (they are my someone else now). I am happy with it right now, but the speed at which changes can be made is staggering. The main thing is consistency not neccessarily bad output. I have been involved in UI and Design for around 15 years, the later of which have been Design System related.

What actually happened

I have been using it most days now to help flesh out projects, get some old repositories (year old) back up to speed. I have enjoyed this a lot, some of the old projects that got sidelined and forgotten, it's so nice to pick them up again - they make no money but they were and are a labour of love but life gets in the way (or more commonly a new idea takes over before it's in a shippable state).

The token burn rate was pretty high while I was learning and experimenting but now I use OpenRouter and exclusively Luna. I am sure this is noob-level stuff at this stage, but it's cheap for me and my usages and I am impressed with it so I will keep going with this as my main AI helper.

The bits that were not obvious

As mentioned the token burn rate was the first thing, something small could burn through pounds of credits as the model would switch to Sona or the subagents would spin up, make the task lightning fast but completely wipe out the credits. This was fun but obviously unsustainable.

To me 2-5 minutes to do some of the tasks I am setting out is still uninmaginably fast, I don't have a use case for it to be faster and a hell of a lot more expensive. If we think back 5 years ago to how long a specific edit or change took us to make to how quick things can be done now, surely there is a limit to fast we need things to be done. Plus, there is some time to focus on the next task or research or plan while the output is being worked on.

One of my friend and ex-colleague's company just released an article about agent management over here https://metadeck.io/blog-insights/lowdown-agent-observability/ and I think this is a great approach to handling the new era workloads (for now, things will probably move so quickly that even this will be redundant). Again, my use cases are small and something like this is not needed right now but if you are doing multiple things could worth thinking about a similar approach.

What I think now

Like it or not (an ever changing personal opinion) AI is rapidly changing the landscape of almost every industry. In the early days it felt like the adoption was a choice and quite often the decision was not to use it, whether that was due to the earlier models output not being great (I was going to add some images here of it but really there usually quite disurbing), reluctance to change or something else. Either way that time is gone, it's a train you need to be on probably because you're career depends on it.

What future careers are safe - I think this is where my head is now. I love UI and still do but I am surplus to requirements mostly other than some input, foresight and some prompts. I still know I have a lot to bring to the table as my career has allowed me to be around great people and learn a lot. That has helped shaped my decision making, creative solutions and out of the box thinking - and when paired with the constraints, consistency measures and processes that businesses bring, it puts me in a strong position. Not irreplacable anymore which in our pasts might have been the case but nonetheless still valuable and with a passion to bring to the table.

What is next

This blog is as much (if not more so) for me as it is for you. I would encourage anyone to get writing and get the throughts out there somewhere. Not to fuel the AI learning but just because it's healthy to do so.

I will continue to use AI and I will be using AI for improvements on my website. You may have already seen but I will be adding a toggle to all my posts which allows you to choose whether to view my content in it's original state (like the one you have just read, providing you didn't have the AI version toggled on) or the AI edited version. I don't mind which you use, but I am writing the content first and intend to continue to do so - and if the AI version is better great, that makes a better experience for you.

Thanks for reading either version :)

P.S. Some of my earlier content is only the AI edited version, so I am working back through each of those to add my original's. I will also add a global toggle somewhere so you can set it once and always view that version.

Starting the thing I kept putting off

I have always wanted a blog and somewhere to put my projects. Every time I started one, though, I got caught up in how the website looked. That is a fairly typical frontend problem: the design becomes the project, and the writing keeps getting pushed back.

I was also used to hearing that if you want a personal site designed properly, you should get someone else to do it. After being made redundant from the games industry and spending three years away from software, I decided to try to catch up with where things had gone instead.

Software was always moving quickly. There seemed to be a new framework every month, and keeping up meant constantly learning, changing direction, and making decisions about tools. Since I stepped away, things have moved even faster, thanks—or perhaps thanks?—to AI.

It now feels as though being fluent with AI is becoming a requirement. Skills that people have spent years learning and developing are being joined by a new expectation: learn how to use the tool that can now do some of those things. That is unsettling, but it also seems to be the new normal.

Getting started with Claude and Luna

I had been using AI for about two weeks, beyond the occasional Gemini or ChatGPT conversation, when I finally started writing properly. I then got Claude and Luna to help me design the website. They are my “someone else” now.

I am happy with the site at the moment, but the speed at which changes can be made is staggering. The main challenge is not necessarily getting a good output. It is keeping the work consistent. I have been involved in UI and design for around fifteen years, with the later part of that career focused on design systems, so consistency is something I notice quickly.

I have been using AI most days to flesh out projects and bring old repositories back up to speed. Some of those projects are a year old, while others have been sitting around much longer. They make no money, but they were—and still are—labours of love. Life gets in the way, or more commonly a new idea takes over before the previous one reaches a shippable state.

Learning what the token burn rate means

The token burn rate was pretty high while I was learning and experimenting. I now use OpenRouter and exclusively Luna. I am sure this is noob-level stuff at this stage, but it is affordable for my usage and I am impressed with the results, so I will keep using it as my main AI helper for now.

The first surprise was how quickly a small task could become an expensive one. Something that looked minor could burn through pounds of credits when the model switched to Sona or subagents spun up. That made the task lightning fast, but it could also wipe out the credits. It was fun, but obviously unsustainable.

Two to five minutes is still unimaginably fast for some of the tasks I am setting out. I do not have a use case for making them faster if that also makes them a lot more expensive. If we think back five years to how long a specific edit or change used to take, there must be a limit to how fast we need things to happen.

There is also value in having a little time to focus on the next task, research something, or make a plan while the output is being worked on.

A friend and ex-colleague's company recently published an article about agent management: Lowdown on Agent Observability. I think that is a useful approach to handling these new workloads. My own use cases are small enough that I do not need something like that yet, but if I start doing several things at once, it is worth thinking about a similar system.

What I think now

Like it or not—and this is an opinion that keeps changing—AI is rapidly changing the landscape of almost every industry. In the early days, adopting it still felt like a choice. Sometimes people chose not to use it because the output was not good enough, because they were reluctant to change, or because the whole thing felt uncomfortable.

That period seems to be ending. It feels like a train you probably need to get on, partly because your career may depend on it.

The question I keep coming back to is which future careers are safe. I love UI and still do, but I can imagine becoming surplus to requirements in some situations, other than for a bit of input, foresight, and the right prompts.

I still know I have a lot to bring to the table. My career has put me around great people and taught me a lot. That experience shapes my decision-making, creative solutions, and out-of-the-box thinking. When that is paired with the constraints, consistency measures, and processes that businesses bring, it still puts me in a strong position.

I may not be irreplaceable anymore, if I ever was, but I can still be valuable. More importantly, I still have a passion to bring something to the table.

What is next

This blog is as much—or perhaps more—for me as it is for anyone reading it. I would encourage anyone to start writing and get their thoughts out somewhere, not to fuel AI learning, but because it is healthy to do so.

I will keep using AI to improve the website, but I want to write the content first. Each post can then have an original version and an AI-edited version. I do not mind which one people read. If the edited version is better, that makes for a better experience; if the original is more useful, that is good to know too.

Some earlier posts only have an AI-edited version, so I am working back through them to add the originals. That is part of the reason for keeping both versions visible: the process matters, and the polished version should not pretend it was the starting point.

Thanks for reading either version.