I wanted to write some thoughts on this topic as a letter to my future self. I will likely have a good time looking back at this in 2035, and compare it to all the progress over the years since then. The current state of AI is the worse it will ever get, and it is already incredible in many ways.

It really feels like where in this AI whirlwind right now. There are new tools and solutions popping up all the time. Models are getting more and more complex, and everyone is talking about AI right now it feels like. I know I am kind of in a bubble here, working at Sana. But still, my parents also see this all folding out. Perhaps not in the same as I do, but it is not something just for the people in tech. In a few years time we will be able to reflect on how AI fundamentally changed the world, for the better I hope.

I have been adoptiong various AI tools gradually over the last two years or so, and now I am at a point where I use them continuously. Sometimes I find myself doing some kind of work that could obviously just be done for me, and I feel stupid. I should have asked AI right away. I use AI a lot, but it is still not the default for me. I actively reach for it very often, but my brain still wants to solve problems “individually”. I am guessing this will change over time, but I have very mixed feelings about this.

This is what my current stack of tools looks like:

Asking questions, bouncing ideas, and getting quick answers

I am using Raycast as kind of general purpose AI UI. I have not used ChatGPT that much, I started using Raycast pretty early. It is very well integrated to my operating system, and most things are at my fingertips. It gives me quick access to a bunch of models, well optimized for quick one-off questions and bouncing ideas. This is often the first things I go to, it has mostly replaced search engines, or going to a specific site to find information. I often trust the model to give me what I need. Sometimes I click through to the sources to understand where the answer came from.

I often type out a question, or even just a few words, and hit Tab on my keyboard. This launches their “quick AI” which is just amazing for getting a quick answer. From there I might go into a more detailed chat, sometimes even with a more complex model. I do not overthink what model I am using. Sometimes I am switching when the response I get is not at the level of detail I was hoping for. Or when it is just completely wrong. But I very rarely pick a model first, and then ask it something. That is not intuitive to me right now. I might work on my understanding here to improve on that, but at the same time I am just looking for a quick answer to something that popped up. If it is important to go deeper, I have time to think.

Producing code, documentation and understanding systems

I do still use Neovim daily, but it is hard to not adopt Cursor, it is likely this form factor I will be working with for the coming years. If neovim had something better available, I would likely use that. I have been loving that setup recently, configuring it to be slim and easy to flow with. On the other hand, neovim is a great tool for writing code. I am not so sure that is what I should be optimizing for.

Overall I am pretty happy with the current state of the AI tooling for writing code. I still feel like I am in charge, and I can easily turn it off. I do not have to be fully locked in just writing prompts. I am still driving the car. I have a feeling though that this will change in the coming months or so. With the rise of tools like Lovable and v0, it is almost like the authored code is an implementation detail. Not really, but in a way it is less important to me as the “architect” what the code is, if the AI tool is always interacting with it.

Producing text

I have recently started dictation more and more. The quality is so good now, I often find the output to be good enough right away. It depends a lot on what I am using it for of course. If I am just vibe coding and going back and forth in Cursor I am not even reading the output. But for a writing an email or a slack message to the team, it gives me a great boilerplate to polish.

Other

I am also interfacing daily with AI-powered code reviews. I actually really like these, but they are not bullet-proof by any means. Sometimes they find useful points are perspectives on the changes I am making, but most of the time it is just helpful as a sanity check. It saves my colleagues some time also, they do not have to context switch back and forth just for me to merge a quick CSS fix.

---

To summarize: I am using AI many times every day. I use it to produce code, raw text, ask it questions and reflect on ideas. I often feel like I am not giving it enough context, so I spend more and more time on that. Preparing and planning, setting up the right context, before actually getting started.