Opinion on AI and the future of IT
Current State
When a statement starts with “people often ask me …”, you know it never happened, but during the past 9 months of job interviews, I was asked “What are your thoughts on AI and agentic development?” on a regular basis. Once I found a job, I forgot about the question until I met a friend at a store who asked me pretty much the same thing. My answer has always been “It’s like having a really fast bike, car, or Formula One”. What it means is that it depends on the driver: if it’s a lousy driver, the crash is going to happen pretty fast, but if the driver is good, everything will work perfectly.
That being said, I consider myself a pretty good AI driver. For the past year or so, I have worked on a Digital Audio Workstation, or DAW for short, in Rust. At first, it was a way to practice what I learned from the Rust book, but then it became a serious project. After reading that book, I needed a problem to solve in order to really understand the language and libraries, so I started rewriting Maolan. I started writing Maolan in January of 2018 in C++, and as I really liked the Rust language, I decided to give the rewrite a try. It went really well, and after two months of writing code, I had an audio backend, an engine, and the foundation in place. I have to admit, I was inspired by my ex-colleague, Alan Somers, who really likes Rust and whose opinion and knowledge I highly respect.
After writing the most important parts by hand, I decided to give AI a try. So far, I have tried OpenAI, Claude, Gemini, and Kimi. Gemini is not something I consider a good AI agent, but the other three are great. To be honest, Kimi costs the least, so I have used it the most. AI writes code insanely fast, and most of the time I have some nitpicks about the code it wrote; sometimes I just say “forget it, I’ll do this manually”. The reason is that AI doesn’t really know what you want, and if you’re not precise, code gets bad pretty fast. As you can see, I keep saying “fast”, and that’s why I think the fast vehicle analogy is good. In any case, Maolan is at the point where I started recording the first song using it, which revealed a bunch of bugs that I fixed.
My background is that I started playing guitar when I was 6, producing trance music when I was 16, working with electronics and schematics since I was 16, and programming since I was 19. Now, in my mid-forties, I have a lot of experience that makes me an expert in the field of audio/MIDI DSP and music production. I know exactly what I want and how to get it, and I know exactly what the hardware and kernel are capable of and how to use them. With such experience, AI is a huge boost, and when it goes sideways, I can tell right away.
As an experiment, I worked on a project in a domain I knew nothing about. I wanted to see how badly it would turn out, even with me as a senior developer and a pretty good IT architect. Within two months, I stopped, learned about the technologies I knew nothing about, read the AI-generated code, and threw it all away. That was me being a bad driver to prove my point to myself, as I hate having an opinion about something I’ve never tested.
I think I found my way of working with and learning from AI. Find your speed and don’t drive above that limit. Don’t use AI only to develop something; ask it a bunch of questions. I think a quarter of the tokens I spent in the first 6 months were for coding, and the other three-quarters went to AI explaining things to me and telling me what to read. Use AI as your teacher, as it can improve your coding skills a lot. And I mean A LOT!!!
As for programming languages, I think stricter ones are going to be more popular. My experience with AI and Rust is that Rust’s strictness makes AI do a better job. For Rust itself, when you add tools like rustfmt and clippy, along with tests, the results are even better. Tools like Miri can additionally guide AI models on how to write better code. Today, most of the new AI-written applications I see are written in Rust, and I think the strictness has a lot to do with it. If some future programming language becomes even stricter, I think it will be even better for AI-assisted development. Maybe some future tool for an existing language will catch bugs by analyzing the code even better; maybe it doesn’t have to be a new language. In any case, the more help AI gets, the better the code it will write. And not just AI, humans, too. When I started with Rust, I hated it because it prevented me from writing code the way I was used to, only to later realize why it does that and what the benefits of writing code the Rust way are.
The Future
The logical question is “what’s next?”, so this is my prediction. Looking at the history of IT, everything starts centralized, then big companies have it in-house, then smaller companies have it in-house, then everyone has it. From ENIAC to home PCs, mobile phones, and tablets, it’s always pretty much the same. That being said, I think something similar will happen with AI; the only question is how exactly. I think the first step will be the widespread adoption of systems with large pools of unified memory, allowing today’s huge models to run on tomorrow’s personal machines. Of course, at first those systems will be more expensive than a gaming PC, so only companies will be able to afford them, but over time they will reach the masses, and it will be normal for your AI model to use most of your RAM during development.
At some point, I think we are bound to switch to optical computers, for two main reasons: power and speed. When I say power, I mean consumption of electricity. Moving electrons through resistive materials consumes energy and generates heat; photons can avoid those thermal losses. As CPU clock speeds have hovered around 5 GHz for quite a few years, much of the performance scaling has come from adding more cores. For processes that can be parallelized, that’s great, but sometimes you can’t do that. For example, in a single chain of audio effects, each effect must be processed in order, so multiple cores do not speed up that chain. Optical computers could potentially reach much higher frequencies. The problem today is that we still lack practical, scalable optical switching components (or optical transistors, if you will) that can replace their electronic counterparts in general-purpose computers. Transistors are fundamental building blocks of digital electronics, so we need an equivalent building block for digital optics. If you ever see news that such optical switching circuits have become practical and manufacturable at scale, that’s when we will gradually switch to optical computers. I can imagine a 1 kW optical computer as powerful as today’s small data center and requiring almost no cooling. I say “almost” because UPSes will still generate heat.
For the more distant future, I think a combination of optical and quantum is the way to go. In essence, the optical portion will do what electronic hardware does today with almost no need for programming to change, while quantum processors will solve specific classes of problems where they offer an advantage. I say “almost no need for change” because firmware will change drastically, but the kernel and userspace will not change as much. To use a quantum processor, or maybe a co-processor, programming will have to change drastically, but for an optical processor, it may not change as much.