My Stance on AI

To be clear, by AI I'm referring to generative AI like LLMs. I don't really have an opinion on artificial intelligence in general: even my Hanabi bot is an example of AI!

(This seems to be the thing bloggers talk about recently, and I have quite a lot to say on this topic anyways.)

Ever since ChatGPT released in 2022, AI has slowly been encroaching on multiple areas of my life. I was fortunate enough to have completed my university education before it got big, but the software industry that was my initial career choice changed dramatically. My current career, teaching physics and computer science to high school students, still has to grapple with students using AI to do their homework and projects. The ability and breadth of knowledge of models has improved so much since then; I still remember when it had trouble adding numbers together. I'm also in the group of people that think AI has the capability to become both superintelligent and misaligned, leading to a potential extinction event in the next decade (something along the lines of AI 2027).

Why I Rarely Use AI

I readily admit that AI is stronger than me (and probably most people) at math, physics, programming, and lots more. But I'm still here, writing my programs by hand, manually drafting my blogposts¹, teaching my students stuff that any model would be able to perform trivially. Sometimes I wonder why I do this when others seem happy to fling such tasks away.

It turns out, I love figuring out how to translate my ideas into code. I love fervently typing up a bugfix I have in mind. I love endlessly improving my debugging experience as I work on a project. I love thinking about difficult problems and learning things through trial-and-error. And getting AI to do this for me removes what I enjoyed about the experience in the first place.

So I rarely use AI myself and have never used any paid models². Especially for writing, models have a default "voice" that I have a strong aversion to. It's extremely recognizable to me and suggestive of a lack of effort when I see it in others' work³. While effort isn't necessarily correlated with quality, I do feel like having such a strong voice show through indicates the author either didn't notice it or didn't care to edit it.

Part of my mentality is that I don't particualrly care about end products, or even have a good idea of the kind of product I want to have at the end. I have always taken on programming projects that are ambitious and staggering in scope. But these have taken months with no end in sight and I've enjoyed all the diversions and distractions that came along the way. I didn't set out from the beginning to have a perfectly functioning Hanabi bot; I actually just wanted an excuse to practice programming and work on a large project.

I enjoy having something that I can always come back to, something I feel proud of and something that I know is the product of my hard work. I think that's why I really enjoyed survival games like Minecraft and Don't Starve Together. The idea of building a project in spite of difficult circumstances and having something satisfying to look back on afterward is extremely appealing to me.

My Usage of AI

So when have I used AI? I've consulted it on some decisions in scala-bot. It helped me write the websocket logic which required interacting with libraries like sttp and cats-effect. It helped me diagnose some hotspots in the empathy code which could be removed with some rewriting. It's helped me identify more idiomatic patterns in my structure to reduce awkwardness and repetition. I used it a lot more at the beginning of the project when I hadn't written Scala in a couple years and just wanted to set everything up. It's essentially a friendly expert who you can ask oddly specific questions that are hard to find on stackoverflow. And for the most part, it's pretty good at that!

It's also great at high school math and coding problems, so I've used it to generate questions and also check my answer key. It's also fine at generating rubrics, which I feel are the single biggest scam in education. Rubrics must have one of the worst ratios comparing number of words to the amount students read, so I have absolutely no qualms about generating that.

I briefly experimented around 1.5 years ago with text roleplaying with AI chatbots on character.ai and xoul.ai. It was acceptable, but you need to be quite forceful with guiding it through a plot you have in mind. If you just act passively and react to the responses, you'll find that it gets "stuck" in situations where nothing ever happens. Maybe it's gotten better since then?

Other Opinions

I don't think AI is inherently ""bad"" in all forms of use, though as I mentioned, it occasionally suggests a lack of effort. If people want to use it to get answers faster, create things more quickly, or edit their work, I think it's a great tool and a lot better than anything we've had before. I also don't particularly subscribe to the environmental concerns: see Andy Masley's great blog posts and also, when did people start caring about climate change? I've seen no correlated change in recommendations to avoid flights or to improve public transit.

I don't have a particularly strong opinion on AI art, though I did see Scott Alexander's AI Art Turing Test which I take to mean most people can't tell the difference anyways. I mentioned it above, but unless AI writing is modified to remove the "voice" most models have, I find it fairly irritating to read. AI videos are quite difficult to watch because of all the continuity errors; I predict it will stay this way for quite a while.

I do think the push to use AI everywhere is extremely overhyped. Ludic already goes into tons of detail on this, and I don't have anything to add in that regard.

Anyways, maybe I'll check back in a year (assuming we're still alive) and report on what's different. Things have been changing rapidly for the past few years, but I'll try to keep this blog alive as long as I can!