LLMs

LLMs are a really, really cool technology. I use them every day, often a lot. They have a lot of issues. LLMs are pretty bad for the environment. I tend to think their issues are over-hyped, but they also can't be denied. LLMs have almost certainly stolen a lot of content. Some of it is philosophically debatable, and some of it isn't (unless you wanna get real skeptical with it).

Nevertheless, LLMs are a really, really cool technology. They speed up the generation of code and can do simple maintenance tasks that relate to text very well. They're also excellent at doing deep (open internet only) research very quickly. They are not incredibly reliable for both alignment problem reasons and general quality issues. Their work needs to be checked in a similar way as a human's would: if it's a really high risk task and you have plenty of time, you're obviously going to want to check it over, but if it's a lower risk task or you simply can't achieve what you want to do without delegating, you have to delegate. Sometimes that means failure, as in the case of delegating to a person.

My goal on this site is usually to communicate ideas. I tend to think that LLMs are doing a pretty bad job at that. The world will be a much different place when LLMs can write the way that I'm writing even here, in my opinion.

My goal on this site is also to be correct. I use LLMs to check my facts. I use LLMs to back up my facts. A lot of the facts on this site are formatted by LLMs. Why? Well, they like to write things like "CO₂e" properly (@clod pls fix). In many cases, they're giving me a research report on something that I've already done research on, and they're returning it exactly as I'd want it to be presented. But whether I write something with LLMs or not is basically aligned with my philosophy on them for this kind of task: I check over their work, and I don't do any sort of gratuitous unnecessary rewriting if I like the way it comes out.

There may be some copy that's left on this site which I didn't mean to leave as a placeholder but there it is. I'm still getting the basic shape of the site's ontology down, so some of this is the LLM trying to pretend to be me, but hopefully by around 2026-07-30 that's no longer the case.

I've also done stuff like "Go on my YouTube account and find videos to add here" or "organize these photos and put them on Felix's page".

Unlike almost every web framework I've used before, this is the first time I've found a web framework that I both really like and wouldn't know how to write manually before I made a website successfully with it (topcoat×topcoatA Rust server-side web framework from the tokio project. This site runs on 0.1.3.github.com/tokio-rs/topcoat->). I read a lot of the code, and I edit the content in neovim, but almost all of the actual code edits are vibe coded, and the whole scaffolding and deployment setup was 100% vibe coded.

I have been alternating between Anthropic's Fable 5, OpenAI's Luna, Terra, and Sol 5.6, and Cursor models (Grok 4.5 and Composer 2.5). I really like how the code's been coming out.