AI and LLMs

I was reading Mitchell Hashimoto’s excellent article, Everyone Should Know SIMD, when I came across this disclaimer: I hate that I have to do this for every post now, but I also want to note this was completely hand-written with no AI assistance. It made me wonder whether a new distinction is beginning to emerge. Imagine Mitchell writes the article himself, then asks a generative model to produce Zig examples illustrating the concepts. Suppose the model produces exactly the same examples he would have written. Has the quality of the article declined? ...

When working with LLMs, the prompt is only part of the problem. The harder part is deciding which context should travel with the request: source files, command output, constraints, previous decisions, and the exact shape of the failure. Most writing about LLMs asks how they make our work more efficient. The more interesting question here is narrower: how do we work more efficiently with the models themselves? In my work, that often means leveraging multiple tools, each with a focused role, rather than treating the LLM as a single chat window that has to hold everything at once. ...

After a small change in my son’s schedule, I wanted to keep everything consistent while making it easy to share. That kicked off a small calendar-planning tool with a few requirements: a clean, printable calendar view, no paid app or subscription for a one-off need, and a way to subscribe or export the calendar so it stays visible across all of my devices. Background I first looked at Apple Calendar, but its printing options are limited and a presentable, readable calendar was non-negotiable. Fantastical would cover everything I needed, yet I was not keen to buy a license for a one-off need. I also considered using something like InDesign with scripting to build a polished calendar, but that felt like overkill and would have required learning additional tooling and writing a fairly complex calendar builder. Google Calendar had similar printing limitations. Given all of that, building a small software package felt like the right choice: I could script the visuals and color coding, export to ICS, and control schedule changes via versioned JSON files. ...

Background For a computer, an image is collection of interpretable instructions that amount to a visual representation. Raster images are composed from using pixels, containing unique colours, whereas vector images keep track of points and equations that join them. In \(\LaTeX\) PGF/TikZ is used to generate vector graphics from algebraic descriptions. TikZ is mostly used to conveniently draw various scientific figures. ChatGPT is capable of generating computer code in majority of popular languages. I wanted to test how far I can get with ChatGPT generating illustrations using TikZ. ...