"OpenCode Go" (a subscription) lets you use lots of hosted open-weights frontier AI models, such as GLM-5 (currently right up there in the frontier model leaderboards) for $10 per month.
I've used it to create little one-off tools that I needed for some specific tasks, without any care whether it's "of substance" or "investable." In the past, I might have Googled around to see if there was an existing app or open-source package to do the job. Quite often, the AI agent will use some existing software packages, but I didn't have to find them and figure out how to use them.
A real example from today: I got tired of the Looney Tunes way-too-colourful screensaver options on my Mac mini, so I asked Claude how to get a screensaver that is a nearly-black uniform monochrome grey. Surprisingly, that's not actually an option in macOS System Settings. So Claude just wrote a little Python script to generate the images I need for my two displays, saving them to the right place on my Mac mini. It used the Pillow package, but I didn't have to spin up a whole Python project and install Pillow; it just used PEP 723 inline script metadata to tell uv to use Python >=3.12 and to install Pillow. Then Claude gave me the uv one-liner to run the script (uv run ~/make_screensaver.py) and instructions for how to tell my Mac mini to use the generated files. The whole process took about 15 minutes from when I started writing the first prompt to the time I had my new screensaver working.
For me in Canada today, Kagi is showing nanoclaw.wrongtld as the third text link, after two different GitHub repos (why two? I didn't have time to sort that out). I clicked the thing to block the link to the site with the wrong TLD; hopefully other Kagi subscribers will do the same.
You wouldn't use an LLM to solve a big Linear Programming problem, because it would cost way more than using the Simplex Method, and you'd be worried that it might be wrong.
reply