Some ideas about LLMs
Large language models have been around for quite some time.
I've been interested in this technology since I saw Eliza at a museum when I was a teenager. I'd been reading science fiction for a long time, still do, and the idea of a communicative computer system was really cool. Still is.
In the beginning of the current century, after moving to Germany, I took some linguistics courses at the University of Marburg. One of my professor's specialties was machine/human interface.
When Siri, Alexa, and their gang came out, (I think it's better for everyone concerned if we let Clippy rest in peace) I enthusiastically jumped on the bandwagon. So, I could turn lights on and off, ask for the weather or sports scores, and perform a few other useful but not altogether scintillating tasks. I still ask Alexa to play music for me from time to time, and turn on the plug for my media center. But the novelty wore off a long time ago.
Now, OpenAI's ChatGPT, Claude, Bard/Gemini and Microsoft's Copilot burst on the scene.
Once again, this sci-fi Ur-nerd jumped in with both feet.
And the time-honored "your mileage may vary" caveat reigns supreme.
I've reawakened an interest in programming (Python) and that's certainly a positive aspect. I've learned that these LLMs are better at helping people code than they are at most other tasks. Because they're computer systems, natürlich.
However, the hype that they are currently marinating in seems to me to be cooked up by an economic requirement, not a functional requirement. If several of the largest companies in the world dump tens of billions of dollars into a service, they are fully capable of generating enough interest in that service to lure enterprises and individuals to pay for said very expensive service.
I'm OK with that, it's how late-stage capitalism works.
So unless you want a computer to create a meme for you (no judgment) you're better off using your native intelligence to interface with the world. Or, use it to edit your code, it's really good at that.
As of July 2026 my current system includes Anthropic's Claude Code, Ollama for local models served to OpenWeb UI, Gemini Notebook, Docker, GitHub, Supabase, in both Mac and Windows. I prefer Mac these days, the Metal API features add to the advantages Apple's silicon gives a setup. I worked with ChatGPT for about a year, and in 2025 I focused on Microsoft's Copilot. However, those systems were not fit to purpose for my use cases, and I'm very happy to have changed to the Anthropic products.
Knowing where these systems fail day to day is what post-editing and the training modules are built on.
Articles
Caveat: these are blog posts. I haven't treated them like research papers. I don't want to exhaustively sprinkle them with links to references. I'd like the reader to read first, cogitate and then deep dive into the background if and when necessary. It's an editorial decision, not a faux pas. If you don't know who Vaclav Smil is or what a subway token looked like, you're here, which leads me to the assumption that you have access to that information. I'm asking you to resist the urge to dive down the rabbit hole while you're reading if you can, I'm prone to it myself, but it's an attentional failing at best. Thanks for considering it.
