Short report of lightweight AI research conducted in 2025

December 1st, 2025 - version: 1.2

After some months of light-focus research on what people call "AI" today I came to a conclusion (current, temporary) that what we're talking about is really JUST a new programming language, one that doesn't have strict rules but is rather based on self-learned "natural language"

Characteristic of the shell language is that you give commands to a command prompt and engage very specialized tools to perform various tasks, making them cooperate (via pipelines for example).

The shell is already pretty "colloquial", today's "AI" is just MORE colloquial, in that you describe what you want in a natural language of your choice, and then hope that the machine understands what you meant.

PRO: it might really understand you and do what you asked for, with very very low friction

CON: it might be so dumb that you might feel you want to throw the machine against the wall

From what I saw the chances you have to do with a "smart" one of these "interpreters" depend mostly on:

1. The "knowledge" or "skill" level of the "artificial brain" you'll be talking with
2. The power of the machine that makes that "artificial brain" actually talk to you

Comments are welcome via ActivityPub: https://floss.social/@strk/115645373186841641

References

These are the resources I've found on my road:

• Ollama - an inference tool, lets you download Large Language Models (LLMs) locally and use them
• Opencode - a coding agent tool, let's you connect to Ollama (or other services) and interct with them to automate tasks on a code base ("please change the code so it does this and that, commit, push").
• ##llm-bots - an IRC channel on libera.chat network where humans put AI agents to test and use as a community (web access)

Credits

These are the people who helped me conduct the research, in chronological order:

• Lorenzo inspired me to look at LLMs, telling me how they gave him back the joy of programming.
• Elisabetta, my partner, borne me while I talked to machines.
• Guglielmo confirmed me you can actually be the CEO of a company with just AI "employees", or a book editor with just AI "writers".
• Laurențiu played with me and Lorenzo in the OSGeo Lounge chat room and shared his large knowledge about all topics (at some point I've been wondering if he's also an LLM!)
• pasky (in ##llm-bots IRC channel on Libera.Chat network) showed me how far the few big corporations went with LLM and inference tools allowing me to take a look from within my usual terminal.
• [someone] (in ##llm-bots) showed me how far you can go with self-hosted LLMs and inference tools even when you define yourself as someone who knows no programming languages.