Trains in Anime (and other random opinions)

Every LLM is a text file generator. You put a text file in, you get another text file out (eventually). Every "technology", "agent" and so on is just a layer around that. It's a whole ecosystem built around "we have a text file generator, what can you even do with that?"

Chat bots? It's easy: Your first text file A is what the user types. Then you get text file B, the response. Then the user types in text file C, and you put, A, B and C together into a text file, feed this text file ABC into the LLM, and get yet another text file.

System prompts are just hidden text files you put in front of everything else. Context or memory? That's just a hidden text file.

AIs that simulate thought? You arrange your text files, include some text about "write the next step". You get a text file back. You attach it to your existing text file and feed it into the LLM again, telling it to write the next step, and so on, until the LLM includes the magic "done" marker in its output text file.

(This is in fact a higher level of how all LLMs work: They don't generate a text file, they generate a "token", a short piece of text that's roughly one word. Then you feed all the input, plus the newly generated token, into the LLM again, and see what pops out, until the LLM decides to include the "end of response" token.)

Coding assistants? Actually just a fancy name for putting literally all your code in the text file, plus some question and some description, and seeing what comes out. The new text file may include instructions to change some code, and your "coding agent tool" may (using normal code) decide to actually implement these changes.

"Skills" and MCP and so on? Those are just more text files, usually formatted in Markdown, a format originally invented to write bitchy pro-Apple blog posts (true story); and particular ways for the client tool to load them and feed them into the next LLM call based on what the previous LLM calls generated.

Agentic AI? Sometimes the LLM generates a command line or web request or equivalent in its text file. The surrounding code then decides to execute this command line and feed the result back into your main text file. Better hope the surrounding code filters out anything you don't want to happen. Or you can write "don't generate evil command lines" in your text file, and hope that the LLM honours that request (and can tell the difference to begin with).

By changing what sort of machine generated command lines you actually chose to execute, you "give the LLM power", but kind of not really. The software between your text file and the LLM, that is what chooses which command lines do and do not get execute, and you better hope it was coded properly.

Your AI got something wrong, maybe got a simple fact wrong, or it generated an unfortunate command line that your client then executed, killing your entire company, and when you pointed that out it said sorry? It's fucking not sorry. Not in any "lying" sense but because "it", as an entity, does not exist. Instead one of thousands, maybe millions of identical servers receives one single text file containing whatever setup the AI company deemed necessary; whatever you told "the AI" (including e.g. all your source code); whatever previous calls to it generated as results; whatever you answered to that and so on. It's very well possible that it's a different physical server from word to word or (probably more likely) from response to response.

Using the power gained from scraping the entire internet constantly, scanning in all the books, using all the power and all the water, this server does a lot of matrix multiplications with some really big matrices (those were the results of all the energy use), and returns a likely result vector that says there's a 90% chance the reply should start with the word "Sorry", because that would be an appropriate continuation to this text file you sent it.

When people say LLMs are just fancy text autocompletion, you probably think that's over-simplified. And it is, but only a tiny bit: LLMs and everything built on them are autocomplete in a loop. Sometimes two loops. And also sometimes some code that takes the output of this autocomplete and acts on it in, in the hope that you'll find that useful.