Aliens: Explaining LLMs and AI agents to my children

archive.org

2 points·onetokenatatime··3 comments

I wrote a little aliens story in an attempt to explain to my children how agentic AI works under the hood. You too may find it helpful if you would like a simple mental model of how concepts such as LLMs, harnesses, SFT, RLVR, RLHF, attention, sub-agents, thinking tokens, prompt injection, and context compaction fit together.

LLMs are vectorial databases with losses. They use the characters(bits) in the input question to index the query, and are quite bad at retaining the context of such query. This is why companies have named the split and shortening of the length of those inputs as "agents".

They named it like this due they follow a marketing that personify the names of the methods, to make people believe they are dealing with a different technology, including the named 'thinking' methods that simply query twice to increase the index of the next query, and so on.

I think introducing aliens and robots will make children believe it is magic, like if intelligent entities or a thinking were behind the answers/procedures, and that such "answers" are smarter than them or even their parents, when in reality, it is a marketing personified model joining statistical characters(bits) from the text with created such a vectorial database with loses, and guaranteed errors presence, with joins like a pachinko machine.

Fair! I really hope kids don't think any of what's happening is magic. I tried to design against it: the alien is just a next-token predictor, has no memory, and can't do anything without the robot. It's painstakingly trained by the scientists. So I was hoping to frame it as a multi-part scientific endeavour rather than magic.

Reply on news.ycombinator.com