What's a Model?
◀ Prev | 2025-09-01, access: Public
alignment basics theory text Gemma hallucination For a change, this talk doesn't focus on a specific paper, but tries to address some basic concepts needed for understanding other material in the lecture series.
In particular, what actually is meant when we talk about a model? I start out by discussing linear regression, which is much simpler than most of the machine learning models people talk about today but illustrates the concepts - a model of a model, if you will. It's even possible to address such things as "memorization" and "hallucination" in the context of linear regression.
Then I talk about the training process used for building much larger models, in particular, the ones that go into chatbots. I cover a bit of how user input goes into a "template" to feed the underlying model of an online chatbot, and how the model can also interact with "tools" to solve problems that the model alone could not do so well.
This talk is my chance to comment a bit on misconceptions about AI, that I wish would be less widespread, and I include some discussion of how heavily "aligned" models can end up morally retarded. This is a problem we really need to address, and the right way to address it is with less so-called safety, not more.
I conclude, instead of my usual research ideas, with some "homework" instead. Those who attempt the homework might like to comment on their results here in the discussion thread system. I've made this video visible at the "Public" level (meaning you don't need an account to watch it), but posting in the comment threads requires creating a free account.
The link for llama.cpp should be live in the PDF file of the slides, but you can also click on https://github.com/ggml-org/llama.cpp; and the story I mentioned that gave Gemma such fits was The Number 13 Road - which, note well, was written over 20 years ago and not originally intended to be a moral puzzle, at least not in the first few paragraphs. It didn't occur to me at the time that one day I'd be arguing with an AI about who did or didn't do the right thing in that story.
At risk of letting things get overheated, I'm making the comment threads for this video a "designated area" where you may talk about environmental and job market impacts, because that's actually relevant to the topic of the video. All other discussion forum rules still apply.
Access level to read comments: Free account (logged in)
Access level to write comments: Free account (logged in)