People arguing about AI risk are usually arguing about different things.
Four walkthroughs that take the vocabulary apart, show where the real disagreements are, and separate what has been measured from what is argued. Every claim is tagged and sourced.
I build things with these systems every day and I wanted to understand the argument about them properly, rather than by absorbing headlines. So I did what I do with anything complicated: took it apart and drew it. This is that, in public, with the uncertainty left visible instead of smoothed out.
Start anywhere
How a language model gets made
Five stages from raw text to a deployed assistant, and which safety work attaches to which stage. The one structural idea most coverage misses: four stages change the model, one wraps around it.
walkthrough · 5 stepsWhere the worry comes in
How bad, and how it happens, are separate questions, and almost every confused argument is two people each holding one axis. Includes the number people quote at you, and why two of them can say ten percent and mean incompatible things.
interactive grid · definition builderThe words
Thirty five terms the argument cannot proceed without, defined plainly, including the speculative ones people state as fact. Test yourself first if you like.
quiz · glossaryHow to read the flags
Every substantive claim on this site carries one of three marks. They are the most important thing here, because the usual failure of writing on this subject is letting the three blur together.
On sources
Every page here ends with its sources, linked to the primary document rather than to coverage of it. Where a person is quoted, the link goes to the place they actually said it. Where the evidence points both ways, as it does on biological risk uplift, both sides are listed rather than one being picked.
None of the frameworks used here are mine. The two axis grid comes from Bostrom and Cirkovic, the causal categories from Zwetsloot and Dafoe and from the International AI Safety Report, the training pipeline from the published method papers. What I have done is arrange them and draw them. Each page says which parts it borrowed and from whom.