Why do I need to use this? Why can't I just use the LLMs directly to generate lectures if I want?
Just assign years worth of reading homework every week, and I am sure all the students will fulfill their assignments of reading every token that went into each video, no matter how many, rather than watch an LLM generated video.
Students simply need to learn how to read text as fast as an LLM can, and have an enormous token budget to use the LLMs directly themselves.
And while you're at it, increase their tuition by the amount those tokens cost, so each student can pay for feeding the same tokens into their own LLM instead of the school doing it once and paying out of their own pocket immediately. Because the idea of schools producing reusable courseware is unprecedented! That way they can fold the token cost into their student loans.
Sarcasm off. Here is what I actually mean, and I think it makes a better case for this than the authors made for themselves.
The prompt is not "give me a lecture on relativity." That is only the part you type. The rest is the lecture description language, the visual grammar, the worked examples, the pedagogical constraints, and the review loop: a program, not a request. Latent training gives you the physics for free. It does not give you the presentation. Asking why you need this when you have an LLM is like asking why you need a compiler and a library when you already have a CPU.
Authoring cost is paid once and amortized over every viewer. That is what a textbook is. "Just prompt it yourself" is the scriptorium: every student copies their own book, pays their own tokens, and collects their own transcription errors. A school paying once and shipping one reviewed artifact to everybody is not a limitation, it is the entire point of courseware.
Their strongest argument is the one they undersold: the source is the artifact. Every recorded lecture on the internet is a render, frozen and unpatchable, and its mistakes are permanent. If the lecture is source code, a mistake becomes a patch, a translation becomes a recompile, and improvements accumulate instead of being lost every time someone re-records from scratch. And because they render deterministically from source with graphics and TTS instead of generating video with a diffusion model, a fix is a fix and not a re-roll. That is the real correctness argument, and it never got made.
Graphics settled this exact split forty years ago. PostScript is the authoring language: loops, conditionals, a full interpreter. PDF is PostScript with the programming removed, an inert render of flat drawing calls with all the loops unrolled, and Distiller produces one from the other by running the program and recording what it draws. Language in the middle, inert artifact at the edge. A lecture wants the same shape, and "lectures as code" is that observation applied to teaching.
Other arguments available to them: a lecture with a git history is a new kind of object, because you can bisect an explanation and read the changelog of a proof. Errata stop being a sheet nobody reads and become a rebuild. Localization, transcripts, and accessibility fall out of the source instead of being bolted onto a render. One source can emit different levels, and different notation conventions for physicists and mathematicians.
The answer to davisp lives in the same place: sign the source. Feynman is trusted because his lectures are texts we can check, argue with, and cite. Publish the language, publish the sources, publish the diffs, let critics file issues against a named maintainer, and "everyone who watches after the fix gets the corrected lecture" becomes a checkable claim instead of a promise. That's the one thing they're not doing. regnull asked for the language and was told it is the company. Right now they keep the LaTeX and the PostScript and ship only the PDF -- and they're PhD students! arXiv wants their .tex, not just their PDF, because in their own field shipping only the render was never good enough.