I've looked a little bit at the RAISE website, and I've looked at the overview of curriculum topics, and I'm finding it a little...sparse, maybe? (I haven't actually looked at the class materials on grasple though, so maybe there's more stuff there.) I'm wondering how realistic it would be for someone to start engaging with MIRI-esque topics after learning just the courses RAISE has outlined.
At least for the prerequisites course, these are all topics covered throughout the first two years of a typical undergraduate computer science degree. And that doesn't seem like quite enough.
EX: TurnTrout's sequence of essays on their journey to become able to contribute towards MIRI-esque topics seems to span a much greater gamut of topics (linear algebra, analysis, etc.) at greater depth, closer to what one might cover in graduate school.
I guess, to operationalize, I'm curious about:
1. What target audience RAISE has in mind (technical people looking for a refresher, people who have had zero real exposure to technical subjects before, etc. etc.) for their materials.
2. What degree of competence RAISE expects people to come out of the curriculum with, either best-case or average-case.
3. In the best case, how many units of material do you think RAISE can product? In other words, is it enough for students to study RAISE's material for a 6-month long curriculum? 1 year long?
****
(Of course, it's also much easier from my position to be engaging/critiquing existing works, than to actually put in the effort to make all of this happen. I don't mean any of the above as an indictment. It's admirable and impressive that y'all have coordinated to make this happen at all!)
Hi, full time content developer at RAISE here.
The overview page you are referring to (is it this one?) contains just some examples of subjects that we are working on.
1. One of the main goals is making a complete map of what is out there regarding AI Safety, and then recursively create explanations for the concepts it contains. That could fit multiple audiences depending on how deep we are able to go. We have started doing that with IRL and IDA. We are also trying a bottom-up approach with the prerequisite course because why not.
2. Almost the same as reading papers, with clear pointers to references to quickly integrate any missing knowledge. Whether this will be achieved in the best case or in the average case is currently under testing.
3. I don’t know about the absolute amount of time required for that. Keep in mind that this remains to be confirmed, but we have recently started collecting some statistics that suggest it's going to be at least comparatively quicker to read RAISE material, compared to having to search for the right papers plus reading and understanding them. This would be the second main goal.
(Of course, it's also much easier from my position to be engaging/critiquing existing works, than to actually put in the effort to make all of this happen. I don't mean any of the above as an indictment. It's admirable and impressive that y'all have coordinated to make this happen at all!)
Thanks :)
Followup to AI Safety Prerequisites Course: Revamp and New Lessons. First post.
These are three new lessons of our online course on math formalizations required for AI safety research:
With these lessons, the student now should:
If you study using our course, please give us feedback. Leave a comment here or email us at raise@aisafety.info, or through the contact form. Do you have an idea about what prerequisites are most important for AI Safety research? Do you know an optimal way to learn them? Tell us using the same methods or collaborate with us.
Can you check if a mathematical proof is correct? Do you know how to make proofs understandable and easy to remember? Would you like to help to create the prerequisites course? If yes, consider volunteering.