# AI Safety Prerequisites Course: Basic abstract representations of computation

Th­ese are three new les­sons of our on­line course on math for­mal­iza­tions re­quired for AI safety re­search:

• Level 10: Re­cur­sive Functions

• Level 11: Set The­o­retic Recursion

• Level 12: The Equiv­alence of Differ­ent No­tions of Computability

With these les­sons, the stu­dent now should:

• Un­der­stand the ba­sic ab­stract rep­re­sen­ta­tions of com­pu­ta­tion.

• Know some of what we can ex­pect from com­put­ers and also what we can’t ex­pect from them.

• Know a lot more set the­o­retic tools like equiv­alence re­la­tions and or­der­ings.

• Have seen the con­struc­tion of the nat­u­ral num­bers from the per­spec­tive of set the­ory.

• Know about math­e­mat­i­cal in­duc­tion, and have used it!

• Know about re­cur­sion, and have used it!

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.

• 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 :)