[Question] Is it harder to become a MIRI mathematician in 2019 compared to in 2013?

Below, I make the dis­tinc­tion be­tween “MIRI math­e­mat­i­cian” and “MIRI en­g­ineer”. (In my mind I tend to think of these as “re­searchers” and “en­g­ineers”, re­spec­tively, but I think MIRI calls both of these classes of peo­ple “re­searchers” these days so I want to avoid us­ing “re­searcher”.) Ba­si­cally I count any­one who has pub­lished a pa­per or post in agent foun­da­tions as a math­e­mat­i­cian, and ev­ery­one else as an en­g­ineer. From the cur­rent team page, I would clas­sify Nate, Eliezer, Benya, Scott, Sam, Abram, and Tsvi as “MIRI math­e­mat­i­cian”, and Jesse, Nick, Buck, Ben, James, and Ed­ward as “MIRI en­g­ineer”. I don’t ac­tu­ally know if this is a rea­son­able clas­sifi­ca­tion given that MIRI’s re­cent work isn’t pub­lic.

As far as I can tell, MIRI has not made any new math­e­mat­i­cian hires since mid-2017; see this table which I made, and this blog post which I be­lieve is the last hiring up­date for math­e­mat­i­ci­ans.

As­sum­ing there are no unan­nounced hires, the lack of hires can be for two broad rea­sons:

  • A change in the de­mand: less need for new peo­ple, more se­lec­tive about who to hire, etc.

  • A change in the sup­ply: fewer peo­ple to pick from, drop in qual­ity of peo­ple that come into con­tact, etc.

I think both changes have prob­a­bly hap­pened, al­though my guess would be that with the take­off of the AI safety field, there are now more peo­ple to pick from, so the lack of hires is mostly due to a change in the de­mand.

One way I’ve been think­ing about these changes is to con­sider the same per­son at­tempt­ing to work at MIRI as a math­e­mat­i­cian in 2013 vs 2019. In par­tic­u­lar, I’ve been think­ing about Nate Soares, who was hired as a re­search fel­low in 2014. To give my own sum­mary, Nate was a full-time Google en­g­ineer who had taken multi-vari­able calcu­lus and real anal­y­sis in col­lege (in ad­di­tion to the math­e­mat­ics ap­pear­ing in his CS de­gree), but oth­er­wise with no ex­pe­rience learn­ing or do­ing re­search in math. He dis­cov­ered LessWrong prob­a­bly in early or mid 2013, and came around to MIRI’s wor­ld­views by July 2013. He then spent five months (Au­gust to De­cem­ber) learn­ing ba­sic cat­e­gory the­ory, set the­ory, model the­ory, com­putabil­ity and logic, and prov­abil­ity logic, then at­tended a MIRI work­shop, stud­ied some more math, and joined MIRI as a re­search as­so­ci­ate (Jan­uary 2014) and then as a re­search fel­low (April 2014). Nate might be an out­lier in his study­ing/​re­search abil­ities, so might be a bad ex­am­ple to use here, but he is the most trans­par­ent ex­am­ple, hav­ing blogged about his ex­pe­riences. See The me­chan­ics of my re­cent pro­duc­tivity, On sav­ing the world, and his other LessWrong posts for more in­for­ma­tion.

I am won­der­ing what the cor­re­spond­ing tra­jec­tory would look like for a Nate Soares who dis­cov­ered LessWrong in 2019. (This ver­sion of Nate would be six years younger. Also dis­cov­er­ing LessWrong later means a lower qual­ity per­son in ex­pec­ta­tion, even ad­just­ing for age, so we can imag­ine some­thing mag­i­cally pre­vented him from dis­cov­er­ing it.) Would he have been hired? If so, how long would it have taken him? If not, what would have hap­pened to him?

I’ve brain­stormed some po­ten­tial differ­ences be­tween 2013 and 2019. Note that I’m not say­ing these are the only rea­sons or that I think any par­tic­u­lar rea­son listed is likely (I’m ask­ing this ques­tion be­cause I’m not sure what’s ac­tu­ally the case).

  • Take­off of the field in gen­eral (see Timeline of AI safety, Timeline of OpenAI, Timeline of MIRI, Timeline of FHI, and Timeline of CHAI for some rele­vant dates)

    • Now there are more AI safety or­ga­ni­za­tions, who could be com­pet­ing for new tal­ent. In other words, a 2019 ver­sion of Nate could be hired by OpenAI in­stead of MIRI.

    • More peo­ple in­ter­ested in en­ter­ing the field. I think this would mean more hires, so doesn’t ex­plain the lack of hires.

    • Po­si­tion of MIRI rel­a­tive to the AI safety field. In 2013 MIRI rep­re­sented the AI safety field as a whole, but in 2019 MIRI rep­re­sents a par­tic­u­lar view within AI safety. There are now more filters for a po­ten­tial hire, where they must not only agree with AI safety as a cause area, but also with MIRI’s spe­cific po­si­tions within that cause area.

  • Less open­ness at MIRI (e.g. nondis­closed-by-de­fault policy, fewer MIRI peo­ple who are ac­tive on pub­lic fo­rums). A 2019 ver­sion of Nate would have a harder time learn­ing about and con­tribut­ing to the cut­ting edge of MIRI re­search, and in­ter­act­ing with MIRI re­searchers.

  • Emer­gence of MIRI’s new re­search di­rec­tions. This might mean MIRI is more fo­cused on hiring en­g­ineers, so a 2019 ver­sion of Nate might have be­come an en­g­ineer at MIRI in­stead of a math­e­mat­i­cian.

  • Lack of MIRI work­shops (I’ve heard the grow­ing size of MIRI’s re­search team as a rea­son for this, i.e. work­shops are less nec­es­sary be­cause there are more peo­ple already at MIRI to bounce ideas off of). A 2019 ver­sion of Nate wouldn’t be able to at­tend a work­shop, un­like the 2013 ver­sion.

  • Emer­gence of pro­grams like MSFP, in­tern­ships, AIRCS, and re­train­ing pro­ject. MSFP in par­tic­u­lar seems to serve a similar re­cruit­ing func­tion for math­e­mat­i­ci­ans as work­shops, but I’m not clear on the differ­ences from a hiring view­point.

  • Big­ger re­search team at MIRI. Maybe the “ba­sic re­search po­si­tions” at MIRI have been filled, so MIRI can be more se­lec­tive about the ex­per­tise new hires have, or their level of ca­pa­bil­ity. Another idea is that MIRI re­search is not very par­alleliz­able, so the value of the marginal hire goes down quickly as a func­tion of num­ber of re­searchers.

  • Fund­ing. MIRI seems to have way more fund­ing now, so this doesn’t ex­plain the lack of hires.

I am in­ter­ested in thoughts on the spe­cific ques­tions I’ve raised, thoughts on which of the rea­sons brain­stormed above are most likely, and any other thoughts peo­ple have on MIRI hiring prac­tices in 2013 vs 2019.

Thanks to Matthew Bar­nett, Louis Franc­ini, and Vipul Naik for feed­back on this ques­tion.