I applied for a MIRI job in 2020. Here’s what happened next.
I recently saw Rob Bensinger asking what MIRI could do to improve hiring. I tried to get a job at MIRI in February 2020 and I have some feedback to give. I am sorry, because I am about to say some harsh things, but I think they might be important for you to hear, and my previous more low-key attempts to point out problems had little to no effect.
tl,dr: MIRI promised to contact me once the pandemic no longer prevents them from hiring people. They never did.
What happened
I applied for a Software Engineer role at MIRI on February 26th. I hoped to end up as a researcher eventually, but this position seemed like the easiest way in for a foreigner. The first stage was a multiple choice test from Triplebyte. I took it on March 2nd. After a week, on the 9th of March (why does it take a week to grade a multiple choice quiz?!), I got the following letter from Triplebyte:
Hi Viktoriya,
Triplebyte partners with Machine Intelligence Research Institute to help candidates that go through their process find the job that’s right for them. If the opportunity with Machine Intelligence Research Institute is not the right fit, we believe we can help you find a great fit somewhere else.
Seeing as you did particularly well on Machine Intelligence Research Institute’s assessment, we are going to fast-track you to the final stage of our application process! All that is left is our technical interview. If you pass, you will be matched with top tech companies and fast-tracked to final round interviews. [...]
This gave me the impression that I passed the quiz, so I was surprised to get a rejection letter from Buck Shlegeris on the same day. The letter left me with a way forward: Buck sent me a list of ways to impress him and get back into the interview. The tasks looked about a week of work each. I wondered if doing one is worth the opportunity cost.
I lived in Russia at the time. Russia was fighting two wars, in Syria and Ukraine. I think it’s fair to estimate that all marginal tax payments were spent on war. It didn’t cost much to hire an additional soldier to fight in Ukraine. There’s a word “15-тысячники” for a mercenary hired to fight for DNR for 15000 rubles a month. I easily paid that sum monthly in taxes. Staying in Russia was costing human lives. Once I stopped trying to overthrow the government the normal way, there was no reason to stay. Wasn’t it better to spend the time applying to other US companies then, to increase the chances of landing a job before the H-1B application deadline? I thought about it and decided to go for MIRI. I didn’t want to spend time on companies that weren’t doing alignment research. I was reasonably confident I’d pass the interview if I was able to get in. (I think this was justified. I interviewed for a FAANG company last year and passed.)
So I solved a task and sent it to Buck while reminding about the H-1B application deadline. He said he’d get back to me about further steps. The H-1B application deadline passed. There was no reply. I later discovered that MIRI can actually apply for H-1B all year round. Well, at least that was true 6 years ago. Buck never told me about this. If I knew, I could have applied for another job. The ML living library position was still open (at least that’s what the job page said), and I’d being working in ML for 2 years.
Two months later I got a letter from Buck. He said that MIRI applications process is messed up by COVID, and that he put me on the list of people to contact when things are starting up again.
I asked Buck if MIRI does green card sponsorship and he said he isn’t sure how all of that works. I asked who I should contact to find out, and got no reply. This is weird, how can an interviewer not know these things? It was very important to me to know whether I’d have a path to a green card. An H-1B visa lasts at most 6 years. Given the things I said about Ukraine online, I’d surely go to prison if I ever returned to my home country. Having a green card means being able to leave your employer to do independent research, or change a job more easily without making the new employer go through another visa application process. It’s a major job perk, so maybe you should figure out if you can do this?
I tried to get in touch with Buck before the new H-1B deadline in 2021. I couldn’t, so I posted on lesswrong (I wrote under the nickname “seed” when I lived in Russia). Buck replied to give me some general advice on getting an AI safety job. Anna Salamon offered to talk to me about AI safety jobs. I sent her a direct message and an email to anna@rationality.org. I got no reply. At this point, I no longer felt that I could rely on rationalists’ help to get out of the country. I found a regular ML job in Germany. Shortly after, I won the diversity lottery. I spent the last year working in Germany, moved to the US last month and got into the SERI MATS program.
Rob Bensinger told me 3 months ago that “for all practical means and purposes, pandemic is already over”. However, I never got contacted by Buck. Maybe you no longer consider it important to keep the promise, since MIRI no longer needs programmers? But it would me important for me if someone contacted me to tell me this, because then I’d know that I can trust MIRI at least a little bit. I have since seen evidence that some MIRI employees are commited to being honest even when it’s inconvenient. I feel a bit more trusting of MIRI now.
However, I cannot believe what MIRI told me to believe: that you wanted to hire programmers, but couldn’t do so for years because of the pandemic, even though you managed to hire at least 3 other people despite the pandemic. Even though lots of IT companies not in the business of saving the world from imminent doom worked through the pandemic. It doesn’t make any sense to me. Here is my best hypothesis of what happened. By the end of February 2020, MIRI did not need more programmers. What you needed was people with new research ideas. Maybe you just filled the Software Engineer position and hadn’t yet taken down the vacancy. That would be consistent with my observation that the job information on MIRI’s website only seems to get updated after I nag you about it. That’s why I was rejected after passing the test. The standard rejection letter happened to include a suggestion to do a week-long task to impress Buck. After I completed it, it became awkward to reject me anyway. Fortunately, COVID was there to take the blame.
Anna probably didn’t reply because she was busy and I didn’t ping her enough times, because my intuition on how many followups are acceptable was just wrong. Sending fewer than 3 is not even trying. I am always able to reach people now that I adopted 3 followups as a rule.
Am I right?
Against metahonesty
When I am trying to figure out how much to trust MIRI employees, I feel really creeped out by the fact that you developed a whole set of norms for deceiving people. Yes, in theory, one was only supposed to lie for really good reasons, such as saving Jews from Holocaust or, erm, covering up that you broke a murderess out of prison based on hearsay. In practice I saw it invoked to lie for PR purposes. While I feel somewhat reassured by Nate Soares saying that he decided to be more honest, I don’t think it’s generally ok when an empoyee cannot just trust her employer, and is instead expected to ask about his honesty protocols.
Metahonesty culture is incompatible with the Charity culture.
In our society, the prevailing norm—which I just called Charity culture—is that we should assume people innocent until proven guilty. Questioning someone’s integrity is offensive, you don’t do that until you have solid proof. By metahonesty norms, if you suspect someone of lying, you should ask about their honesty protocols. If you don’t, the loss of trust is considered partially your fault, because you could have just asked. Can you actually ask the question without making your interlocutor feel suspected and interrogated, though? People in both Eliezer’s thought experiments didn’t manage to. It’s tricky, especially when your conversation partner actually has something to hide.
Metahonesty norms ignore status differentials
You may notice that in both Eliezer’s examples, it is the higher status person questioning the lower-status one. It is harder to imagine a citizen questioning a Nazi officer than vice versa. Asking about honesty protocols is potentially offensive, and offending a powerful person is dangerous. It is often high-status people who have honesty protocols in this community. Meanwhile, people hesitate to even ask regular questions, and Eliezer has to be careful to ask other employees’ opinion before offering his own lest he biases everyone. What are the chances that an employee will question her boss then? I wager a guess that this never actually happens. I never saw anyone ask about honesty protocols on this forum. I worry that in practice, these metahonesty norms amount to community leaders giving themselves permission to lie to us.
Metahonesty is not scalable
The more community grows, the harder it will be to keep track of everyone’s honesty protocols. It becomes harder for new people to join. And imagine if people actually started asking metahonesty questions all the time, you’d have to deal with hundreds of emails after every public announcement!
Metahonesty fails to preserve trust
Once you see someone lie, you ought to start questioning every claim they made, including their claim of being metahonest. You need other evidence than just their word that they’re metahonest and not a simple liar. It could be a long history of interactions (doesn’t scale!) where you questioned their honesty policies (which I never saw happen) and it always turned out they were metahonest.
I am not insisting on some form of extreme pedantry here. I don’t consider it deception to say “Good morning” when the morning is actually shitty. I also don’t care who hides in your attic. I am just making a modest proposal that we shouldn’t lie to each other about our work and research.
Outside impressions
Suppose the genius miracle person we all hope for decides to join alignment research today. What can they learn about MIRI online?
MIRI isn’t publishing research. One proxy a person can use to judge whether a company has its act together is to look at the website. Sometimes things on your website are years out of date. When I see those, I lose confidence that anything is up to date. Kudos for updating recently. Though the Type Theorist job ad is still confusing. It says at the bottom that you are not actively seeking to fill this role at this time. Why keep the job ad then? Do you think you may want to hire Type Theorists in the future? Do you still want them to get in touch? Is it just there as a historical artefact? (Actually, maybe simply taking it down is a bad idea. A comment says that you can do year-round H-1B applications, and this is the only place where a job seeker can learn this fact.) Lots of info never makes it to the careers page. E.g. when MIRI was looking for Haskell programmers, you could learn about it from twitter or reddit, but not from the careers page. Now MIRI wants to hire competent executives who can run their own large alignment programs mostly-independently, but you can only learn that by reading lesswrong comments. I wondered at some point if you are trying to keep out outsiders. Have you considered hiring someone to maintain the website?
What can one learn about MIRI workplace atmosphere from online sources?
There is my story.
There is a story of a job applicant who felt “extremely awkward” at the interview.
There is a story of a researcher who felt afraid that her employers would assasinate her.
Zvi vouches for MIRI researchers based on personal experience with them.
There are 2 positive reviews at the Glassdoor page.
There are many negative impressions, but there can be a self-selection bias here. I think it could help if people shared more stories.
Personal note
Another thing I wish I knew earlier, is that lots of people are actually willing to help newcomers enter the field. JJ Hepburn from AI safety support, David Manheim, Ryan Kidd and other people were willing to talk to me and help me figure out how to get started. For the first time in a very long time, I feel like maybe I am not alone in this. I just wanted to thank you for making me feel this way.
Related:
Seems like the communication with MIRI can be a frustrating experience.
What is the standard solution to this kind of problem? You need some kind of database of ongoing tasks. When something happens, for example Viktoria applies for a job, you create a new ticket. (Ticket type = job application. Title = “Viktoria applies for Software Engineer role”. Created = 2020-02-26. Assigned to = Buck. Status = evaluating the applicant.)
Then you create some queries, for example “Open job applications”. There should be a person responsible for checking this query once in a while. If you move Buck to a different role, you check his open tickets and assign them to someone else. There could be some alarm queries such as “Open tickets where no update was made during the last month”. There should be a person responsible for checking this, too.
Dear people in MIRI, do you have anything resembling this? Could you please show Viktoria her ticket?
*
The parts about metahonesty etc. seem to me like red hering. I suspect that most people in MIRI simply never had a job outside of academia or a non-profit, and as a result they are quite unfamiliar with how to do things efficiently and reliably. Incompetence, not malice.
*
The most primitive version of a ticket system would probably be a combination of a shared spreadsheet and a wiki. The wiki would have one page per ticket, the spreadsheet would contain the metadata. For example, the spreadsheet would have tabs “Job Applications (open)” and “Job Applications (closed)”; the first one would have columns like “Created”, “Last Update”, “Applicant Name”, “Role”, “Assigned To”, “Current Status”, “Wiki Page”, and the corresponding wiki page would be called like “JOB-001″ and would contain the entire history of what happened.
We did when I was there, and I wouldn’t be surprised if it was still in use. (It was in Asana, iirc.) I don’t think there was a meta-system of “check whether or not people are falling thru the cracks”. When I left and stopped handling one of the inboxes, it got handed to someone else, but I don’t think we ever had targets for things like ‘response time’ or ‘fraction of applicants responded to’ or so on that we checked how well people were doing on. A bunch of people who we didn’t have form responses to, or weren’t quite sure how to evaluate, I think spent a lot of time waiting for replies (or never got them).
[FWIW my sense is that Buck was better than I was at communicating with candidates, and I think the “rejection but here’s how to impress me” is a better response for everyone than “hmm I’m not quite sure what to do with this person, I’ll come back to this later”.]
One problem is that I don’t think keeping the public-facing messaging up-to-date was a priority for anyone, or a lot of things were in ‘maintenance mode’ and only barely getting maintained. Like, we had a MIRIx program, where we would pay for meals/snacks for people to talk about MIRI research; I don’t remember who started it, but I took over the ‘handle requested changes’ function (like if an organizer emailed us to change their email, or get taken off the list, or to get their reimbursement or w/e). But the page looks the same as it did when I left, and I’m pretty sure a lot of those are out of date (I’d be surprised if Tom Everitt is still running events in Canberra, given that he lives in the UK now, for example). It’d probably be good to update the page to reflect the actual level of interest from current-MIRI, but I’m not sure who would think that’s a high-priority thing for them to do (and so probably it wouldn’t seem good taking opportunity cost into account).
Why not just use JIRA or a free equivalent?
If you have the capacity to set up JIRA/whatever and maintain it, yeah, it will work.
But it will not work out of the box. You need to set up the ticket types, define workflows, etc. Later redefine them, if something changes or you find out that the original setup did not work well. Setting up JIRA is a bit complicated and requires some technical skills. You may find out that one member of your team now became a part-time JIRA administrator. For a small organization this may be an unacceptable overhead. On the other hand, incorrectly set up system (workflows that do not correspond to reality) will be a pain to work with.
That is why I offered a poor man’s version that can be set up in an afternoon.
A few quick impressions:
That sucks you got your hopes up but nobody got back to you.
Regarding getting back to people, I agree it’s sometimes pretty disruptive when someone says they will and then don’t. At the same time, I don’t think that most of civilization has still figured out how to reliably respond to all the emails that they are supposed to, even important work ones? Most competent and reliable people I know are not on top of their email (and even now just thinking about this I opened my email and replied to an email I’m two weeks late on). I don’t think it’s as costly or unethical as breaking a promise.
“At this point, I no longer felt that I could rely on rationalists’ help to get out of the country.” ← Alas, but not that I disagree.
I have some sense reading this that you put MIRI in the same bucket as a MANGA companies (my new name for ‘FAANG’), in terms of what to expect operationally. I think most people don’t have a good bucket for how much smaller organizations work, and that leads you to say things like “why does it take a week to grade a multiple choice quiz?” and why they might still have old job pages up. In small orgs, one person is responsible for everything that in larger orgs many people do, and so many thinks just don’t get done on the same time scales. (I think that’s fine.)
I guess I’ll go read some of the linked threads on metahonesty. [Done.] Seems fine to be creeped about the notion of metahonesty as an emotional reaction, though I personally think the semi-autistic desire to reason explicitly about norms and when to be dishonest is valuable and worthwhile, though I agree it can be unsettling and sometimes lead someone to unwise and unethical choices.
Actually, I spent a while confused about how this related to the first part of the post (“What happened”). My guess you think that Buck lied to you and that rationalists would say “that’s fine as long as he would’ve also openly explained he would lie in that situation”, and that’s what you’re arguing against. I think that’s not what many rationalists would say, they’d say “that sucks that he didn’t get back to you. it sounds like the MIRI hiring process is fairly inconsistent in its work”. I think they’d agree that this was an instance of someone not modeling-themselves very accurately, but also that it’s not a surprising level of bad self-modeling, which itself makes it less costly.
To me, the metahonesty thing feels “related in topic” but not “upstream” of being unreliable in getting back to potential hires. Perhaps you disagree.
By the way, if you end up as part of the scholars program in SERI MATS, then I am currently the person putting together office space for you in Berkeley at the Lightcone Offices. I’ll be making sure the office can support you and the other scholars do your work, so maybe see you then :)
Added: This whole thread has helped remind me to get back on top of my emails and try another round of inbox zero. Thanks for that :D
That’s a pretty amusing comparison because MANGA companies are also notorious for (despite all the supposed professionalism, tooling, high volume, and fulltime specialists involved in one of the most important & limiting-step things they do) hiring processes which dick around people, tell them wrong outdated things about immigrations or what they would be working on, require many rounds, and wind up wasting their time and just leave them hanging for months or ghosting them in the end. Lots of bellyaching on HN & elsewhere about this. (Should this be taken as a defense or a criticism of MIRI’s ops...?)
I don’t believe that this is explained by MIRI just forgetting, because I brought attention to myself in February 2021. The Software Engineer job ad was unchanged the whole time, after my post they updated it to say that the hiring is slowed down by COVID. (Sometime later, it was changed to say to send a letter to Buck, and he will get back to you after the pandemic.) Slowed down… by a year? If your hiring takes a year, you are not hiring. MIRI’s explanation is that they couldn’t hire me for a year because of COVID, and I don’t understand how could that be? Maybe some people get sick, or you need time to switch to remote working, but I don’t see how does this delays you more than a couple of months. Maybe they don’t give visas during COVID, then why not just say that. And they hired 3 other people in the meanwhile, proving they were capable of hiring.
I formed a different theory in spring 2020: COVID explains at most 2 months of this, it is mostly an excuse. MIRI just does not need programmers, what they want is people with new ideas. My theory predicted that they will not resume hiring programmers once the pandemic is over, and that they will never get back to me. MIRI’s explanation predicted the opposite. Then all my predictions came true. This is why I have trouble believing what MIRI told me.
And this is why I started wondering if I can trust them. It seemed relevant that MIRI has mislead people for PR reasons before. Metahonesty was used as a reason why an employee should’ve trusted them anyway. I explained in the post why I think that couldn’t work. The relevance to hiring is that having such a norm in place reduces my trust. I wouldn’t be offended if someone lied to a Nazi officer, or, for that matter, slashed their tires. But California isn’t actually occupied by Nazis, and if I heard that a group of researchers in California had tire-slashing policies, I’d feel alarmed.
I agree that it is hard to stay on top of all emails. But if the system of getting back to candidates is unreliable, it’s better to reject a candidate you can’t hire this month. If I’m rejected, I can reapply half a year later. If I’m told to wait for them, and I reapply anyway, the implication is that either I can’t follow instructions, or I think the company is untrustworthy or incompetent (and then why am I applying?). That could keep a candidate from reapplying forever.
oooh
This makes me wonder if operations really is the big bottleneck with EA/AIS/MIRI in particular. Anyone good at operations, would not have let this situation happen. Either it would’ve ended quicker (“Sorry, we’re not hiring right now, try again later.”) or ended differently (“You’re hired”).
FWIW I think that accepting/rejection emails are unusually hard to write, because it’s such a big event for someone else’s life, that I always spend a lot of cognitive cycles trying to do a good job, and it’s a massive cost when you have to process like 30 of them. Quite stressful. Even for people who are strong operationally, I think this is a quite different and unique task.
Large companies that do send rejection emails (e.g. Google) keep them short and blunt. Y Combinator iirc had a message (or link?) basically saying “our rejection likely doesn’t mean anything wrong with you or your idea, there’s just a ton of applicants”. If rejection emails are being personalized and it’s taking a while, then whatever process is used for them sounds like a bad process (even if the process is “just” somebody spending lots of mental cycles doing it). Unfortunately, this moves my mental model of MIRI another micron towards “an organization of procrastination / prioritization issues / ???”.
For me, the difficulty was basically never “saying no” to someone, or figuring out how to politely say “no” to people, and was almost always deciding whether or not to say no to someone. Lots of people were borderline in various ways, or it wasn’t clear which category of people should get the next unit of attention (and so if their category isn’t going to get any units of attention in the next month I should just say no to them, but if they are going to get attention then maybe I’ll look at them more closely later, etc.).
I think now I would respond with an email that’s much more like “you’re waitlisted; going off of past numbers, that means this is ~90% a no and ~10% that I’ll get back to you in more detail in 1-3 months” or Buck’s “by default you’re rejected, but if you want more consideration do XYZ” or w/e.
[EDIT]: Like, for workshop attendance, you know you have 24 slots (or w/e), and you can just score people, and have a bunch of definite invites, a long maybe list, and the rest definite rejects, you can tell everyone their status. [Ideally you have the meeting to resolve the maybe list quickly and so you can wait until you have a yes/no/”if someone else drops out” waitlist decision for everyone and don’t need to email anyone twice.] But if someone is like “hey I’m interested in helping, here are some facts about me” they either fit into one of your buckets and you can send them the material for that bucket, or not, in which case it’s unclear. [I had less context for the engineering roles, but my sense is that there were the standard small team problems of “hmm, does this person fit the place where we most want to grow next?” vs. “would we want this person a year from now?” and so on, whereas MANGA is much more “oh yeah, we need a hundred more Xs across the whole org, we’ll hire them and let them find a team and if they don’t find a team, then we’ll let them go.”
FWIW, I think MIRI’s operations team is unusually good compared to peer EA orgs, but mostly weren’t in charge of recruiting.
I’m confused about how the first paragraph connects to the second paragraph. This feels like it’s missing some connecting section that explains what the last paragraph has to do with the rest of the process.
Oh sorry looks like I accidentally published a draft.
I’m not sure I make the connection between what you described and meta-honesty norms. The flakiness and unresponsiveness you described is unfortunately endemic to company recruiters. I attribute it to the fact that corresponding with lots of job applicants is a lot of work which is not fun, and bad incentive structures that don’t appropriately reward/punish doing a good/bad job with hiring. It’s too bad that MIRI hasn’t managed to outperform that status quo, but I guess they didn’t.
This is a really important post I think, and I hope it gets seen by the right people! Developing a culture in which we can trust each other is really essential, and I do wish there was more focus on progress being viewable from an outside perspective.