Update: After the departures of Mira Murati, Barrett Zoph, and Bob McGrew, it's time to revisit this analysis. By my accounting:

Among the first 505 to sign the letter: at least 24 departures (4.8%)

Signed by 702 but not by 505: at least 13 departures (6.6%)

Didn't sign by 702: at least 19 departures (27.9%).

Of course, that's finer-grained than the criterion for this market, which looks likely to resolve to the 80%+ bucket (with a combined attrition of 12% so far). In retrospect I wish I'd separated things out differently (e.g. 95-100, 90-95, 80-90, 60-80, 30-60, 0-30).

Update: Daniel Kokotajlo adds the previously un-noted departures of safety researchers Jan Hendrik Kirchner (signed between 668th and 702nd), Collin Burns (did not sign as of the 702 version), Jeffrey Wu (did not sign as of the 702 version), Jonathan Uesato (did not sign as of the 702 version), Steven Bills (signed between 668th and 702nd), Yuri Burda (did not sign as of the 702 version), and Todor Markov (among the first 505 signers).

Assuming all of these are validated, the lower bound on departures:

Among those who hadn't signed by 702: at least 11 departures out of an imputed 68 (at least 16% attrition)
Among those who had signed by 505 but not by 702: at least 13 departures out of 197 (at least 6.6% attrition)

And I think the lower bound is very loose, given that this is a report about safety researchers specifically. (Though I do expect them in particular to have been pushed out.) I hope to find more data as we get closer to 1 year.

In retrospect, I was overly pessimistic and should have sliced the outcomes up more finely, i.e. "will attrition be less than 5%, 5-10%, 10-20%, 20-40%, or more than 40%".

With John Schulman's departure, the list slowly continues to grow, but I now doubt it will reach 20% by 11/22.

If this spreadsheet tracking departures from OpenAI continues to be updated, I plan to use it in conjunction with the list of 505 original signatories in order to decide the market in November (I will be cross-checking names with other sources in order to validate names as technical staff who joined before-crisis and departed after-crisis): https://docs.google.com/spreadsheets/d/1B1iFZO26Dpl2iN1Jv1ZugyLktX_qxEQRaYkUHaRR8GY/

Lots of false positives here (people who self-reported "working at OpenAI" because they customized their GPTs, etc), so there will be extra legwork for me. And I worry there are also false negatives. But as a starting estimate, of people who were marked as having left in 2024, 13 were among the first 505 to sign (an attrition rate of 2.6%), another 10 were in the first 702 but not the first 505 (an attrition rate of 5.1%), and I've been able to validate 7 departures among the remaining 68 (an attrition rate of 10.3%).

I may have been too pessimistic about the absolute attrition rate (unless there's a major exodus in the next few months), but qualitatively I was right to be worried that one's prospects at OpenAI were strongly correlated with how quickly one signed the letter. I'll let others trade on this info.

Original leaked version with 505 signatures, by the way: https://www.documentcloud.org/documents/24172246-letter-to-the-openai-board-google-docs

@PatrickLV how will you know which ones signed late? I can’t imagine that this data is available

@Soli I found a later copy of the letter with 667 signatures; I'm still looking for a copy with the 730+ signatures that was reported by the end, or (even better) another means of tallying the reported 770 people who were at OpenAI at the time. Any help would be appreciated!

@PatrickLV oh interesting - where did you find this copy?

@Soli https://www.documentcloud.org/documents/24172377-letter-to-the-openai-board

@PatrickLV Ah, and here's one with 702 signatures: https://gwern.net/doc/reinforcement-learning/openai/2023-openai.pdf

@PatrickLV wow can't believe this is public information - thank you for sharing

@Soli All of those iterations were leaks, of course.

Andrej Karpathy, one of the most prominent employees who had not signed the letter by the time it was leaked, has now left OpenAI. He denies it was the result of "any particular event, issue or drama", and I think he was not part of the superalignment team. https://twitter.com/karpathy/status/1757600075281547344

I think some people didn't sign immediately simply bc it was late at night, etc

@JonasVollmer Certainly that adds some noise, but it seems very unlikely to be a coincidence that the superalignment team (save for Ilya) was largely absent from the letter at that point while the rest of the company was mostly present.