I have no domain relevant info. I’m also poorly calibrated. Regardless, here’s my uninformed Fermi estimate, which leads me to a 75% interval with 25 on the low end and 1k on the high end. Most of the probability there is from 25 to 250. 100 on the low end, 1k on the high end.
Thus, a total of 25k at the big American labs; I don’t know how big China is at this, but I’m just going to go with the heuristic that I would’ve heard more about them if they were competitive.
Okay, but it’s not like every one of those 25k are contributing equally. Some people are subpar at their job; others revolutionize the field. They may not even all be researchers, though I’m not getting numbers on this so I’ll pretend like they are all doing ML research.
Looking at the original Attention is All You Need paper, there are 8 authors with “equal contribution”, and like half of them have Wikipedia pages, many seem to have done things like cofounded startups, etc. OpenAI supposedly had ~1k employees in 2023.
Of course, people talk with others in the company, and that’s definitely a benefit; but for the purposes of this I’ll just pretend like, say, the top 1-10% of the employees are contributing way more than the rest.
There are then 250 in the top 1% of all top AI-co employees. Having the top 100 leave would probably reduce by at about 2x, then? On the one hand, “replacement work” makes the slowdown lower; on the other hand, the best few are again probably pulling extra weight.
So on the low end, the top 100 people leaving would do the trick. My gut says this is a bit high, and I think it’s due to putting more weight on the “the top do more” factor, so I’ll say 25.
On the high end, it’s more like 1k people needed.
The main uncertainty here is “how is impact distributed among the employees? How much more impactful are the top few researchers, compared to the rest?”. Someone who’s actually worked there, or who’s familiar with recent AI research, could help answer those.
I have no domain relevant info. I’m also poorly calibrated. Regardless, here’s my uninformed Fermi estimate, which leads me to a 75% interval with
25 on the low end and 1k on the high end. Most of the probability there is from 25 to 250. 100 on the low end, 1k on the high end.(edit: I just saw that apparently 50 top staff at OpenAI left, according to this 1 year old article. I think it’s plausibly this slowed them down by at least 2x since then! But it makes me think that my intuitions weren’t properly multiplying by the number of labs, so I scaled up my estimate).
Details:
First, how many employees do the top labs have?
OpenAI: 4.5k, though they plan to double by the end of the year, so I’ll just pretend it’s at 8k.
Anthropic: 2.5k. Given what it sounds like OpenAI is doing, I’m just going to double this to 5k.
Google Deepmind: 6k. Doubling it to 12k.
Thus, a total of 25k at the big American labs; I don’t know how big China is at this, but I’m just going to go with the heuristic that I would’ve heard more about them if they were competitive.
Okay, but it’s not like every one of those 25k are contributing equally. Some people are subpar at their job; others revolutionize the field. They may not even all be researchers, though I’m not getting numbers on this so I’ll pretend like they are all doing ML research.
Looking at the original Attention is All You Need paper, there are 8 authors with “equal contribution”, and like half of them have Wikipedia pages, many seem to have done things like cofounded startups, etc. OpenAI supposedly had ~1k employees in 2023.
Of course, people talk with others in the company, and that’s definitely a benefit; but for the purposes of this I’ll just pretend like, say, the top 1-10% of the employees are contributing way more than the rest.
There are then 250 in the top 1% of all top AI-co employees. Having the top 100 leave would probably reduce by at about 2x, then? On the one hand, “replacement work” makes the slowdown lower; on the other hand, the best few are again probably pulling extra weight.
So on the low end, the top 100 people leaving would do the trick. My gut says this is a bit high, and I think it’s due to putting more weight on the “the top do more” factor, so I’ll say 25.
On the high end, it’s more like 1k people needed.
The main uncertainty here is “how is impact distributed among the employees? How much more impactful are the top few researchers, compared to the rest?”. Someone who’s actually worked there, or who’s familiar with recent AI research, could help answer those.