Disagree. Before I started at METR I dropped out of Caltech and had written two random interp papers. Formally, I knew only as much as the average undergrad about experimental design, stats, or data viz, but 4 months later the time horizon paper was out and I had significantly contributed to the methodology and writing. Mostly I just needed general intelligence plus enough knowledge about AI safety to have research taste, and someone with more intelligence than me could pick up knowledge faster.
Many of my colleagues generate heaps of utility with other skills: being cracked engineers, highly reliable generalists, good at maintaining relationships with labs, or having research taste in other areas like human uplift experiments. AIs are now good tutors for many technical areas so the barrier to acquiring many technical skills is lower.
On top of all this, the best out of 100 candidates is much better than the best of 10 candidates. It’s potentially 10 good hires instead of 1, and 1~2 exceptional hires instead of 0! Safety orgs are elastic to the quality of the hiring pool and will hire more if it increases. So the expected impact of someone smart or otherwise cracked is potentially huge.
Before I started at METR I dropped out of Caltech and had written two random interp papers
Not to belabor your specific case—I agree directionally that there’s no intense gatekeeping or connections needed—but you worked at Jane Street and MIRI and then did MATS before joining METR right . . . OP’s claim isn’t that specific time-consuming credentials are needed but that even “smarts + unpredictable fit” things are hard to signal in an application from outside (without PhDs or work experience).
Fair enough. OP sounded like they were talking about domain specific experience though, rather than credentials. And I think my past experience transferred poorly.
I only interned at Jane Street but that does have a lot of signaling value.
How long ago did you do this? I’m a cracked engineer / highly reliable generalist with non-zero research taste and I can barely get AI safety jobs or fellowships to even talk to me, regardless of what type of position I apply for. It’s fine if they only want the 1-2 exceptional candidates, but that’s “friends with a PhD in machine learning”, not “smart friends”.
Building a network and creating artefacts seem particularly important. It’s not just about being smart. No idea what you’ve already done but I feel like I see plenty of decent people get in and cracked people not get in, so I don’t think the bar is impossibly high. Higher for technical research than some other parts though
I agree that it’s possible to get in, but it has a lot to do with legible experience and networks, and being smart without those things is mostly a recipe for frustration
Edit: And my point is that a field that needs more applicants doesn’t look like this.
People are still hiring now though. METR is hiring for evals execution, and probably more things. Epoch is building out several teams including benchmarking. Labs are hiring. All orgs have many people without a PhD in machine learning. The bar is high but it’s much more about smarts + unpredictable fit things than any formal qualification.
Disagree. Before I started at METR I dropped out of Caltech and had written two random interp papers. Formally, I knew only as much as the average undergrad about experimental design, stats, or data viz, but 4 months later the time horizon paper was out and I had significantly contributed to the methodology and writing. Mostly I just needed general intelligence plus enough knowledge about AI safety to have research taste, and someone with more intelligence than me could pick up knowledge faster.
Many of my colleagues generate heaps of utility with other skills: being cracked engineers, highly reliable generalists, good at maintaining relationships with labs, or having research taste in other areas like human uplift experiments. AIs are now good tutors for many technical areas so the barrier to acquiring many technical skills is lower.
On top of all this, the best out of 100 candidates is much better than the best of 10 candidates. It’s potentially 10 good hires instead of 1, and 1~2 exceptional hires instead of 0! Safety orgs are elastic to the quality of the hiring pool and will hire more if it increases. So the expected impact of someone smart or otherwise cracked is potentially huge.
Not to belabor your specific case—I agree directionally that there’s no intense gatekeeping or connections needed—but you worked at Jane Street and MIRI and then did MATS before joining METR right . . . OP’s claim isn’t that specific time-consuming credentials are needed but that even “smarts + unpredictable fit” things are hard to signal in an application from outside (without PhDs or work experience).
Fair enough. OP sounded like they were talking about domain specific experience though, rather than credentials. And I think my past experience transferred poorly.
I only interned at Jane Street but that does have a lot of signaling value.
How long ago did you do this? I’m a cracked engineer / highly reliable generalist with non-zero research taste and I can barely get AI safety jobs or fellowships to even talk to me, regardless of what type of position I apply for. It’s fine if they only want the 1-2 exceptional candidates, but that’s “friends with a PhD in machine learning”, not “smart friends”.
Building a network and creating artefacts seem particularly important. It’s not just about being smart. No idea what you’ve already done but I feel like I see plenty of decent people get in and cracked people not get in, so I don’t think the bar is impossibly high. Higher for technical research than some other parts though
I agree that it’s possible to get in, but it has a lot to do with legible experience and networks, and being smart without those things is mostly a recipe for frustration
Edit: And my point is that a field that needs more applicants doesn’t look like this.
December 2024, so 1.6 years ago.
People are still hiring now though. METR is hiring for evals execution, and probably more things. Epoch is building out several teams including benchmarking. Labs are hiring. All orgs have many people without a PhD in machine learning. The bar is high but it’s much more about smarts + unpredictable fit things than any formal qualification.