Thanks! I only know a handful of people in this category, but for what it's worth, it again feels like people who were predisposed to thinking that working on pretraining would be okay rather than them being "corrupted." 

E.g., I recently talked to someone who told me that their main takeaway from a safety fellowship was realizing that they didn't fit in because they actually weren't worried about existential risk in the same way that the other attendees were.

People seem surprised and bewildered when AI folks defect away from AI safety towards capabilities. People trust that as AI companies grow, those gaining power and money from shares will not be adversely influenced by that power and money.

fwiw I don't actually know many examples of this, and the ones I hear cited often seem uncompelling to me. E.g.:

• Greg Brockman doesn't seem like a true believer in OpenAI's nonprofit mission who got corrupted but rather someone who went into it wanting to make a profit
• Mechanize's founders don't seem like EAs who got corrupted by AI money but rather EAs with unusual moral and empirical views which result in them thinking that the best course of action is the exact opposite of what most EAs think

(Counterexamples appreciated, though!)

And credit to the AI skeptics that they seem to mostly have updated in light of the new evidence (or at least claimed that they never actually believed in long timelines, which is maybe less noble, but ends up in the same place).

Yeah I agree that if you only have one bit of detail that you can store, then saying it is "hard" rather than "easy" is probably the correct bit. However I would suggest that for something as important as your career you should investigate in substantially more detail. If you do so I expect you will come up with a range of needed skills/attributes for these jobs, some of which you might find easy, others of which you might find hard. 

I no longer work at METR. I would guess that they'd be excited about applicants who have done this, but don't want to speak for them.

Cool! Impressive numbers.

Table 1 shows the techniques used; the teams which were allowed to use SAEs (an interpretability technique) used them; the one which was prohibited from using them searched the data.

Also note that "training data" does not mean "instructions". Section 3 describes their training process. 

I see, thanks! I'm not sure exactly what you'd consider as evidence here, but e.g. here's citation count on papers from the past year vs. AI Lab Watch safety rating^[1]

1. ^{^}

   Raw data. Note that anthropic doesn't use arxiv, which affects their citation counts. This is just coming from a dumb search of semantic scholar; I expect a lot of disagreement could be had over the exact criteria for considering something "interpretability" but I expect the Ant/GDM > OAI >> * ordering to be true for almost any definition.

I suspect that I'm still misunderstanding you, but: eg interpretability tools are empirically able to identify misalignment, which feels like a (somewhat simple example of) the thing we want. Neel Nanda's 80k podcast goes over the state of the field; tldr is roughly that there are pretty meaningful advances but also he's skeptical that it will be a silver bullet.

I agree with Ben Stewart that there's a galaxy-brain argument that these positive impacts are outweighed by accelerating progress, but it seems hard to argue that things like interpretability aren't making progress on their own terms.