It can vary enormously based on risk factors, choice of car, and quantity of coverage, but that does still sound extremely high to me.  I think even if you're a 25-yo male with pretty generous coverage above minimum liability, you probably won't be paying more than ~$300/mo unless you have recent accidents on your record.  Gas costs obviously scale ~linearly with miles driven, but even if your daily commute is a 40 mile round-trip, that's still only like $200/mo.  (There are people with longer commutes than that, but not ones that you can easily substitute for with an e-bike; even 20 miles each way seems like a stretch.)

This inspired me experiment with a "loyal Vizier"-themed system prompt.  Not amazing, but maybe there's something there.

I expect this has most of the costs of keeping them running full-time with few of the other benefits.

(iirc Eric thinks that the difference for Bores was that it was ~20% better to donate on the first day, that the difference would be larger for Bores than for Wiener, and that "first day vs. not first day" was most of the difference, so if it's more than few percent more costly for you to donate now rather than 2 months from now I'm not sure it makes sense to do that.)

Oh, alas.  Thank you for the correction!

(I still expect OpenPhil the LLC to have been paying comparable amounts to its most-remunerated employees, but not so confidently that I would assert it outright.)

OpenPhil's 5th-highest-compensated employee earned about $184k in 2023^[1], which gives you a ceiling.  Anthropic currently extends offers of ~$550k to mid-level^[2] engineers and researchers.  Joe's role might not be on the same ladder as other technical roles, but companies like Anthropic tend to pay pretty well across the board.

Edit: retracted first half of the claim, see this reply.

1. ^{^}

   According to their public Form 990 filing.

2. ^{^}

   I realize the job title says "Senior Software Engineer", but given the way their ladder is structured, I think mid-level is probably closer (though it's fuzzy).

Curated.  I disagree with some stronger/broader forms of the various claims re: missing "mental elements", but I'm not sure you intend the stronger forms of those claims and they don't seem load bearing for the rest of the piece in any case.  However, this is an excellent explanation^[1] of why LLM-generated text is low-value to engage with when presented as a human output, especially in contexts like LessWrong.  Notably, most of these reasons are robust to LLM output improving in quality/truthfulness (though I do expect some trade-offs to become much more difficult if LLM outputs start to dominate top human outputs on certain dimensions).

1. ^{^}

   To the point where I'm tempted to update our policy about LLM writing on LessWrong to refer to it.

No.  It turns out after a bit of digging that this might be technically possible even though we're a ~7-person team, but it'd still be additional overhead and I'm not sure I buy that the concerns it'd be alleviating are that reasonable^[1].

1. ^{^}

   Not a confident claim.  I personally wouldn't be that reassured by the mere existence of such a log in this case, compared to my baseline level of trust in the other admins, but obviously my epistemic state is different from that of someone who doesn't work on the site.  Still, I claim that it would not substantially reduce the (annualized) likelihood of an admin illicitly looking at someone's drafts/DMs/votes; take that as you will.  I'd be much more reassured (in terms of relative risk reduction, not absolute) by the actual inability of admins to run such queries without a second admin's thumbs-up, but that would impose an enormous burden on our ability to do our jobs day-to-day without a pretty impractical level of investment in new tooling (after which I expect the burden would merely be "very large").

I don't think we currently have one.  As far as I know, LessWrong hasn't had any requests made of it by law enforcement that would trip a warrant canary while I've been working here (since July 5th, 2022).  I have no information about before then.  I'm not sure this is at the top of our priority list; we'd need to stand up some new infrastructure for it to be more helpful than harmful (i.e. because we forgot to update it, or something).

Expanding on Ruby's comment with some more detail, after talking to some other Lightcone team members:

Those of us with access to database credentials (which is all the core team members, in theory) would be physically able to run those queries without getting sign-off from another Lightcone team member.  We don't look at the contents of user's DMs without their permission unless we get complaints about spam or harassment, and in those cases also try to take care to only look at the minimum information necessary to determine whether the complaint is valid, and this has happened extremely rarely^[1].  Similarly, we don't read the contents or titles of users' never-published^[2] drafts.  We also don't look at users' votes except when conducting investigations into suspected voting misbehavior like targeted downvoting or brigading, and when we do we're careful to only look at the minimum amount of information necessary to render a judgment, and we try to minimize the number of moderators who conduct any given investigation.

1. ^{^}

   I don't recall ever having done it, Habryka remembers having done it once.

2. ^{^}

   We do see drafts that were previously published and then redrafted in certain moderation views.  Some users will post something that gets downvoted and then redraft it; we consider this reasonable because other users will have seen the post and it could easily have been archived by e.g. archive.org in the meantime.