dspeyer

Eliezer does not curate any web forum.

He says he is not planning to bomb any data centers, but he's also said that if he were planning to bomb data centers, he'd lie about it in public, so...

The author and most of his associates generally qualify as techno-utopians, and speak routinely of the "glorious transhumanist future".

Smarter-than-human AGI really is different from all previous technologies, in much the same way that homo sapiens is different from all preceding life on earth.

Off the top of my head...

    * A plan exists
    * The plan itself predicts non-obvious results in smaller systems, and those tests have passed (prediction written *before* running test)
    * A bunch of smart people have looked at it and said, "Yes, this looks plausible"

The closest thing to a plan is RLHF, which has failed every toy problem its been thrown at and made everyone in the field say "even if this worked in toy problems, it wouldn't generalize".

I have yet to see a large static typed program that didn't -- somewhere -- run into the limits of static typing and contain a set of workarounds, using void* or linguistic equivalent. That's code a dynamic language doesn't need.

The only code you can be sure isn't buggy is code that doesn't exist.

This isn't a question a non-engineer phone screener can ask. Coming up with first pass filters that don't require an engineer to interpret is harder.

I zoomed the image in and out a bunch -- no change. When I zoomed in enough, I could see the checkerboarding in A, but the overall color stayed the same as B.

I also had trouble calling either "more even". The first one has greater division at the darker end, and the second at the lighter, but they about balance out.

The extreme for me was figure 12. A and B are so similar I can't see the line between them, but C (the "corrected" square) is a completely different shade.

I'm viewing on a data projector. That's probably the reason. Still, it makes me skeptical that there's anything display-agnostic you can do for gamma.

Last time I manually packed structures I was doing GPU programming. I forget the details, but the CPU and GPU had different alignment requirements so anything other than a manually packed structure broke in weird ways.

Is there a good reason for making day 42 (out of 84) the primary endpoint? Did they at least decide to do that before looking at the data?

Because passwords are terrible.

Lots of users pick guessable ones unless you have extensive password rules to stop them, and those rules are a huge pain to your users.

Lots of users have terrible password hygiene, leaving them all sorts of places they shouldn't.

Enough users are going to forget their passwords that you need a recovery mechanism. Which means an attacker needs to break either the password or the recovery. The most common recovery method is email, but Google often is the email provider.

I suspect most people who don't use a password manager reuse passwords. A lot.

If you get the password to one of my unimportant accounts, you get the password to most of them. They're unimportant, but still...

Not so easily. If you're trying to learn really difficult things, classes are a huge help.

Often they don't. Maybe by nominative tuition, but not by median.

They don't mention Researcher Allegiance Effect. Therefore I suspect they did nothing about it. Given the nature of the study, I feel comfortable ignoring the results on that alone, without bothering to find specific methodological flaws.