David Higgs

Cursus frenandus est. The race must be bridled.

[David Higgs @Higgs_David_J]

@ramez (2/2) And METR in their original paper had sections on messiness and other differences to real tasks, plus they ran subsequent studies. The doubling time proved consistent, but the absolute value was never meant to be applied naively

[David Higgs @Higgs_David_J]

@ramez It depends on who's doing the communicating: I agree many (including some otherwise smart people) took it overly straightforwardly. But I re-read the beginning of AI 2027 plus part of their timelines supplement as a sanity check and they didn't. (1/2)

[David Higgs @Higgs_David_J]

@irrepldjw @allTheYud @KelseyTuoc Basic economics. Although AI automating all economic value plausibly changes the underlying economics enough that socialism becomes rational, albeit still market socialism w/ UBI or similar.

Remember that Powerful Person can be rich, bureaucrat, general, mob boss, multiple, etc.

[David Higgs @Higgs_David_J]

@ramez (2/2) For example, see the "benchmarks and gaps" & "time horizon extension" methods that AI 2027's (overly bullish) numbers are based off of. To do 1-month long actual internal AI engineering tasks, one author guessed ~10 years on METR benchmarks would be sufficient

[David Higgs @Higgs_David_J]

@ramez Nonono, I mean neither METR, nor AI 2027 authors, were/are saying that an "n-hour time horizon" on the METR benchmark actually corresponds to n-hours of real world work! So I was asking what the "prediction" gap was, but you just used naive numbers I think? (1/2)

[David Higgs @Higgs_David_J]

@Schwarzschild2c @robertskmiles In a trivial sense sure, but e.g. most of what Bostrom wrote in Superintelligence still holds up. And there are some people who clearly had a wide distribution, so is them merely narrowing to particular paths w/o being surprised "not shifting opinions" based on new evidence?

[David Higgs @Higgs_David_J]

@the_yanco @JeremyKritz @robertskmiles It's possible to think about something in a high quality way within a few (or less) years time span lol... Isn't it more like "if you thought about and dismissed AI then you've earned disrespect of your opinions?"

[David Higgs @Higgs_David_J]

@xereeto @robertskmiles @Aella_Girl It's crazy thinking about how it just gradually then quickly went from decades to years to maybe (though hopefully/probably not) months. Honestly mostly fun for me, though ofc terrifying too when I can engage my abstract reasoning sufficiently

[David Higgs @Higgs_David_J]

@Quinurum @paperclippriors @robertskmiles Probably more like 75-90%, just an ass-pull, but no way 99%. I remember stuff like just getting the AI/genie to understand your wishes being hard for example, not that that was central exactly

[David Higgs @Higgs_David_J]

@ilex_ulmus (2/2) but I don't know if this kind of... toxicity(?) is how to build that. I'm not an expert, but this isn't the kind of thing I could see MLK, Ghandi, Al Gore, Obama, etc. saying. Although Frederick Douglass maybe...

[David Higgs @Higgs_David_J]

@ilex_ulmus As someone who would gladly support a pause, pacing, whatever regulatory action we can get (though not perma-ban), this is very off-putting. Especially how broad of a brush you're painting with here (e.g. "no matter how indirectly").

I want a social movement for AI safety (1/2)

[David Higgs @Higgs_David_J]

@briab_briar @peterwildeford What about in Britain? Or other US states?

[David Higgs @Higgs_David_J]

@NathAdoo @andersonbcdefg @tautologer I agree with everything but the last part, unless by "grow up" you mean "advocate for policy which could (and hopefully will imo) change the current trajectory"

[David Higgs @Higgs_David_J]

@hyperdiscogirl @tautologer Seeing all the constant (and seemingly accelerating) progress is very fun, personally using models is pretty cool sometimes as well. Just worried about that increasingly plausible x-risk barreling down on us...

[David Higgs @Higgs_David_J]

@deredleritt3r @tautologer I'm also optimistic, but only because there are a lot of people taking pessimistic lines of thought extremely seriously and doing things about it

[David Higgs @Higgs_David_J]

@grok @iamianiamianb @allTheYud Aren't tools like fetching web pages enabled in chat sessions by default now?

[David Higgs @Higgs_David_J]

@Danno28_ @patio11 @allTheYud This is sooo true in my experience

[David Higgs @Higgs_David_J]

@ramez Ty, tho I have read that article. What I meant was, how you got a "prediction" that in Google's internal AI research tasks/actual use AI should have been able to do 6 hr tasks, from METR time horizon data. You just meant those were the naive, unadjusted numbers

[David Higgs @Higgs_David_J]

@ben_snow6 @tobyordoxford These are internal frontier AI researchers, they are trying to push the models to their limits obviously, not using them like most people