Matthew Barnett posts analysis of Eliezer Yudkowsky's calibration on AI doom predictions centered on a 2016 statement about Turing test timelines before the end of the world
NathanpmYoung replies that the statement offers limited evidence on timelines.
I think this isn't such a bad take from Yud. I am pretty sure I could suss out any current gen LLM in an hour of an unconstrained Turing test. I know what to look for. They're not devious enough to scheme out *with human-like responses*. @allTheYud do you think you can?
Matthew Barnett@MatthewJBarTo assess whether Eliezer Yudkowsky is calibrated on AI doom, it seems relevant that in 2016 he said he'd be "pretty shocked" if an AI could pass an unrestricted one-hour Turing test before the end of the world.
@MatthewJBar @inductionheads im iffy on eliezer bc it's clear many of his priors were formed well before the transformer paradigm. he argues against "alignment by default" but in ways often entirely unrelated to transformers as they are post-2022
Matthew Barnett@MatthewJBarTo assess whether Eliezer Yudkowsky is calibrated on AI doom, it seems relevant that in 2016 he said he'd be "pretty shocked" if an AI could pass an unrestricted one-hour Turing test before the end of the world.
Huh, seems pretty reasonable to me, maybe depends on exactly what you're imagining by 'no holds barred'? Seems plausible with spiky capability profile there'll still be things (that can be tested in conversation) where models are detectably worse than humans, or weird behavioral artifacts from training they can't suppress, even when they're pretty superhuman overall
Matthew Barnett@MatthewJBarTo assess whether Eliezer Yudkowsky is calibrated on AI doom, it seems relevant that in 2016 he said he'd be "pretty shocked" if an AI could pass an unrestricted one-hour Turing test before the end of the world.
@Jsevillamol This prediction was how he operationalized one of the three central premises of his argument with Bryan Caplan about AI doom. I think it's relevant to his track record, even if not conclusive.
Jaime Sevilla@Jsevillamol@MatthewJBar He has previously owned this was a bad prediction, and has also has made some surprising claims that turned out to be correct eg >16% probability of IMO gold by 2025. I love holding people accountable as much as anyone, but let's not bash people based in a single example.
@davidmanheim @Jsevillamol I'd welcome a more comprehensive evaluation of his predictions. What other falsifiable predictions has he made that directly pertain to AI doom (as opposed to unrelated predictions about other topics)?
David Manheim@davidmanheim@MatthewJBar @Jsevillamol Of course it's relevant, and filtered to support your motivated position about his accuracy. If you wanted to do any kind of evaluation of his track record properly, you'd want to collect a large set and evaluate them, instead of picking out an example where he performed poorly.
To be clear, I'm not sure if current AI could actually pass this test. A lot hinges on how such a test is conducted. I do think the prediction will ultimately end up being wrong though.
Matthew Barnett@MatthewJBarHe said this in the comment section of this post: https://www.econlib.org/archives/2016/03/so_far_my_respo.html
@MatthewJBar He has previously owned this was a bad prediction, and has also has made some surprising claims that turned out to be correct eg >16% probability of IMO gold by 2025.
I love holding people accountable as much as anyone, but let's not bash people based in a single example.
Matthew Barnett@MatthewJBarTo assess whether Eliezer Yudkowsky is calibrated on AI doom, it seems relevant that in 2016 he said he'd be "pretty shocked" if an AI could pass an unrestricted one-hour Turing test before the end of the world.
@MatthewJBar Seems like some evidence but not a lot.
Matthew Barnett@MatthewJBarTo assess whether Eliezer Yudkowsky is calibrated on AI doom, it seems relevant that in 2016 he said he'd be "pretty shocked" if an AI could pass an unrestricted one-hour Turing test before the end of the world.