Users in the replies enthusiastically cheer Scott Alexander to harshly rebut opponents misusing epistemic humility claims in AI safety debates.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@slatestarcodex @ramez ROAST HIS ASS SCOTT
I've previously tried to rebut this argument as a kind of rebranding of arrogance as humility, especially around COVID (see https://slatestarcodex.com/2020/04/14/a-failure-but-not-of-prediction/ ). As COVID was ramping up, people kept saying it would be insufficiently "humble" to take any action against it, because we hadn't yet "proven" that it could become a worldwide pandemic. As there became a 10% chance, a 25% chance, and then a 50% chance that it would, people continued not to take action, because there still wasn't the "proof", and it would be "arrogant" to prepare for something that merely had a 25% chance of killing millions of people. But this is the wrong way of thinking about things. The people who were implicitly acting as if there was a 100% chance that COVID wouldn't be dangerous, were calling the people assuming a 75% chance that it wouldn't be "arrogant" and accusing them of being "absolutely sure" that it would be dangerous! If you allowed a 10% or 25% or even a 1% chance that AI could be dangerous, I think you would be extremely invested in building the global capacity for a potential off switch - not necessarily in pulling that off switch now, but in making sure it existed and can be pulled on short notice. That's AI2040's current ask - building the technology to verify an agreement if it needs to happen later. It's the only thing that needs to be done before the situation clarifies in 2027 or 2028. I think if you are indeed trying to being humble and leave room for all possibilities, you should refocus on supporting it. Also, we at least have lines on graphs suggesting our position might be right. Where's your argument that regulating chips the same way we regulate every other product would lead to a "global surveillance state"? When it's about blocking AI safety, all people require is hand-waving; when it's about trying to create the capacity for it, no argument, no matter how strong, will ever be good enough.
With COVID, the world had seen global epidemics and pandemics before, and we had concrete evidence that they could kill tens of hundreds of millions. Risk of that scale from fast spreading disease is well established from actual examples of it happening. With AI x-risk, there is no such empirical evidence. There are speculative forecasts of it, that depend on a host of assumptions that are debatable and grounded more in thought experiments than empirical evidence. I favor quite a bit of safety work, and the enactment of safety policies that are no-regret. To justify more severe and restrictive policies, I would need to see actual evidence that AI x-risk exists and is a looming danger. I don't.
I've written extensively about this kind of rebranding of arrogance as humility, especially around COVID. As COVID was ramping up, people kept saying it would be insufficiently "humble" to take any action against it, because we hadn't yet "proven" that it could become a worldwide pandemic. As there became a 10% chance, a 25% chance, and then a 50% chance that it would, people continued not to take action, because there still wasn't the "proof", and it would be "arrogant" to prepare for something that merely had a 25% chance of killing millions of people. This is crazy logic. The people who were implicitly acting as if there was a 100% chance that COVID wouldn't be dangerous, were calling the people assuming a 75% chance that it wouldn't be "arrogant" and accusing them of being "absolutely sure" that it would be dangerous! If you allowed a 10% or 25% or even a 1% chance that AI could be dangerous, I think you would be extremely invested in building the global capacity for a potential off switch - not necessarily in pulling that off switch now, but in making sure it existed and can be pulled on short notice. That's our current ask, and I think your platform would be better spent in supporting that than in arguing that anyone who talks about it must be overconfident because their confidence in your position isn't exactly 100%. Also, we at least have lines on graphs suggesting our position might be right. Where's your argument that regulating chips the same way we regulate every other product would lead to a "global surveillance state"? When it's about blocking AI safety, all you require is hand-waving; when it's about trying to create the capacity for it, no argument, no matter how strong, will ever be good enough.
Investor Ramez Naam pushed back, citing a lack of empirical precedents for AI extinction.
@slatestarcodex @ramez ROAST HIS ASS SCOTT
Users in the replies enthusiastically cheer Scott Alexander to harshly rebut opponents misusing epistemic humility claims in AI safety debates.
Based on 1 visible X reactions from 1 accounts; directional sample.
Ask a question below.
Published answers will appear here.