Dan

mental healthcare, AI, music

[Dan @DTLawhon]

This is probably a long shot, but does anyone have a contact at @Rippling they could ping to help with an urgent matter for an employee? Standard support has said they cannot help further. @parkerconrad

[Dan @DTLawhon]

@SilasMahner A Ro spotting in the wild!

[Dan @DTLawhon]

@signulll @ianmisner Not 100% sure how this would play out but if SI displaces human labor because it's lower cost, that should eventually show up in lower cost goods, which would allow human labor wages to deflate as well into a new equilibrium.

[Dan @DTLawhon]

@FourVork Unironically leetcode is great for this.

[Dan @DTLawhon]

@mignano If we use the compute efficiency of the human brain as a known achievable lower bound, there are still several orders of magnitude costs could ultimately come down. Once it's too cheap to meter, just define complexity and usage limits to allow 3-5 tiers of price discrimination.

[Dan @DTLawhon]

@mignano Just depends how low inference costs ultimately get. The average person doesn't need anything nearly as intensive as a programmer or enterprise AI user. I could see a world with $20 subscriptions that are 90% gross margin. Basically back to the software economics status quo.

[Dan @DTLawhon]

@DTLawhon But what about ppl who can’t afford a subscription.

[Dan @DTLawhon]

If inference costs go towards zero then keeping with a subscription model seems best all around, with maybe a few different tiers depending on complexity/quantity of tasks.

Gives companies recurring revenue which is much more valuable than usage-based revenue, and it gives consumers a predictable monthly cost instead of varying by each task.

The cognitive overhead of deciding whether $ x price is worth it per task would introduce a lot of friction into using the product.

[Dan @DTLawhon]

@SilasMahner if we can get 9 other billionaires to make a similar commitment that would be $2500. Literally lifechanging for multiple generations when compounding over the decades.

[Dan @DTLawhon]

@SilasMahner @jameszimmermann Hey, this could be arranged 😂

[Dan @DTLawhon]

@alexandrosM @foomagemindset Haven't read Tetlock directly so I can't evaluate, but that would be a fair and strong pushback if so.

[Dan @DTLawhon]

@tracewoodgrains I would think in that case pushing for a private discussion instead of a public debate would be more fruitful, but maybe they've had those discussions.

[Dan @DTLawhon]

Idk if that warrants the approach he takes in the piece, or if that approach is likely to persuade Pinker, but it's a fair assumption that highly prominent public intellectuals who are specifically respected by the ~2% most influential people in society will play maybe a determinative role in things like regulation and broader elite attitudes towards AI.

[Dan @DTLawhon]

@tracewoodgrains This would be, I think, a very high quality use of time because I suspect your questions will overlap with most reasonable people's questions, and you have the platform to get those questions answered.

[Dan @DTLawhon]

@foomagemindset @alexandrosM I think separately from the merit of tetlock's approach, Scott is making the point that Pinker accepts and endorses Tetlock's methodology explicitly elsewhere, so to criticize it only when used for AI predictions seems like a blindspot.

[Dan @DTLawhon]

@lymanstoneky We already use computers to do computations that would be completely infeasible by hand, but they never perform any operation that is in principle impossible for us to do by hand.

In other words, human brains are Turing complete if aided by external memory.

[Dan @DTLawhon]

@lymanstoneky I don't think so, no. Because the AI proofs aren't fundamentally inscrutable, they are just so lengthy and meandering that it is difficult for a human to hold all of the pieces in their working memory. That limitation has always been the case.

[Dan @DTLawhon]

Not to speak for him, but I don't think that's exactly his argument. Eliezer believes that the threat of AI is overdetermined, meaning there are multiple converging paths that lead to the ~same outcome.

A slow takeoff scenario (slow enough for capabilities to show up in economic growth) is maybe less risky than a fast takeoff scenario all else equal, but in practice such prosperity will incentivize people to press the gas even harder. The benefits of AI will be highly concrete while the risks are still intangible. So you still eventually hit the capabilities point-of-no-return, but maybe it takes 7 years instead of 2 years, or whatever. Better in some sense, but if the end result is the same, not that much better.

So I don't think it would actually move Eliezer's beliefs all that much.

On the other hand, my understanding of Paul's views is that such a lag in economic growth should update him towards a much higher risk assessment because he is counting on added wealth unlocking exponentially more safety resources.

I could be wrong about either of their views, though.

[Dan @DTLawhon]

@RogerThisdell Very good explanation.

[Dan @DTLawhon]

@redaction I feel this. I think it's the same answer under any uncertainty, which is follow whatever engages you enough to get really good at it. Worst case scenario you're satisfying your own curiosity.