"Additionally, another implication from AH’s model based on point 4 is that if the probability of discovery changes over time, it is unclear whether this will lead to more research being conducted." doesn't point 4 explicitly state otherwise? i get that this doesn't say anything about the inverse (lower probability means lower level of research) and i hear your point about the extremes, but to continue on your example, some of the ai breakthroughs have been through "just" scaling ie bigger models on bigger machines. i guess the capex requirements could arguably make the discovery hard, but seems like the driving factor is much the monopoly profits, as shortly held as they may be, over fear of losing said monopoly due to easy future discoveries
Good question and catch. To elaborate on that point (I skipped it in the article for brevity) - AH, in their paper, proposed a world in which we transition between 'discovarbility' states. For example, assume there are 3 possible states of the world - high discovery chance, medium discovery chance and low discovery chance. We transition between the states probabilistically.
Suppose the state "high discovery chance" changes and becomes "even higher discovery chance". Point 4 tells us that this change will result in increased research conducted whenever the world is in this "even higher discovery chance" state. But, in the other states - "medium discovery" and "low discovery" - the amount of conducted research could fall. This is because firms would be worried about conducting too much research in these lower states, since when the world transitions into the "even higher discovery" state, even if the firm managed to discover a new technology, they will get to be a monopoly for shorter time now.
Therefore, overall across the three states, total conducted research might fall.
ah, so it's a distinction between a state evolving versus transitioning between states, and seemingly not wanting to "waste resources" in the lower states. thanks for the explanation, fascinating insight
"Additionally, another implication from AH’s model based on point 4 is that if the probability of discovery changes over time, it is unclear whether this will lead to more research being conducted." doesn't point 4 explicitly state otherwise? i get that this doesn't say anything about the inverse (lower probability means lower level of research) and i hear your point about the extremes, but to continue on your example, some of the ai breakthroughs have been through "just" scaling ie bigger models on bigger machines. i guess the capex requirements could arguably make the discovery hard, but seems like the driving factor is much the monopoly profits, as shortly held as they may be, over fear of losing said monopoly due to easy future discoveries
also, phenomenal piece
Good question and catch. To elaborate on that point (I skipped it in the article for brevity) - AH, in their paper, proposed a world in which we transition between 'discovarbility' states. For example, assume there are 3 possible states of the world - high discovery chance, medium discovery chance and low discovery chance. We transition between the states probabilistically.
Suppose the state "high discovery chance" changes and becomes "even higher discovery chance". Point 4 tells us that this change will result in increased research conducted whenever the world is in this "even higher discovery chance" state. But, in the other states - "medium discovery" and "low discovery" - the amount of conducted research could fall. This is because firms would be worried about conducting too much research in these lower states, since when the world transitions into the "even higher discovery" state, even if the firm managed to discover a new technology, they will get to be a monopoly for shorter time now.
Therefore, overall across the three states, total conducted research might fall.
And glad you liked the article :)
ah, so it's a distinction between a state evolving versus transitioning between states, and seemingly not wanting to "waste resources" in the lower states. thanks for the explanation, fascinating insight
Have you seen the new paper on AGI by Restrepo? It's more of a thought experiment, but it's interesting.
Just had a look at it. Definitely interesting. Funnily enough, I would 'challenge' Restrepo's paper with the Acemoglu/Restrepo paper - https://www.nber.org/system/files/working_papers/w24196/w24196.pdf
But regardless, an interesting way of thinking through the implications
I have similar sentiments. I think it depends on if this level of AGI is actually possible, but it's all hypothetical right now.
Thank you, I just learned something! Knowledge is power and you just empowered me!
Thank you and very glad you found it informative!