Ryan Greenblatt

In AI 2027, we predicted that AI would take over the world or irreversibly concentrate power. In AI 2040: Plan A, we've laid out our positive vision for what should happen instead.

Overall, it seems like our most important disagreement is probably about the AI/robots dominating humans and growing massively in population. Our views on economic growth from AI come from these arguments, not the modelling. On the specifics of the econ model, we aren't confident in our modelling choices, but i'll explain them below: > On sigma 1.1: This is only on the top K / L_eff CES layer, its not a huge crux, e.g., at 0.6 the model has 180x GWP by 2040 instead of 200x, but 1.1 seems more reasonable to me because AI/robots enter through L_eff, and don't seem like they should be bounded hard by K. > On human-inherent tasks = 0: This is wrong in the plan A world with slowdown + regulation, in 2040 we have 1% of tasks for humans (even then i think they will have large AI uplift, and it'll be more from regulation than human-preference). We do think its true about a default world though because we believe either in AI takeover, or just human-preference demand becoming a very niche/negligible part of the new economy (e.g., analogous to counting bartering tribes in today's GWP). > on frozen task space: sorry about this, it was a remnant of an earlier version. I actually intend for the exogenous automation curves to be representing % of current-year tasks, so including new tasks/goods. I will rename the cases of "% 2024 tasks". The model/numbers don't change from this, because the autoamtion curve is exogenous. > on segmented thing / nothing is scarce point: There are scarce inputs. It is all consistently in real 2025 $ to anchor the 'real' growth level on average but relative prices will move a lot, and we discuss this in the supplement. We imagine e.g., land, positional goods, etc. going up in price but goods/services going down a lot. It's pretty unclear how 'real' would get measured in practice though, but at the end of the day the growth is driven by the explosion in highly capable AI/robot labor.

Do you have a take on the economic model yet? I still think it is quite misleading and flawed. In some cases the modeling choices are caveated in margin notes (e.g. sigma = 1.1) with often handwavy reasoning, but given all the main text/graphs take the most aggressive possible modeling choices, it looks like a form of narrative spin to me. Still going through it but for example: - sigma is set at 1.1, on the substitution side of CES, with an arguably circular argument for why - human-inherent tasks are defined to provide 0 contribution to real output by assumption and are entirely outside the production function - no new tasks or goods are added to the economy by assumption; it's a one-good model - GWP calculations are systematically wrong: pre-2025 is calculated as normal, but post-2025 is calculated from the industrial explosion's output valued at frozen 2025 marginal prices in a one-good model which assumes there is nothing scarce. so it's like a segmented time-series that produces an explosion only by changing the calculation method. I played around with the simple model parameters to get some set of results which seem actually likely under an AGI and robotics powered high growth regime: 2040 GWP = ~2x 2025 GWP; 12yr fastest GWP doubling; tasks ~70% automatable by 2040; sigma = 0.6

Overall, it seems like our most important disagreement is probably about the AI/robots dominating humans and growing massively in population. Our views on economic growth from AI come from these arguments, not the modelling. On the specifics of the econ model, we aren't confident in our modelling choices, but i'll explain them below: > On sigma 1.1: This is only on the top K / L_eff CES layer, its not a huge crux, e.g., at 0.6 the model has 180x GWP by 2040 instead of 200x, but 1.1 seems more reasonable to me because AI/robots enter through L_eff, and don't seem like they should be bounded hard by K. > On human-inherent tasks = 0: This is wrong in the plan A world with slowdown + regulation, in 2040 we have 1% of tasks for humans (even then i think they will have large AI uplift, and it'll be more from regulation than human-preference). We do think its true about a default world though because we believe either in AI takeover, or just human-preference demand becoming a very niche/negligible part of the new economy (e.g., analogous to counting bartering tribes in today's GWP). > on frozen task space: sorry about this, it was a remnant of an earlier version. I actually intend for the exogenous automation curves to be representing % of current-year tasks, so including new tasks/goods. I will rename the cases of "% 2024 tasks". The model/numbers don't change from this, because the autoamtion curve is exogenous. > on segmented thing / nothing is scarce point: There are scarce inputs. It is all consistently in real 2025 $ to anchor the 'real' growth level on average but relative prices will move a lot, and we discuss this in the supplement. We imagine e.g., land, positional goods, etc. going up in price but goods/services going down a lot. It's pretty unclear how 'real' would get measured in practice though, but at the end of the day the growth is driven by the explosion in highly capable AI/robot labor.

how do you deal with the consumption argument? essentially it seems like the longer your robots focus on just building more robots, the longer your payoff of investment, and in the limit it's unclear who buys all the downstream goods if you're making exponentially more of them but meanwhile if you have robots this capable there will be gigarevenue from just selling them directly it's plausible there's enough belief to justify the spend regardless, but that depends on how much investment would be required and expected payback period

very often my response to something like rsi is "we've already been rsi-ing for some time" or my response to superexponential gdp is "gpd has always been best-regressed by a hyperbola" or my response to the project is "we've already been collaborating with the government on important tech for some time" or my response to agi is too powerful to ship is "we've been shipping the power of gods for a century"

Exciting political developments in the AI Safety space