I think you’re basically right[1], but make a pretty wrong argument:
The current value added to the economy from minerals is worth less than 1% of total Gross World Product. In other words, the scarcity of minerals (such as metal ores) is simply not a big factor bottlenecking economic value in the world. As a result, mining asteroids for their minerals and bringing them back to Earth, even if it could be done profitably, would likely not result in much improvement to human welfare.
Sure, but SOTA generative AI is also currently a small fraction of GDP. The relevant question is how much it would advance GDP/welfare if made extremely cheap/abundant. It think the fraction of GDP on mineral rents is only a mediocre proxy for how much of a bottleneck it is.
(I agree that a natural reference class for generative AI is “all human labor”, but we can also draw a reference class for minerals like “all things which are at all constrained by materials cost”.)
I think making all minerals nearly free would result in a large increase in GDP though I agree the delta is much smaller than for AI. Minimally, you can make gold free and gold has a $15 trillion mkt cap. (Unclear how much we should care about the welfare benefits of gold though...)
Though I don’t know about the claim about misleading rhetorical tactics. I’m pretty unsure here, people often due emphasize and mention benefits if they think that benefits are likely. (And people who place very high probability on catastrophe don’t think this is likely. I personally think massive benefit is >50% likely.)
Sure, but SOTA generative AI is also currently a small fraction of GDP. The relevant question is how much it would advance GDP/welfare if made extremely cheap/abundant.
The fact that it’s a small fraction of GDP indicates that it’s not a big bottleneck to creating consumer value. This applies to both (current) SOTA AI and minerals.
If value were critically bottlenecked by some key resource, then producers of that resource could charge high prices for it, or consumers could try to purchase it in larger quantities, pushing up its contribution to GDP. The fact that minerals are sold for relatively cheap relative to the rest of the economy indicates they aren’t a very important input, in the sense of bottlenecking consumer value [ETA: assuming they aren’t ~perfect substitutes for other goods]. This is not true for things that do take up lots of GDP, like healthcare and housing.
Regarding your specific point: I expect current SOTA generative AI would continue to be a small fraction of GDP even if it were made 10x or 100x cheaper (in terms of inference cost). Do you disagree?
I think making all minerals nearly free would result in a large increase in GDP though I agree the delta is much smaller than for AI. Minimally, you can make gold free and gold has a $15 trillion mkt cap.
Gold is almost entirely used as a store of value rather than a generator of value. As a result, the net welfare effect of mining gold is probably either slightly positive, or negative depending on the cost of mining it. Doubling the supply of gold would be like doubling the supply of Bitcoin: it would result in a windfall to whoever got the new supply, but would make all previous holders of Bitcoin poorer. These effects cancel each other out; it’s a zero sum game.
Real (social) wealth generation is achieved by creating valuable goods and services that satisfy human preferences. This mainly involves utilizing labor and capital to create technology, manufacture items, or provide professional services. Wealth does not primarily come from the amount of raw resources available in the world.
I agree with everything you wrote in your comment, but I’m still not sure I buy the claim:
The fact that it’s a small fraction of GDP indicates that it’s not a big bottleneck to creating consumer value.
I agree it correlates and is evidence, but I still think that the exact claim is easy to counterexample.
In particular, I think the argument you make ignores substitutes. E.g., suppose that quinoa became 1000x cheaper in all regions (cheaper than even current transportation costs). I think quinoa as a fraction of GDP would grow massively despite quinoa currently being a small fraction of GDP. For sufficient reductions in cost, we’d start using quinoa for totally different purposes, e.g. burning it as fuel.
You can also construct toy cases where a given industry/sector is 0% of GDP, but if the price halved, then it would be >50% of GDP due to substitution effects.
It’s probably not worth getting into this further.
Regarding your specific point: I expect current SOTA generative AI would continue to be a small fraction of GDP even if it were made 10x or 100x cheaper (in terms of inference cost). Do you disagree?
I don’t disagree with this. Yeah, my generative AI example wasn’t great and conflated “cheapness is a quality of its own” with the notion of quality itself. What I should have said is that if generative AI training and inference was made 1000x cheaper then generative AI would greatly boost GDP (same as the quinoa case). (Though I agree the analogy is now less tight.)
Or more precisely, if ML GPUs were made 1,000,000x more powerful (at the same cost and energy) (for an overall 1,000,000x reduction in flop/$), this would be reasonably likely to massively increase GDP despite not currently being that high a fraction of GDP (still <1% I think?).
I agree it correlates and is evidence, but I still think that the exact claim is easy to counterexample.
In particular, I think the argument you make ignores substitutes. E.g., suppose that quinoa because 1000x cheaper in all regions (cheaper than even current transportation costs). I think quinoa as a fraction of GDP would grow massively despite quinoa currently being a small fraction of GDP.
Sorry for handwaving some details away: I agree you can construct toy models in which the claim is not true. I think in pretty much any realistic CES production function you construct (and trivially in Cobb-Douglas functions), the substitution effect will not be strong enough to outweigh this consideration for minerals. The essential quantity here is the elasticity of substitution, and more specifically, how substitutable the item is for other things in the economy.
Minerals are not great substitutes for lots of things in the economy: they can’t be eaten, they can’t be used as fuel, they can’t substitute for labor etc. As they become more abundant, I expect the income effect and substitution effect will roughly cancel out, causing their contribution to GDP to neither rise nor fall by much.
I asked GPT-4 to demonstrate the precise effect on total utility of a good falling in price by a factor of a million in two separate toy models, if you’re interested (the conversation is here). But ultimately I agree my language was a bit sloppy and you’re right to point out that there are worlds in which the claim I made is technically not true. (I guess I should have stated it less confidently too.)
I think you’re basically right[1], but make a pretty wrong argument:
Sure, but SOTA generative AI is also currently a small fraction of GDP. The relevant question is how much it would advance GDP/welfare if made extremely cheap/abundant. It think the fraction of GDP on mineral rents is only a mediocre proxy for how much of a bottleneck it is.
(I agree that a natural reference class for generative AI is “all human labor”, but we can also draw a reference class for minerals like “all things which are at all constrained by materials cost”.)
I think making all minerals nearly free would result in a large increase in GDP though I agree the delta is much smaller than for AI. Minimally, you can make gold free and gold has a $15 trillion mkt cap. (Unclear how much we should care about the welfare benefits of gold though...)
Though I don’t know about the claim about misleading rhetorical tactics. I’m pretty unsure here, people often due emphasize and mention benefits if they think that benefits are likely. (And people who place very high probability on catastrophe don’t think this is likely. I personally think massive benefit is >50% likely.)
The fact that it’s a small fraction of GDP indicates that it’s not a big bottleneck to creating consumer value. This applies to both (current) SOTA AI and minerals.
If value were critically bottlenecked by some key resource, then producers of that resource could charge high prices for it, or consumers could try to purchase it in larger quantities, pushing up its contribution to GDP. The fact that minerals are sold for relatively cheap relative to the rest of the economy indicates they aren’t a very important input, in the sense of bottlenecking consumer value [ETA: assuming they aren’t ~perfect substitutes for other goods]. This is not true for things that do take up lots of GDP, like healthcare and housing.
Regarding your specific point: I expect current SOTA generative AI would continue to be a small fraction of GDP even if it were made 10x or 100x cheaper (in terms of inference cost). Do you disagree?
Gold is almost entirely used as a store of value rather than a generator of value. As a result, the net welfare effect of mining gold is probably either slightly positive, or negative depending on the cost of mining it. Doubling the supply of gold would be like doubling the supply of Bitcoin: it would result in a windfall to whoever got the new supply, but would make all previous holders of Bitcoin poorer. These effects cancel each other out; it’s a zero sum game.
Real (social) wealth generation is achieved by creating valuable goods and services that satisfy human preferences. This mainly involves utilizing labor and capital to create technology, manufacture items, or provide professional services. Wealth does not primarily come from the amount of raw resources available in the world.
I agree with everything you wrote in your comment, but I’m still not sure I buy the claim:
I agree it correlates and is evidence, but I still think that the exact claim is easy to counterexample.
In particular, I think the argument you make ignores substitutes. E.g., suppose that quinoa became 1000x cheaper in all regions (cheaper than even current transportation costs). I think quinoa as a fraction of GDP would grow massively despite quinoa currently being a small fraction of GDP. For sufficient reductions in cost, we’d start using quinoa for totally different purposes, e.g. burning it as fuel.
You can also construct toy cases where a given industry/sector is 0% of GDP, but if the price halved, then it would be >50% of GDP due to substitution effects.
It’s probably not worth getting into this further.
I don’t disagree with this. Yeah, my generative AI example wasn’t great and conflated “cheapness is a quality of its own” with the notion of quality itself. What I should have said is that if generative AI training and inference was made 1000x cheaper then generative AI would greatly boost GDP (same as the quinoa case). (Though I agree the analogy is now less tight.)
Or more precisely, if ML GPUs were made 1,000,000x more powerful (at the same cost and energy) (for an overall 1,000,000x reduction in flop/$), this would be reasonably likely to massively increase GDP despite not currently being that high a fraction of GDP (still <1% I think?).
Sorry for handwaving some details away: I agree you can construct toy models in which the claim is not true. I think in pretty much any realistic CES production function you construct (and trivially in Cobb-Douglas functions), the substitution effect will not be strong enough to outweigh this consideration for minerals. The essential quantity here is the elasticity of substitution, and more specifically, how substitutable the item is for other things in the economy.
Minerals are not great substitutes for lots of things in the economy: they can’t be eaten, they can’t be used as fuel, they can’t substitute for labor etc. As they become more abundant, I expect the income effect and substitution effect will roughly cancel out, causing their contribution to GDP to neither rise nor fall by much.
I asked GPT-4 to demonstrate the precise effect on total utility of a good falling in price by a factor of a million in two separate toy models, if you’re interested (the conversation is here). But ultimately I agree my language was a bit sloppy and you’re right to point out that there are worlds in which the claim I made is technically not true. (I guess I should have stated it less confidently too.)