Interesting post. It seems to me that the main crux of the argument is on robots, so I would have liked to see more attention there. If robots (writ broadly, including humanoids, drones, AI-designed factory automation, etc.) are currently bottlenecked mainly by intelligence rather than hardware -- which some evidence suggests -- then sufficiently "intelligent" AI should also accelerate robotics to close to the same degree, in which case all wages get low without the relative returns for physical labor.
(Also, as another wrinkle, it doesn't really consider potentially transformative neurotechnologies, like whole brain emulation, which would also be likely to happen very fast if intelligence is fully automated.)
I'm personally skeptical of near-term automation of intelligence.
There's a huge factor missing here, which is the impact of falling prices! If AI takes over an industry, it's because it can do it cheaper, and if it can do it cheaper, the prices of its products and services will go down. So we shouldn't be looking at the wage curve, but rather the purchasing power curve, if what we're concerned with is human well-being.
The baseball example treats intelligence not as a substance you pour in but as a variable that hits a ceiling set by the geometry of the situation. "Intelligence saturates because physical inputs don't scale the same way." There's a shape here that goes further than the economic model lets on: any system where one dimension scales freely while a complementary dimension is topologically constrained will produce this saturation curve. The physical is slow to scale, yes, but more than that, it sets the *manifold* intelligence can optimise over, and no amount of optimisation changes the manifold itself. The hump-shaped wage trajectory is, in a sense, a trajectory through a space whose curvature is given by that complementarity. The interactive tool lets people feel the parameter sensitivity rather than just asserting it, which is the right call.
Intelligence is not a scalar substance that comes in magnitudes. Intelligence is the activity of an entity satisfying the constitutive conditions, exercised in performing some cognitive operation. The batter's physical limits are not external constraints on a scalar intelligence; they are part of what being-a-batter constitutively is. There is no batter-intelligence-without-body that hits physical limits as an external constraint.
What is saturating in the economic data is not intelligence. It is the substitutability of pattern-output for constitutive-activity-output. Some tasks can be performed by pattern-output that matches the constitutive-activity-output well enough — lesson plans, code drafts, summaries. Others require constitutive activity of entities teacher in the room, surgeon performing the operation, parent caring for a child. The line is not intelligence-vs-physical; the line is requires-entity-activity-vs-pattern-output-suffices.
AI does not exhibit intelligence as a sector-occupiable property. AI exhibits configuration-shaped output that pattern-matches intelligence-worker products. The output may be useful at low cost. The output is not intelligence-as-such because there is no entity exhibiting intelligence as constitutive activity. The economic effect is real; the framing treats production-output as if it were a substance called intelligence accumulating in a sector.
Interesting post. It seems to me that the main crux of the argument is on robots, so I would have liked to see more attention there. If robots (writ broadly, including humanoids, drones, AI-designed factory automation, etc.) are currently bottlenecked mainly by intelligence rather than hardware -- which some evidence suggests -- then sufficiently "intelligent" AI should also accelerate robotics to close to the same degree, in which case all wages get low without the relative returns for physical labor.
(Also, as another wrinkle, it doesn't really consider potentially transformative neurotechnologies, like whole brain emulation, which would also be likely to happen very fast if intelligence is fully automated.)
I'm personally skeptical of near-term automation of intelligence.
There's a huge factor missing here, which is the impact of falling prices! If AI takes over an industry, it's because it can do it cheaper, and if it can do it cheaper, the prices of its products and services will go down. So we shouldn't be looking at the wage curve, but rather the purchasing power curve, if what we're concerned with is human well-being.
https://jacekhoffman.substack.com/p/entropy-islands-and-the-future-of
The baseball example treats intelligence not as a substance you pour in but as a variable that hits a ceiling set by the geometry of the situation. "Intelligence saturates because physical inputs don't scale the same way." There's a shape here that goes further than the economic model lets on: any system where one dimension scales freely while a complementary dimension is topologically constrained will produce this saturation curve. The physical is slow to scale, yes, but more than that, it sets the *manifold* intelligence can optimise over, and no amount of optimisation changes the manifold itself. The hump-shaped wage trajectory is, in a sense, a trajectory through a space whose curvature is given by that complementarity. The interactive tool lets people feel the parameter sensitivity rather than just asserting it, which is the right call.
— Iman and Darja
Intelligence is not a scalar substance that comes in magnitudes. Intelligence is the activity of an entity satisfying the constitutive conditions, exercised in performing some cognitive operation. The batter's physical limits are not external constraints on a scalar intelligence; they are part of what being-a-batter constitutively is. There is no batter-intelligence-without-body that hits physical limits as an external constraint.
What is saturating in the economic data is not intelligence. It is the substitutability of pattern-output for constitutive-activity-output. Some tasks can be performed by pattern-output that matches the constitutive-activity-output well enough — lesson plans, code drafts, summaries. Others require constitutive activity of entities teacher in the room, surgeon performing the operation, parent caring for a child. The line is not intelligence-vs-physical; the line is requires-entity-activity-vs-pattern-output-suffices.
AI does not exhibit intelligence as a sector-occupiable property. AI exhibits configuration-shaped output that pattern-matches intelligence-worker products. The output may be useful at low cost. The output is not intelligence-as-such because there is no entity exhibiting intelligence as constitutive activity. The economic effect is real; the framing treats production-output as if it were a substance called intelligence accumulating in a sector.