I recently attended a talk at Manifest by Chad Jones on the economic effects of AI. Much of it was sensible. Unlike many economists, he gives careful consideration to AI becoming pretty powerful soon. But his main scenarios predict much slower growth than I expect.
His paper Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion clarifies the parts of his talk that puzzled me. This post explores where our assumptions differ.
The fastest scenario that he considers (figure 6 – The Future if AI = ‘Moore’s Law Everywhere’) has economic growth rising to 13% by 2040. Whereas I expect at least 30% growth by then, due to automation happening earlier than he’s willing to imagine.
The key areas where I disagree with him are beliefs about the extent to which growth will be constrained by weak links, which likely stems from differing beliefs about how general-purpose AI will be.
Weak Links
Jones assumes a fixed set of tasks that are needed for production.
He uses a century of historical data to show a trend of exponential decay of not-yet-automated tasks. This pattern implies a long tail of tasks that are slow to automate. These tasks are the weak links that ensure that economic growth is constrained by the need for human labor to accomplish those tasks.
The automation rate (~2% per year) is how quickly tasks move from labor to capital; the elasticity (? = 0.2) is how much the tasks that haven’t moved yet bottleneck output — and Jones’s slow growth comes almost entirely from the second, not the first.
My response is that the elasticity ceases to matter once the weak links are no longer human labor, which is why I focus on the automation rate and the arrival of general-purpose robots rather than on ?.
AI has altered that long-term trend, due to crossing an important threshold of general-purpose learning about a decade ago (the shift from single-task agents to systems whose learning transfers across tasks). It has not yet scaled up enough to have much effect on the data that Jones is focused on. Maybe it would be visible if he looked carefully at recent data, but his data sources (Table 1) end in years ranging from 2017 to 2023.
One pattern leads me to dispute Jones’ expectation is the famous METR graph of increasing task lengths that an AI can complete. Many other AI benchmarks hint at roughly the same pattern of AI rapidly matching or beating humans on cognitive tasks. I claim it generalizes to physical tasks.
Jones gives examples of tasks that he imagines won’t be automated for several decades:
helping an elderly patient with dementia through a confused night, rewiring the electrical system in a renovated building, running a kindergarten classroom, negotiating a delicate business deal, or playing professional sports.
Whereas I predict that AI will exceed human productivity by about 2030 at whichever of those tasks are needed for economic growth.
Plugging my assumptions into Jones’ model should generate growth predictions that match my expectations.
Human Level Robotics
The key question is whether we’ll soon have robots that can do general-purpose labor as well as a typical human. I expect that to happen around the end of this decade. Jones’ slow calibration implies human-level general labor doesn’t arrive until the 2040s or later. It seems likely that this disagreement explains approximately all of our disagreement about growth rates.
AI has become sufficiently general-purpose that software is no longer the main bottleneck to replacing human labor with robots. Robotic software is still incomplete, but it has learning abilities that are comparable to humans, and it mainly needs scaling of compute and data. It may still be expensive to train a robot for many particular tasks, but that training is mostly a matter of throwing resources at the training, and not constrained by a need for new insights.
Robotic progress is currently constrained mainly by relatively ordinary hardware engineering tasks.
Some of you may object that dexterity looks hard. Software shows signs of overcoming those hardware limits. See ALOHA Unleashed: A Simple Recipe for Robot Dexterity and ?0: A Vision-Language-Action Flow Model for General Robot Control.
I expect that there will still be weak links after 2030 that limit growth, but they mostly won’t involve human labor. They will be more like the energy limits described in this LessWrong post.
There may still be areas where something prevents automation (dementia patient care? sports?), but they won’t be tasks that are needed for expanding Elon Musk’s empire.
Here are some Manifold markets with forecasts that seem somewhat closer to my model than to the Jones and Tonetti model: