The piece is right that productivity studies measure the wrong layer, but the part it treats as the AI-era twist (value capture moving off the worker) is nothing new.
The divergence between productivity going up and who gets paid for it is as old as the industrial revolution. Codifying expertise, cutting the worker out of the customer relationship, value landing on a platform instead of a person. Whether it's the assembly line, the ATM, or Expedia, it has always happened as new tech enabled it.
The value never flowed to the raw tech provider, and I don't think that will change now. LLM architecture is probabilistic AND it's an approximation by design. It gets closer to reliable but never gets there, so anything that depends on being right needs an intermediary that wraps the model and supplies the trust. Reliable systems have always been built out of unreliable parts.
So it's the same disruption model as ever. Incumbents get first shot, usually miss, (for reasons unrelated to tech but rather how capital gets allocated) and many get replaced. The one constant is that the workers don't end up owning the value.
The only question imo is whether this version of AI meaningfully different in degree from previous waves. That's still TBD.
I feel like the insights about value migration and unit of work shifts are similarly insightful to The stack fallacy - gets to the fast assumptions we accidentally build in based on our own lived context.
I'm interested in whether we have an adequate theoretical basis for studying AI. Do you think there is such a basis and that it is known to some people but not others who write about AI; that there is such a basis and that nobody who writes about AI knows it yet; or that there is no such basis?
I don't think we have a full theoretical basis, to be honest.
The best we can do right now is draw better boundaries on where and to what extent we can leverage the current theory base. A lot of it was drafted for a different set of assumptions that no longer hold true.
This puts us in a place where we can draw boundaries on what does not apply but doesn't get us to a place where we have a full basis on claiming what does.
It's still a step in the right direction away from many of today's blind spots which get propagated simply on the basis of citation discipline.
The piece is right that productivity studies measure the wrong layer, but the part it treats as the AI-era twist (value capture moving off the worker) is nothing new.
The divergence between productivity going up and who gets paid for it is as old as the industrial revolution. Codifying expertise, cutting the worker out of the customer relationship, value landing on a platform instead of a person. Whether it's the assembly line, the ATM, or Expedia, it has always happened as new tech enabled it.
The value never flowed to the raw tech provider, and I don't think that will change now. LLM architecture is probabilistic AND it's an approximation by design. It gets closer to reliable but never gets there, so anything that depends on being right needs an intermediary that wraps the model and supplies the trust. Reliable systems have always been built out of unreliable parts.
So it's the same disruption model as ever. Incumbents get first shot, usually miss, (for reasons unrelated to tech but rather how capital gets allocated) and many get replaced. The one constant is that the workers don't end up owning the value.
The only question imo is whether this version of AI meaningfully different in degree from previous waves. That's still TBD.
I feel like the insights about value migration and unit of work shifts are similarly insightful to The stack fallacy - gets to the fast assumptions we accidentally build in based on our own lived context.
I'm interested in whether we have an adequate theoretical basis for studying AI. Do you think there is such a basis and that it is known to some people but not others who write about AI; that there is such a basis and that nobody who writes about AI knows it yet; or that there is no such basis?
I don't think we have a full theoretical basis, to be honest.
The best we can do right now is draw better boundaries on where and to what extent we can leverage the current theory base. A lot of it was drafted for a different set of assumptions that no longer hold true.
This puts us in a place where we can draw boundaries on what does not apply but doesn't get us to a place where we have a full basis on claiming what does.
It's still a step in the right direction away from many of today's blind spots which get propagated simply on the basis of citation discipline.