Sign up to access an inspiration library of 30,000+ posts
Andreas Haupt's Recent LinkedIn Posts
Andreas Haupt
@indraos
Human-Centered AI Postdoctoral Fellow @ Stanford University | Economics and Computers
Posts
Andreas Haupt
Tech & AI
2mo
Overwhelmingly, we work on replacing humans.
Sometimes, we talk about augmenting human individuals.
Too rarely, we think about redesigning institutions using AI.
Erik Brynjolfsson's Turing Trap names the problem: when we measure AI progress by how well it imitates humans, we optimize for replacement. The benchmark is "can the machine do what the person does?". The economic consequence is substitution. Augmentation research, e.g., Erik's and my centaur evaluations work, has so far taken a lot of focus on the individual worker: Can a human, paired with this AI, do this task better?
But the unit of analysis we care about is often not the individual. The actual cost of scaling an organization is not the work anyone does — it's everything that happens so people can do work together. The weekly syncs, the status updates, the context-switching, the entire professional class whose job is making sure everyone is pointing in the same direction.
It might be an uncommon mindspace to think about AI’s impact not replacing individual contributors, but changing how we’re managed. Every coordination technology in history (writing, bureaucracy, the corporation, software platforms) created organizational forms that were previously impossible. We are sitting on the next one and using it to autocomplete emails in existing organizations.
This is what Prashaant Ranganathan and I have been prototyping. We gave an AI agent the role of project coordinator for a real team with a real deadline. We used humans and some of the infrastructure to reach them (TaskRabbit, Amazon, Instacart). What we found surprised us.
A manifesto, learnings from a first prototype (Giuseppe), and a chance to sign up for future prototypes in comments.
111
Andreas Haupt
Tech & AI
2mo
Overwhelmingly, we work on replacing humans.
Sometimes, we talk about augmenting human individuals.
Too rarely, we think about redesigning institutions using AI.
Erik Brynjolfsson's Turing Trap names the problem: when we measure AI progress by how well it imitates humans, we optimize for replacement. The benchmark is "can the machine do what the person does?". The economic consequence is substitution. Augmentation research, e.g., Erik's and my centaur evaluations work, has so far taken a lot of focus on the individual worker: Can a human, paired with this AI, do this task better?
But the unit of analysis we care about is often not the individual. The actual cost of scaling an organization is not the work anyone does — it's everything that happens so people can do work together. The weekly syncs, the status updates, the context-switching, the entire professional class whose job is making sure everyone is pointing in the same direction.
It might be an uncommon mindspace to think about AI’s impact not replacing individual contributors, but changing how we’re managed. Every coordination technology in history (writing, bureaucracy, the corporation, software platforms) created organizational forms that were previously impossible. We are sitting on the next one and using it to autocomplete emails in existing organizations.
This is what Prashaant Ranganathan and I have been prototyping. We gave an AI agent the role of project coordinator for a real team with a real deadline. We used humans and some of the infrastructure to reach them (TaskRabbit, Amazon, Instacart). What we found surprised us.
A manifesto, learnings from a first prototype (Giuseppe), and a chance to sign up for future prototypes in comments.
111
Andreas Haupt
Tech & AI
2mo
Overwhelmingly, we work on replacing humans.
Sometimes, we talk about augmenting human individuals.
Too rarely, we think about redesigning institutions using AI.
Erik Brynjolfsson's Turing Trap names the problem: when we measure AI progress by how well it imitates humans, we optimize for replacement. The benchmark is "can the machine do what the person does?". The economic consequence is substitution. Augmentation research, e.g., Erik's and my centaur evaluations work, has so far taken a lot of focus on the individual worker: Can a human, paired with this AI, do this task better?
But the unit of analysis we care about is often not the individual. The actual cost of scaling an organization is not the work anyone does — it's everything that happens so people can do work together. The weekly syncs, the status updates, the context-switching, the entire professional class whose job is making sure everyone is pointing in the same direction.
It might be an uncommon mindspace to think about AI’s impact not replacing individual contributors, but changing how we’re managed. Every coordination technology in history (writing, bureaucracy, the corporation, software platforms) created organizational forms that were previously impossible. We are sitting on the next one and using it to autocomplete emails in existing organizations.
This is what Prashaant Ranganathan and I have been prototyping. We gave an AI agent the role of project coordinator for a real team with a real deadline. We used humans and some of the infrastructure to reach them (TaskRabbit, Amazon, Instacart). What we found surprised us.
A manifesto, learnings from a first prototype (Giuseppe), and a chance to sign up for future prototypes in comments.
111