notes

The general vibe is that progress will continue, unsure if this results in recursive self improvement or just continued progress, but progress will continue, and soon something qualitatively new will be here: the ability to automate many jobs end-to-end, or at least mostly end-to-end.

• Capital share of the economy vs labor share of the economy - how much money gets paid to people for doing labor vs how much money gets paid to people for owning / using capital (landlords, investors, loaners, etc)
  • Since the industrial revolution, very surprisingly, the labor share of the economy has remained in the 60% range - and only recently dropped to the mid fifties range.
  • Philip remarks on the remarkable stability of this share, but also notes that something qualitatively different is happening.
• David Ricardo, one of the most famous economists of the time, predicted that all jobs would be automated by the industrial revolution and that there would be mass unemployment. His first prediction was correct - almost all the jobs that existed at the time would be automated, but the second prediction was not. They note that it’s hard to make these predictions so far into the future, and that there is a level of uncertainty we have to be aware of.
• One class of jobs that humans will retain their advantage in is relational jobs: jobs that benefit from having humans in them. Performance, for example, ballerinas, masseuses, and so on. But this class of jobs may be wider than those: it is any job where having a human in the loop of the job increases consumer’s willingness to pay for the job: even if you can automate most of what a doctor does (filing insurance claims, ordering medicine), if consumers prefer that real humans sit with them and deliver the news, doctors may also be included in relational jobs.
• For the remaining non-relational jobs, the two outcomes I can draw from this are that 1) you get complete end-to-end automation, and 2) there remains some parts of a workflow that humans retain their comparative advantage in.
  • The second part speaks to the O-ring theory of automation. This theory was named after the Challenger rocket, which failed because just one part was malfunctioning. The theory says that automation of 9/10th’s of a workflow could make the remaining 1/10 of the workflow more valuable, because it becomes the bottleneck. This is part of why we still have high employment levels after the industrial revolution.
  • It remains to be seen if this phenomenon, or if we get full end-to-end automation. This is dependent on both the nature of the job and the timeline. Dwarkesh notes that this theory of automation could go the other way: it may be tenuous to let humans participate in the 1/10th of the workflow they supposedly have a comparative advantage in, because there may be transaction costs unaccounted for (for example, what if the 9/10ths of the job that is automated is fully done in uninterpretable neuralese or simply has too much context for the human to participate in?)
• They talk about the Moores law, which says that the cost of compute approximately halves every 18 months. There is a pessimistic interpretation of Moore’s Law from Philip, which goes, “the value of compute halves every 18 months”. But this may not be true, as while supply goes up, demand also contributes to pricing, and the demand for compute has gone up as people have found more uses for it: the cost per hour to rent an H100 has gone up even though we not have much more advanced technology.
• Fast or slow takeoff better for politics? The say that the worst case scenario is ‘slow drip’ - where there is enough change to cause job displacement but just over time, so the political will cannot be well-catalyzed
• Redistribution really dependent on political officials; they say that redistribution of capital may be worse than some equity-based redistributive scheme
• On the Citrini post: the conditions that they assume are hard to meet; they require very little consumption, as well as investment, from elites
• A twist on O-ring automation: Dwarkesh asks, even if humans are comparatively better at one part of the supply chain, the transaction costs of keeping humans in the loop may be too much - there may be transaction costs like converting neuralese to something human-interpretable and then something back (this seems like a very narrow band of possiblities? the transaction costs surely cannot be that high, and so humans would have to be just marginally better than the models for this to matter, and at that point we’re talking about very little marginal gains in production from being able to switch to the human at lower transaction costs)
• There is also interesting discussion on how much of the economy will be directed towards satisfying human vs AI preferences.