Summary, for those who don't want to watch (even at the 2x speed I did):
- GPT3 and large language models seeming to generate text with deep understanding is fundamentally different than the past. Aka foundation models.
- Most AGI prediction dates have been pulled closer to the present.
- Brynjolfsson et al. assessed ML capability vs current tasks. In almost every occupation, ML could do some of the tasks better than humans. In none could they do all of the tasks.
- Q: Are we going to need to see regulation and law updated when jobs transition from human to machine, and there's no human to hold responsible for task failure? A: Law is lagging practice. Human in the loop is the near future, for both regulatory and practical purposes. Augment rather than replace.
- Q: What about humanoid robots? Specifically re: +10-20 years labor needs vs demographics? A: Controlled settings (e.g. factories) are the best first environments. General purpose environments are much more difficult. We live in constructed environments (e.g. street signs). Likely to see more of this to specifically support automated task execution.
- "Moravec's paradox is the observation by artificial intelligence and robotics researchers that, contrary to traditional assumptions, reasoning requires very little computation, but sensorimotor and perception skills require enormous computational resources." https://en.m.wikipedia.org/wiki/Moravec%27s_paradox
- Q: On productivity paradox, why aren't we seeing measured productivity increases as a result of ML deployment? A: Part is a measuring problem (e.g. zero price goods like Wikipedia don't show up in GDP).
Potential explanations for the missing productivity that he rules out: (1) Mismeasurement: We've always mismeasured things though. See: "consumer surplus." (2) Many technologies are shifting the pie vs making it bigger (e.g. targeted advertising). This doesn't calculate up to missing productivity.
A good explanation: economic equivalent of Amdahl's law. Speed up a portion of the process, yet the metaprocess still only runs at the speed of its slowest step. There is a fair amount of linkage in the economy, and we're not likely to see huge productivity boosts until ML is properly digested and problem tasks are reengineered to leverage the new capabilities. Which happens on the order of decades (aka as managers die / retire).
- Q: What kinds of transformations in the workplace need to happen to realize the full value? A: Look at Amazon's rearchitecting of the "retail commerce" process, and how long that took. Example: A/B tests on automation-augmented call-center agents (via recommendations in ear) were ~20% more productive.
- Q: Talk about the J-curve. A: The idea that investing and adopting new technologies decreases measured productivity (because its impacting unmeasured intangibles), before it begins to demonstrate growth. These have appeared historically (steam engines, early computers). The more impactful the technology, the strong the J curve effect is. Measured in years and decades.
- Q: How can you more accurately measure the intangibles that cause the J curve? A: One example is in acquisitions vs stock market value, as you're acquiring a company's assets + its intangible assets...
(Stopped at 01:08:00, as flight's about to leave)
IMHO, meh. At least for most of the crowd here.