I only know two outside views that result in projections of time to AI, which are (1) looking at AI as a specific technology like Kurzweil or Nagy, and (2) AI as a civilizational transition like Hanson. Of these, I’ve recently decided I prefer Hanson’s approach because it seems to match the scale of the transition better, and it seems misguided to view machines that beat humans at all capabilities as one specific technology anyway as opposed to a confluence of hardware and various software things (or brain scanning if you’re an ems fan).
A problem with Kurzweil’s approach is that modern computer growth trends only started—according to William Nordhaus anyway—around 1940. So who is to say this thing could not just shut off? Indeed, if you take it as one of the performance curves analyzed by Nagy, then maybe as an unusually long-lasting trend it has a higher chance of stopping than if you naively said, “Moore’s Law has lasted 70 years, so there’s a 50⁄50 chance of it lasting 70 more years.”
Ideally a pretty good way to go, making use of your suggestion of model combination, might be to let Hanson’s reasoning about GDP doubling times form 3⁄4 of your probability distribution, and then somehow splice in 1⁄4 based on specific trends happening now—although not just hardware. But that last requirement means I in practice don’t know how to do this.
If anybody thinks they have done better I’d really like to hear about it.
I only know two outside views that result in projections of time to AI, which are (1) looking at AI as a specific technology like Kurzweil or Nagy, and (2) AI as a civilizational transition like Hanson. Of these, I’ve recently decided I prefer Hanson’s approach because it seems to match the scale of the transition better, and it seems misguided to view machines that beat humans at all capabilities as one specific technology anyway as opposed to a confluence of hardware and various software things (or brain scanning if you’re an ems fan).
A problem with Kurzweil’s approach is that modern computer growth trends only started—according to William Nordhaus anyway—around 1940. So who is to say this thing could not just shut off? Indeed, if you take it as one of the performance curves analyzed by Nagy, then maybe as an unusually long-lasting trend it has a higher chance of stopping than if you naively said, “Moore’s Law has lasted 70 years, so there’s a 50⁄50 chance of it lasting 70 more years.”
Ideally a pretty good way to go, making use of your suggestion of model combination, might be to let Hanson’s reasoning about GDP doubling times form 3⁄4 of your probability distribution, and then somehow splice in 1⁄4 based on specific trends happening now—although not just hardware. But that last requirement means I in practice don’t know how to do this.
If anybody thinks they have done better I’d really like to hear about it.