Possible erratum: the Metaculus graph is misplaced. I believe it's meant to be right after the colon in "On the the forecasting platform Metaculus, the predicted date for AGI’s development has dropped by over two decades since 2022:"
There is another way to look at Moravec's paradox. Economists refers to tasks that can’t be easily automated as "weak links". A job is a bundle of tasks, and AI can automate only some of them. A classic example is the radiologist. Geoffrey Hinton predicted that AI reading scans would make radiologists obsolete. However, the numbers actually grew. This happened because the rest of what radiologists do - like talking to patients and working with colleagues - is the part AI cannot easily replace.
> The public will react negatively to AI and constrain its development through regulation
> AI is too unaligned and is the source of a lot of risk. Businesses and consumers become iffy about adoption, insurance with high premiums becomes required, and it becomes heavily regulated
> The supply chain for chips is heavily concentrated, and at least one major piece, TMSC, is in a very geopolitical vulnerable spot
> It will take longer for the economy to allocate a sufficient amount of resources to achieve enough comptue for an intelligence explosion
> Even if we got an intelligence explosion there could still be other bottlenecks, primarily resources and physical capital
great summary, Sarah! Reading this article a year after its publication, and it looks like the case for short timelines is getting stronger and stronger.
Possible erratum: the Metaculus graph is misplaced. I believe it's meant to be right after the colon in "On the the forecasting platform Metaculus, the predicted date for AGI’s development has dropped by over two decades since 2022:"
awesome summary!
There is another way to look at Moravec's paradox. Economists refers to tasks that can’t be easily automated as "weak links". A job is a bundle of tasks, and AI can automate only some of them. A classic example is the radiologist. Geoffrey Hinton predicted that AI reading scans would make radiologists obsolete. However, the numbers actually grew. This happened because the rest of what radiologists do - like talking to patients and working with colleagues - is the part AI cannot easily replace.
Some other arguments for longer timelines:
> The public will react negatively to AI and constrain its development through regulation
> AI is too unaligned and is the source of a lot of risk. Businesses and consumers become iffy about adoption, insurance with high premiums becomes required, and it becomes heavily regulated
> The supply chain for chips is heavily concentrated, and at least one major piece, TMSC, is in a very geopolitical vulnerable spot
> It will take longer for the economy to allocate a sufficient amount of resources to achieve enough comptue for an intelligence explosion
> Even if we got an intelligence explosion there could still be other bottlenecks, primarily resources and physical capital
(I don't necessarily endorse these.)
great summary, Sarah! Reading this article a year after its publication, and it looks like the case for short timelines is getting stronger and stronger.