It sounds like you are saying the public’s awareness is underwhelming compared to what it should be, but for purposes of grading AI 2027, we need to compare to what AI 2027 said it would be. AI 2027 said it would be far far less than it should be.
My point is not just that it’s far less than what it should be, but that your internal felt sense of AI seeming more salient compared to the metric is being selection-biased and you haven’t corrected for it in the way that you corrected AI capabilities benchmarks being selection biased towards estimating more progress in AI than reality, because your internal intuition/simulation is disproportionately sampling the sorts of people who would be excited/concerned/react in any way to something new in AI, at least right now.
Put another way, I trust the metric more than you, because I trust intuition/internal simulations less than you do in this domain.
Another point made by someone else on Twitter (I think Stefan Schubert?) -- public awareness is a metric relevant for grading AI 2027, but not relevant for grading AI 2027′s timelines predictions specifically. AI capabilities don’t go faster or slower depending on whether the public is aware of them. So we maybe should have just excluded this category entirely from our tracking, which would have resulted in shorter timelines from this estimation method.
This is entirely fair, TBH. If I were to track it, I’d grade it seperately, and make clear that this doesn’t feed into AI timeline forecasts at all.
I agree that knowing nothing else, people should trust the 1 metric we chose to forecast more than they trust our subjective impression of how things are going in the underlying variable that metric was a proxy for. Because subjective impressions are biased.
But if you independently have noticed a lot more people talking about AI, taking AI risk more seriously, etc. recently, and/or if you have other metrics in mind (like # of US congresspeople explicitly talking about catastrophic AI risks) that have been going up recently, then you don’t need to take my word for it.
My point is not just that it’s far less than what it should be, but that your internal felt sense of AI seeming more salient compared to the metric is being selection-biased and you haven’t corrected for it in the way that you corrected AI capabilities benchmarks being selection biased towards estimating more progress in AI than reality, because your internal intuition/simulation is disproportionately sampling the sorts of people who would be excited/concerned/react in any way to something new in AI, at least right now.
Put another way, I trust the metric more than you, because I trust intuition/internal simulations less than you do in this domain.
This is entirely fair, TBH. If I were to track it, I’d grade it seperately, and make clear that this doesn’t feed into AI timeline forecasts at all.
I agree that knowing nothing else, people should trust the 1 metric we chose to forecast more than they trust our subjective impression of how things are going in the underlying variable that metric was a proxy for. Because subjective impressions are biased.
But if you independently have noticed a lot more people talking about AI, taking AI risk more seriously, etc. recently, and/or if you have other metrics in mind (like # of US congresspeople explicitly talking about catastrophic AI risks) that have been going up recently, then you don’t need to take my word for it.