I found this super helpful thank you, probably the best thing I’ve read about AI timelines in the last year actually. So so well communicated with small words and minimal jargon thank you!
I know you’re mainly taking about the best thinking approach here, but how does this translate to communication about AI timelines? Distributions make a lot of sense to me but are very hard for most people to think in. This wouldn’t be useful to communicate with for most of my friends, unless I maybe had an hour and a large napkin… I wonder if there is a way to communicate in a “distributy” like way with people who just aren’t statistically minded?
If some regular person asks me when i think the AI apocalypse is coming, what’s a good way to communicate? I don’t want to just guess a year for all the reasons you’ve stated, but a distribution won’t be understood either. In the past I’ve said something like ” I really don’t know but it could well be between 2030 and 2040″, but my impression has been this seems pathetically vague and unhelpful to most people. Any ideas on communicating AI timelines with integrity to non-statsy folks?
As a side note it seems strange that the guy who wrote the AI 2027 story’s 50 percent point is at about 2031ish? Why wasn’t the story then AI 2031?
How to communicate this is a good question, and I don’t yet know the best answer. I think admitting uncertainty is generally good — it is both honest and actually appreciated by many audiences. But there is still the question of how to do it. The scientific norm is usually to stay silent until the evidence for something (or some piece of the puzzle) is strong enough (e.g. reaching p = 0.05). I don’t think that is the right norm here. We are in a very high stake situation and policy-makers need the partial evidence that we do have. But communicating it is hard.
I think your expression is pretty good, and could be made a little better. e.g. “Leading AI forecasters can’t rule out it happening before 2030 and think it will probably happen before 2040.”
Re AI 2027, there is a good explanation of how their views have changed here.
Thanks Toby interesting one on the communication. For policy makers I think that communcation style can work OK, less so with my friends haha.
I’m still confused by why they picked 2027 even in 2025. Back when they made it, Daniel’s median forecast was 2028 and Eli’s 2031. Surely you then pick 2029 or 2030 for your scenario? Picking the “most likely year for it to happen” still feels a bit disingenous to me.
I’m still confused by why they picked 2027 even in 2025. Back when they made it, Daniel’s median forecast was 2028 and Eli’s 2031. Surely you then pick 2029 or 2030 for your scenario? Picking the “most likely year for it to happen” still feels a bit disingenous to me.
I’m not sure why picking the mode feels disingenuous to you, it feels fine to me as long as it’s between roughly 15th and 85th percentiles and you are transparent about it.
The causual history of why it was 2027 is that this was Daniel’s median when we started writing it, and it would have been a lot of work to rewrite our near-final draft to make it 2028 after Daniel changed his view. The reason it was based off of Daniel’s view and not other authors’ is that AI 2027 was ultimately supposed to represent his view rather than amalgam that include others’ views. Giving a single person final say seems better than design by committee. That said, we had few strong disagreements.
The other authors considered 2027 plausible enough and close enough to a modal scenario that they (including me) felt happy to help with the project.
Edit 2: It probably would have been reasonable for me to push for the timelines to be in between our views rather than Daniel’s. I didn’t really consider it because I thought 2027 was plausible enough and Daniel was leading the project. I think I also gave some weight to Linch’s point about it being important to communicate that things could get crazy very soon, but I’m not sure if this was cruxy. However, memetic fitness wasn’t an (explici)t consideration.
From what Daniel said I thought his median was 2028 when he started to write it? But that’s perhaps a bit nitpicky.
I think there might be a wider EA/Rationalist Comms issue here when communicating with the general public. Communicating projects like this isn’t just about whether it “feels” fine—I think its important to think about how it might come accross and future implications. To the general public, this scenario even in 2030 still feels mega-soon and sci-fi. The problem is if we go past 2027 now, many people will say “those tech-bro idiots they’re always wrong” and might miss the point o the thing
If anything I think here picking a more conservative, tail end of the timeline (2028-2030) would have been better, to keep it relevant for longer.
To be clear, I agree that we should make comms decisions based on what we think the effects will be, I wasn’t using “feel” intending to mean otherwise.
I think they were optimizing for a combination of concreteness (so there’s an exact story to point to, where the 2027 story is “things go roughly as they expected” whereas 2028 and 2031 were pricing in different types of individually unexpected delays[1]), and for memetic value.
Compare: My best estimate is that this project will take me 6 months. However, if you ask me to write it out step-by-step, it’d take me 4 months.The 6 months include buffers for various delays, some expected and some unexpected.
I think for project time estimations as part of a larger plan, the 6 month reply is more useful. But for someone following along on my thinking process, or a manager/collegue/direct report trying to help me optimize, the 4 month step-by-step report might be easier to follow along and/or more useful to critique or improve.
Isn’t then somewhere between 2028 and 2031 really “things go roughly as expected” and 2027 is “things go faster than expected if every AI improvement rolls out without roadblocks?” I feel like if you’re going to put something out there in the public sphere as a leader in AI, a bit of timeline conservatism might be prudent. Not the biggest deal though I suppose
“I feel like if you’re going to put something out there in the public sphere as a leader in AI, a bit of timeline conservatism might be prudent.”
I see and respect that position, but you can imagine someone saying the opposite: “I feel like if you’re going to put something out there in the public sphere as a leader in AI, it’s probably prudent to warn people of significant risks that happens much sooner than people expect, even if you think it’s less than 50% likely to happen then.”
Plausibly you can get away with reporting 3-5 numbers.
For 3 numbers, 25th percentile, median, 75th percentile. This is the approach (“interquartile range”) used for reporting SAT acceptance ranges in the US. So we have at least a prior example of a widely reported figure that people don’t think “normal people”/high-schoolers and their parents would have too much trouble understanding.
For 5 numbers, something like 5th percentile, 25th percentile, median, 75th percentile, 95th percentile.
3-5 numbers obviously harder to communicate than 1 number, and less precise than the full distribution. But hopefully it’s clear and useful enough to be good here.
I’m not sure I can get away with that? I would say for over 90% of people 3 numbers would add even more confusion than 2. The SAT example is encouraging, although Americans make up a small proportion of my friends and acquaintances.
I found this super helpful thank you, probably the best thing I’ve read about AI timelines in the last year actually. So so well communicated with small words and minimal jargon thank you!
I know you’re mainly taking about the best thinking approach here, but how does this translate to communication about AI timelines? Distributions make a lot of sense to me but are very hard for most people to think in. This wouldn’t be useful to communicate with for most of my friends, unless I maybe had an hour and a large napkin… I wonder if there is a way to communicate in a “distributy” like way with people who just aren’t statistically minded?
If some regular person asks me when i think the AI apocalypse is coming, what’s a good way to communicate? I don’t want to just guess a year for all the reasons you’ve stated, but a distribution won’t be understood either. In the past I’ve said something like ” I really don’t know but it could well be between 2030 and 2040″, but my impression has been this seems pathetically vague and unhelpful to most people. Any ideas on communicating AI timelines with integrity to non-statsy folks?
As a side note it seems strange that the guy who wrote the AI 2027 story’s 50 percent point is at about 2031ish? Why wasn’t the story then AI 2031?
Thanks Nick!
How to communicate this is a good question, and I don’t yet know the best answer. I think admitting uncertainty is generally good — it is both honest and actually appreciated by many audiences. But there is still the question of how to do it. The scientific norm is usually to stay silent until the evidence for something (or some piece of the puzzle) is strong enough (e.g. reaching p = 0.05). I don’t think that is the right norm here. We are in a very high stake situation and policy-makers need the partial evidence that we do have. But communicating it is hard.
I think your expression is pretty good, and could be made a little better. e.g. “Leading AI forecasters can’t rule out it happening before 2030 and think it will probably happen before 2040.”
Re AI 2027, there is a good explanation of how their views have changed here.
Thanks Toby interesting one on the communication. For policy makers I think that communcation style can work OK, less so with my friends haha.
I’m still confused by why they picked 2027 even in 2025. Back when they made it, Daniel’s median forecast was 2028 and Eli’s 2031. Surely you then pick 2029 or 2030 for your scenario? Picking the “most likely year for it to happen” still feels a bit disingenous to me.
I’m not sure why picking the mode feels disingenuous to you, it feels fine to me as long as it’s between roughly 15th and 85th percentiles and you are transparent about it.
The causual history of why it was 2027 is that this was Daniel’s median when we started writing it, and it would have been a lot of work to rewrite our near-final draft to make it 2028 after Daniel changed his view. The reason it was based off of Daniel’s view and not other authors’ is that AI 2027 was ultimately supposed to represent his view rather than amalgam that include others’ views. Giving a single person final say seems better than design by committee. That said, we had few strong disagreements.
The other authors considered 2027 plausible enough and close enough to a modal scenario that they (including me) felt happy to help with the project.
Edit: Daniel discusses his perspective here
Edit 2: It probably would have been reasonable for me to push for the timelines to be in between our views rather than Daniel’s. I didn’t really consider it because I thought 2027 was plausible enough and Daniel was leading the project. I think I also gave some weight to Linch’s point about it being important to communicate that things could get crazy very soon, but I’m not sure if this was cruxy. However, memetic fitness wasn’t an (explici)t consideration.
Thanks that’s helpful
From what Daniel said I thought his median was 2028 when he started to write it? But that’s perhaps a bit nitpicky.
I think there might be a wider EA/Rationalist Comms issue here when communicating with the general public. Communicating projects like this isn’t just about whether it “feels” fine—I think its important to think about how it might come accross and future implications. To the general public, this scenario even in 2030 still feels mega-soon and sci-fi. The problem is if we go past 2027 now, many people will say “those tech-bro idiots they’re always wrong” and might miss the point o the thing
If anything I think here picking a more conservative, tail end of the timeline (2028-2030) would have been better, to keep it relevant for longer.
I agree not the biggest deal though.
To be clear, I agree that we should make comms decisions based on what we think the effects will be, I wasn’t using “feel” intending to mean otherwise.
I think they were optimizing for a combination of concreteness (so there’s an exact story to point to, where the 2027 story is “things go roughly as they expected” whereas 2028 and 2031 were pricing in different types of individually unexpected delays[1]), and for memetic value.
Compare: My best estimate is that this project will take me 6 months. However, if you ask me to write it out step-by-step, it’d take me 4 months.The 6 months include buffers for various delays, some expected and some unexpected.
I think for project time estimations as part of a larger plan, the 6 month reply is more useful. But for someone following along on my thinking process, or a manager/collegue/direct report trying to help me optimize, the 4 month step-by-step report might be easier to follow along and/or more useful to critique or improve.
The concreteness is fine makes sense for sure
Isn’t then somewhere between 2028 and 2031 really “things go roughly as expected” and 2027 is “things go faster than expected if every AI improvement rolls out without roadblocks?” I feel like if you’re going to put something out there in the public sphere as a leader in AI, a bit of timeline conservatism might be prudent. Not the biggest deal though I suppose
“I feel like if you’re going to put something out there in the public sphere as a leader in AI, a bit of timeline conservatism might be prudent.”
I see and respect that position, but you can imagine someone saying the opposite: “I feel like if you’re going to put something out there in the public sphere as a leader in AI, it’s probably prudent to warn people of significant risks that happens much sooner than people expect, even if you think it’s less than 50% likely to happen then.”
Plausibly you can get away with reporting 3-5 numbers.
For 3 numbers, 25th percentile, median, 75th percentile. This is the approach (“interquartile range”) used for reporting SAT acceptance ranges in the US. So we have at least a prior example of a widely reported figure that people don’t think “normal people”/high-schoolers and their parents would have too much trouble understanding.
For 5 numbers, something like 5th percentile, 25th percentile, median, 75th percentile, 95th percentile.
3-5 numbers obviously harder to communicate than 1 number, and less precise than the full distribution. But hopefully it’s clear and useful enough to be good here.
I’m not sure I can get away with that? I would say for over 90% of people 3 numbers would add even more confusion than 2. The SAT example is encouraging, although Americans make up a small proportion of my friends and acquaintances.
Just the 25th and 75th percentile?
Yeah I tihnk that’s soemthing like the approach Toby and I were discussing!