At the time we did most of the writing for AI 2027, 2027 was my median. It really was a very representative scenario for me, and the whole thing was my project, my baby, so we went with what I thought instead of what some of the other authors thought.
So when people publish extremely short timeline estimates, especially when those estimates are unendorsed "modal" years, it feels like we are playing the "guess the number out of ten" game, our opponents picked "17", and we countered with "1".
I am aware of this dynamic. If I was playing political games optimally, I would have picked a year for the scenario that is short enough to be shorter than what most people expect, but otherwise as long as possible -- maybe 2033? and wrote the scenario to have AGI and takeoff happen in that year. Then, when things predictably went faster than my scenario, I could keep milking that saying "I was already concerned, but now unfortunately things are going even faster than I, the concerned expert, expected." This does seem like a politically better strategy to take, and it's one that I did briefly consider at the time.
But the scenario I wrote would have been trash; it wouldn't have represented what I actually expected and it wouldn't have been as close to what actually happened as AI 2027 in fact was.
I also could have, like, tried to conceal the fact that AI 2027 was actually what I expected. I could have said for example "This is just a scenario, it's not what I actually expect, don't worry I'm not crazy." That way if 2028 comes and AI still seems normal and not recursively self-improving, I can shield myself better from criticism of the form "you cried wolf."
I considered that strategy too and rejected it. I thought it would be better to just say what I actually thought.
I think I made the right choices. I think too many people too often try to play political games, and instead should just make quantitative, falsifiable forecasts. Yes, this makes it easier for people to criticize them if and when the forecasts fail to occur. That's (part of) the point. Society would be better off if this was standard practice, if this was the norm, if this was encouraged.
At the time we did most of the writing for AI 2027, 2027 was my median. It really was a very representative scenario for me, and the whole thing was my project, my baby, so we went with what I thought instead of what some of the other authors thought.
Alright, fair enough.
I am aware of this dynamic. If I was playing political games optimally, I would have picked a year for the scenario that is short enough to be shorter than what most people expect, but otherwise as long as possible -- maybe 2033? and wrote the scenario to have AGI and takeoff happen in that year. Then, when things predictably went faster than my scenario, I could keep milking that saying "I was already concerned, but now unfortunately things are going even faster than I, the concerned expert, expected." This does seem like a politically better strategy to take, and it's one that I did briefly consider at the time.
Was not really what most annoyed me about AI 2027 just FYI. I agree a 2033 timeline would have been ridiculous, what annoyed me originally was that the timeline was faster than the median of the researchers involved. If the reason it was 2027 anyways was because it was your project, then that makes sense.
Thanks. Yeah I had the idea to write it while still at OpenAI, and gathered a team to help me do it (Eli & Thomas were the first). They had somewhat longer timelines but I was the boss and insisted on telling the story I thought was most plausible; I told them we would do other scenarios that represented their views later -- and indeed, AI 2040 has a somewhat slower timeline in which the Plan D race trajectory takes off in 2030, and that's because it was Thomas's baby not mine. We also had plans to do an AI 2035 story representing Eli's timelines, but then Eli's timelines got shorter and now they are shorter still...
Here's a graph of our timelines-over-time:
Median AGI (TED-AI) forecasts by AI 2027 authors over time, updated from the August 2026 version in our blog post. The Nov 2026 points are the TED-AI medians from "Overall forecasts" above, after METR's estimate of ~1.7x AI-attributable acceleration: Daniel Sep 2027 (from late 2028), Eli Sep 2028 (from mid-2032), Brendan Apr 2028 (from early 2031), Thomas Apr 2028 (from early 2032), Romeo Sep 2028 (from early 2035).
It seems like your conclusion in the title operates from a theory-of-mind in which outsiders are basically uninformed good Bayesians. I've kind of thrown out all such models of public opinion dynamics recently.
I don't hold that much credence in [singularity didn't happen in 2027, people in 2028 sneer], as opposed to say, [generally rising panic as information percolates through society and the world catches up to LW with a 2-year delay, while barely tracking the facts on the ground], which is closer to what I'm observing in normie circles.
Imo this would be way more convincing if the community didn't just catastrophically fuck up by pretty much everyone having overly long timelines, including the AI 2027 folks. Many people, myself included, made research, career, and life decisions that intuitively made sense under "broad timelines" without interrogating whether decisions that were "reasonable" under 3 to 10 year timelines made any sense conditional on 2027 PONR.
This post reads to me like a biosecurity person in March 1st of 2020 complaining about how some people in the community think COVID has a 10% chance of being a pandemic and those are the loudest voices who got amplified, even though the median is like 3%, and actually we should worry about pandemics if COVID has a 0.1%, or indeed 0% since other pandemics are possible.
I mean I agree in some technical sense that your points are true but also this is also far from the biggest or most decision-relevant mistake here!
Yeah. It seems like a mistake strategically, and from the public-communications perspective, I'd say it's downright deontologically irresponsible. If you're taking on the mantle of keeping the public informed, if you want to earn their trust and deserve that trust, you have a duty to represent the danger accurately, instead of taking any opportunity to elevate it in their mind. Yet, I think people are being really epistemically sloppy on the topic, in a way that biases towards the latter. They can and should do better.
Edit: Also, is it just me, or cute dark gallows humor like this thread no longer feels appropriate? Like, those jokes were fine before the perceived situation got this stark, ha ha we may not see the next Christmas, "'how much do I have to live?' / 'five' / 'five years?' / 'four...'". Hilarious, I did laugh. But now that this is an actual position some serious people espouse, I don't find these funny at all, they're just contributing to a sense of panic in an unhelpful manner.
(Not to police how people in general are coping with the situation, but from high-profile public AI-risk speakers specifically this seems inappropriate.)
It is, at the very least, tonally off. To use one of our stock metaphors, imagine the pilot of a falling plane making similar rhetorically cute setup-into-payoff jokes to passengers over the speakers, with the setup creating a momentary sense of hope that the payoff then dashes. "Good news: there's a good chance that we survive... the next ten minutes!"
while throwing shade at people who are vaguely saying "long timelines"
I don't think that's it at all? I think it's just "haha the really short timelines to humanity dying are late-2027 now".
I feel like you just disagree with them? Seems extreme to say they're deontologically irresponsible for stating what they believe is accurate (aside from the post's point about AI 2027 being misleading wrt their timelines, but afaik they've learned from that mistake)
I think everyone's better off just stating (or linking to) actual criticisms and their own differences of opinion. Nate seems like a pretty good example here fwiw

What I'm getting at is it can be unhealthy to view people as part of your "faction" since it leads to a desire to control what they say. Maybe I'm naive but my vision is closer to a flourishing market of ideas.
Like, those jokes were fine before the perceived situation got this stark
It was fine to joke about it when it was likely that ASI would kill everyone within our lifetimes, just not in the next couple years? I'm being purposely obtuse but to an to outsider that probably looks like a odd delineation.
I feel like you just disagree with them?
Yes, and I'm irritated precisely because I don't think it's a factional values conflict at play here. I think they're making an epistemic-rationality mistake by not having sufficient over-model uncertainty, in a way they're more than smart enough not to do. Then, downstream of that mistake, they're disastrously over-committing to short-timeline worlds, in a way that'd likely compromise our chances in other worlds (per the original post here).
What I'm getting at is it can be unhealthy to view people as part of your "faction" since it leads to a desire to control what they say.
Mm, you might be somewhat right here, I may be making a this-shaped mistake too somewhat.
It was fine to joke about it when it was likely that ASI would kill everyone within our lifetimes, just not in the next couple years?
Yes. There's such a thing as a joke about a painful matter being "too soon", because the emotional pain is still too raw. Same logic here: it's too late for those now, the matter is too near-term and too emotionally (instead of only intellectually) real.
I think they're making an epistemic-rationality mistake by not having sufficient over-model uncertainty
I think I basically agree: RSI is much more unknown relative to progress driven by humans and compute growth where you can check your model against the past.
I think the linked post is probably wrong as a conclusion from METR's report, even though the statements about the graph are true. Basically the report itself describes the slope change as primarily due to non-AI factors at the start, even though it's ~80% AI by the end. Technically this could be a one-time jump in AI acceleration somewhere like Mythos, but I doubt it. They said "due to non-AI acceleratory factors in 2025, and due to the growing influence of AI over time."
If true the graph is super misleading (I totally didn't realize any of this until rereading the report from a different perspective), which is somewhat METR's fault but mostly downstream of how few bits of information they were allowed to reveal. Also quite weird that the non-AI factor would fairly suddenly appear and then decrease in unison with AI acceleration increasing.
See the latter part of my post here: https://www.lesswrong.com/posts/Gsj36znHDKAihs9qG?commentId=SabuamSLsF7Kztr3Y
If we make it to 2030, my expectation is that 1) a pause informed by short timeline predictions will have been a relevant causal factor in it and 2) we'll have faced at least one major AI-enabled catastrophe that catapulted AI safety into the center of public perception and debate. So even if ASI doesn't happen by 2028, which still seems like a plausible possibility, I think there's a decent chance no faction associated with AI safety will lose respect anyway, as the title suggests, due to other events happening by then. I'd be very surprised if we continue to see a more or less smooth sailing where large parts of daily life still seem pretty normal for most people for another two years.
Last year (by which I mean, [checks watch] "a few months ago"), I had pretty similar complaints. I wasn't very involved with the naming of AI 2027 (I maybe argued with habryka about it a bit). And I did argue a lot for naming "AI 2040: Plan A" something else, in part to alleviate this problem. (ultimately the team didn't find that good alternative name and settled for the site saying "PLAN A" in big letters and "AI 2040" in small letters).
I still definitely think it was at least an intense gamble, and making a choice that indeed spent a shared resource. And I've heard complaints from people who get asked "Wait, so you're one of those AI Safety people that all think AI is coming in 2027?")
But, uh, I currently do just bet that even if AI hasn't fully taken off by 2028, the public will broadly think "well, shit, AI sure did something in 2027."
But I can’t help but feel personally annoyed that the attention economy is structured such that all of the people that state aggressive timelines end up representing the clique.
I don't think this is due to the structure of the attention economy, rather it is a product of the fact that the short timelines people made a website with a big marketing campaign, while the people with longer timelines made some tweets and LessWrong posts. If the "AI 2040" domain had been bought by you instead of the group persuading people "AGI will happen in 2027 unless we stop it" then perhaps the clique would be represented differently.
I agree that if AI hits a wall really soon, some people in 2029 will sneer at the AI safety community for worrying over nothing, but I think the case for short timelines is so strong that it would be genuinely shocking if we didn't have extremely powerful AI in 2028. I really would lower my p(doom) in that case.
I'm not sure the overall public would be less worried about AI, though, since that 2028 is still likely a place with powerful and widespread AI. Unless we make a lot of alignment progress, there will be plenty of events to make people worried about AI.
I think I have like three models:
For the middle one, I feel a lot of hope for just showing up and arguing with people and being like "Here are concrete accurate predictions from AI 2027. Please give me some concrete inaccurate predictions." Perhaps you can also quote whatever Gary Marcus was saying at the time as a point of comparison.
Perhaps you think that everyone is in the middle camp, and the other two are small. Personally I think the first one is much bigger than the other two. I will admit that "If Anyone Builds It, Everyone Dies" is a title that will hold up better as it is more robustly true.
The very dumb one: people have still been listening to Gary Marcus while he's been getting everything ass-over-tea-kettle backwards. The reason? AI is a big deal and he is one of the people talking about it.
Not sure how true this is, because I have no real model of whether/why Gary Marcus has gotten increased read. Maybe that could describe Ed Zitron? But there it seems like the reason they have been getting "bigger" over time is because the demand for cope is increasing over time.
For the middle one, I feel a lot of hope for just showing up and arguing with people and being like "Here are concrete accurate predictions from AI 2027. Please give me some concrete inaccurate predictions."
I have extremely little hope that this would work.
For the middle one, I feel a lot of hope for just showing up and arguing with people and being like "Here are concrete accurate predictions from AI 2027. Please give me some concrete inaccurate predictions."
I have extremely little hope that this would work.
Well I think you should be more hopeful! The world has swung massively in the direction of respecting the kinds of conversations we have been having on LessWrong. Obama talking about the dangers of RSI, Nate Soares winning debates watched by millions of people, the New York Times editorial board advocating for an agreement with China for a slowdown, to list just three. LessWrong-style arguments are often winning the day!
The middle one seems closer to: people didn't read the report, got a sense that it was just a single prediction of how crazy a year will be regarding AI due to the title 15 seconds of commentary they heard about it, the year will come, people will think, "wow this AI stuff is crazy...as written!"
To use an ML analogy, I don't think the embedding/place in vector space people have for AI 2027 is predictions for when AI kills everyone
I think people have a fairly good understanding that timelines are uncertain. Even with things like "when will this game/movie come out", (where it's more of a promise than a prediction) when the original date passes people don't jump to it being totally fake and never happening.
Obviously you lose some credibility but what matters is if it feels closer than before, in which case people saying it was very far away might lose more credibility.
Heavily marketed it as the work of the smartest, most rational, most well-informed people that we had to offer.
Oh come on. Nowhere did I say I was the most rational, most informed, smartest person in the community. You are exaggerating.
Probably not worth getting into.
As anybody who's interacted with the AI risk case will realize, timelines are (at least in theory) mostly orthogonal with your beliefs on AI safety.
Your general point is well-taken. But there are some reasons to think that timelines would be somewhat correlated with AI safety, at least in principle, because time can be spent improving safety. The most likely reason for very long timelines is that we would have paused in order to do a lot of safety research. If it took one thousand years to get to ASI, for example, it seems most likely that we would have spent almost all of that time trying to get it right. Conversely, if we pushed the button tomorrow, it's more likely to go wrong because we rushed it.
group of people and I are working on a piece to address exactly this with hopefully some useful actionable insights, totally agree with you here
I happen to think the aggressive timelines are accurate, but the degree to which the timeline discussions in this community mirror the "Jesus is coming back on this date" ravings in certain Christian bubbles is increasingly stark. And in both cases it comes from insularity - we think we know something other people don't, and get so far forward in discussing it among ourselves that we get fixated on timeline details. Note that the Christian communities which avoid this best are those focused on evangelizing gospel basics. AI communities do not evangelize AI potential, good and bad, with anything like the regularity or fluency that the best missionaries have. We tend to assume people should be on our level, we let the basics get covered for the broad populace by all the people trying to make money from AI, and keep ourselves focused on the parts that are most interesting, getting sucked into ever more refined arguments that look from the outside like trying to date the machine rapture.
If you ignore timeline details, however, your theory becomes extremely difficult to falsify. Timeline details are something you can provably get wrong (or right). And messing up the timeline details does make your side lose credibility, but that's not a pesky detail that gets in the way of convincing people, it's a warning sign that your side is doing something wrong.
I am pretty confused about what arguments you are actually making here.
This comment feels like it's making some blanket aspersions without really arguing what the problem is or why it's bad, and then I also don't get why you believe the last couple sentences have much to do with the way we're relating to timelines.
By any reasonable definition I am an AI doomer. It has nevertheless been historically very strange for me to see how closely associated "AI doomers" and short timelines have been in the popular imagination. As anybody who's interacted with the AI risk case will realize, timelines are (at least in theory) mostly orthogonal with your beliefs on AI safety. Yes, timelines have some impact on what kinds of interventions are going to be viable, but among people who "believe in AGI" the variance in their optimism about the problem is more immediately dependent on underlying technical questions than anything else. Before GPT-2 or AlphaGo, there was still a much smaller, nascent "AI safety community", and these people were certainly not basing their worries on some data about AI uplift that we only got in 2026.
However, for years people on general social media have taken it on faith that the "AI safety position" and "short timelines position" were the same, and even argued against short timelines as if in doing so they were arguing against "doomers." I actually don't think the AI safety people did much to promote this association in the beginning. Instead people would just assume that, if you were worried about AI, you must be worried about it happening in the next ten years, and therefore you must think LLMs scale to superintelligence all that there.
Personally, I don't have any inside view on how long it will take to get to superintelligence. I see remarkable progress, but I don't have a model for how optimizing verifiable metrics like pretraining loss perplexity or whatever will actually cash out into getting AIs to generalize much further, so I can't say whether current systems are going to get us to superintelligence, even if they can solve hard math problems. When I have encountered anti-short-timelines-conflated-with-AI-safety-people, I have historically responded[1] to them by explaining that we will eventually develop superintelligence, even if it takes twenty years, and that most of the theoretical reasons you should be worried about AI alignment will probably still apply. I then make the separate and distinct observation that skeptics of all stripes have essentially gotten the pace of AI progress wildly wrong, and so if your skepticism of AI risk is just that powerful AI will not happen soon, that seems really tenuous.
In the last few years, though, it sure has gotten a lot harder to make that argument, because a group of people who represent the AI safety community:
It was quite annoying to have to explain this last bullet to upwards of a dozen friends and relatives throughout the last two years, whom I do not think did anything wrong in getting a different impression. It was even more annoying to see the METR report come out, and almost everybody start to move toward ludicrously fast 6-12 month timelines based off of evidence that they're not even able to read.
I suppose it's possible that these people are right, and if they're right then I salute them for their bravery and intelligence. But I can't help but feel personally annoyed that the attention economy is structured such that all of the people that state aggressive timelines end up representing the clique. The only AI predictions people remember, on both sides, are the extreme Ege Erdil/Daniel Kokotajlo-type predictions that are surprising and/or turn out extremely wrong. Mostly people just remember the year that was quoted and not the probability you attached to it. And the people we most disagree with, in general, tend to have ludicrously long timelines, extending out decades. So when people publish extremely short timeline estimates, especially when those estimates are unendorsed "modal" years, it feels like we are playing the "guess the number out of ten" game, our opponents picked "17", and we countered with "1".
In real life, not on Twitter.