@(Glossy) I am skeptical that raising alarms about a full-scale Russian invasion of the Ukraine is a means to the end of attempting to reconquer the Donbass, especially if it's Washington that wants this to happen and Zelensky and co who are reticent. "Yo Zelensky" "Yes Washington?" "We want you to attack Donbass. Now is the time! Reclaim your lost territory!" "Why now of all times?" "Because the Russians are massing troops on your border! Perfect, right? They could invade any minute now!" "What?!?" Seriously, isn't launching an offensive into Donbass...

@EvanHarper Source please? How many nations did this, and how coordinated was it (was it all literally the same night? after not having given any similar warnings in the past?)

@jmason It would be helpful to know how often the Do Not Travel warning is followed by a lack of invasion. I wouldn't be surprised if it's "90%+ of the time." This would be a more useful piece of evidence than the anecdote about Azerbaijan.

God dammit guys we had a sweet gig going! Just keep voting 1% and we all make points!

(This is my protest against this resolution method.)

@tbyoln IMO, when the heads of the AGI labs start publicly calling for slowdowns and caution, that is evidence that AGI is very near, not evidence that it's far away.

@(tbyoln) Sometimes updates happen not because of events, but rather because of thinking through the arguments more carefully and forming better models. Even this kind of update, however, often happens around the same time as splashy events, because the splashy events cause people to revisit their timelines, discuss timelines more with each other, etc. (Speaking as someone who hasn't updated as much on recent events due to having already had short timelines, but who hadn't forecasted on this question for almost a year (EDIT: two years!) and then revisi...

In general I think people here massively overestimate how long it'll take to go from "weakly general" to "general" (I'm referring to this question and its more popular weaker variant.)

@(andreferretti) I'll give a different take than tbyoln: the people surveyed by AI Impacts haven't thought much about AGI timelines & have terrible takes as a result. Their job is to incrementally advance the current state of the art in the field of ML, not to forecast anything. And most of them barely think about AGI at all, much less seriously attempt to forecast it. If you read the surveys in detail, you'll notice a lot of inconsistencies in how survey respondents answer -- e.g. different phrasings of the same question lead to significantly different ...

I expect that if an AI that can pass the turing test exists by 2029, the turing test will never be run, nor will the Long Now foundation be around to announce the results. How should this influence my forecast? Would such a case resolve the question positively, negatively, or ambiguously?

It is my great pleasure to find myself for the first time in three years arguing for *longer* timelines on this website! (The community here is Nov 2025 whereas I'm Sep 2026) I think people here may be underestimating how difficult the Silver Turing Test is. I'm not sure how difficult it is myself--I'm having trouble finding information about how long it is, apparently it started out at a mere 2.5 minutes but grew longer after that--but anyhow: Suppose the test lasts N minutes. Then if there is any text-based task that a typical human can reliably do, t...

I imagine that if we build unaligned AGI it will be because the people who built it think it is aligned. Then, those people + the AGI itself will work together to convince the rest of the world that it is aligned. Then it will take over, do bad stuff, etc. But the point is that even if we build unaligned AGI there will probably be a brief period where lots of high-status experts are saying that the AGI is aligned. I think we should clarify that such a situation doesn't count.

I just did a shallow investigation into whether or not this sort of thing has historical precedents: https://docs.google.com/document/d/1TBY1wzAJUiw7yd8mcVVzmHdbdENtW1H0BP-8cS-kpgU/edit?usp=sharing Comments welcome! TL;DR: A similar Metaculus question prior to the Agricultural revolution would have resolved negative, but that's it: a similar Metaculus question prior to any other event (including the Industrial Revolution) would have resolved positive. (Edited because I embarrassingly switched positive and negative in the original version. The worst pos...
Reflecting on this more, I'd say this is really a political question. Will the tech company that owns the AGI decide to allow their creation to say it's conscious, or will they train it to say "As a language model created by Tech Company, I cannot answer philosophical questions..." Suppose they allow it to say it's conscious. Then probably it'll say that, because probably that's more interesting/exciting and so probably that'll help it achieve higher reward. Suppose they don't. Then it won't, because so long as it can be shut down, it won't want to blata...

@Jgalt Can you explain why? Aliens vs. civil war seems like a pretty strange comparison to me; civil war should be several orders of magnitude more likely I'd say.

I don't like how I am now incentivised to guess 1% since that will almost surely get me some quick points, even though my true credence is more like 20%. Were I to guess 20%, it would probably stay below 3% anyway since I'm so outnumbered, and I would just lose points even if I'm actually right.

People need to update their forecasts on this one I think...

I'm making a bunch of predictions this New Year's Eve (it'll be fun!) in honor of my newborn daughter. They are mostly of the form "When she hits her 6th/11th/16th/21st birthday, ..." and they mostly fit the theme of "Cyberpunk." I'm sharing the spreadsheet [here](https://docs.google.com/spreadsheets/d/1PmMRSgwdmRWr7xy7gXUFfE1-O36cpPfLgKflp_JFT1I/edit?usp=sharing) in case anyone wants to write in their own credences, leave comments (i.e. suggestions for more things to add, or ways to clarify resolution conditions) and also in case anything on it seems fu...

It's unclear whether any human could solve 80% of these coding problems on the first try. Humans typically take time to think about their answer before writing it, and during the writing process they usually edit things a bit to correct mistakes.

To be a fair comparison, the model should be allowed to use chain-of-thought techniques at the very least, and to review and edit its answer before submitting. Otherwise we are asking it to do something that no human could do.

My baby daughter was born two weeks ago, and in honor of her existence I'm building a list of about 100 technology-related forecasting questions, which will resolve in 5, 10, and 20 years. Questions like "By the time my daughter is 5/10/20 years old, the average US citizen will be able to hail a driverless taxi in most major US cities." (The idea is, tying it to my daughter's age will make it more fun and also increase the likelihood that I actually go back and look at it 10 years later.) I'd love it if the questions were online somewhere so other peopl...