We already see lagging open-weight models as strategic plays by competitors using a fraction of their available resources, and I’m not sure that stops making sense when the fraction is a bigger absolute number. Sovereign investment could also make a comeback but seems unlikely.
In any case I don’t think investors are under any illusions about what happens if bigger training runs stop paying off
Mainly informed by this sort of thing bsky.app/profile/epoc..., and I can believe in straight lines of costs falling for equal capability more easily than in those of more capability for more training
It’s also still unclear to me whether they are recouping training costs per model or just serving inference profitably.
It’s more like “can’t” than ”don’t want to”, right? As you say, investors would flee (and open models would catch up in 6-12 months, margins would vanish…). I think the urgency makes it more than quantitatively distinct from ordinary R&D spending.
Will Elon Musk be the world's richest person on December 31, 2025? RESOLVED Yes Both the community forecast and user forecasts hover around 75% for most of the year, but the user forecasts dip indecisively to 50 around August until November, by which time they should have long since known better.
I think my worst was on Musk’s wealth. There was a moment when he was losing ground and Ellison was gaining on him, and I got a little carried away with wishful thinking. But I only lost 3.6 points in peer score in the end.
Will there be a bilateral ceasefire in the Russo-Ukraine conflict before 2026? RESOLVED No The community forecast peaks in March 2025 around 75% and decreases smoothly from there. The user forecasts form a relatively dense curve consistently around 2/3 of the community forecast.
If they bring back the view sorted by my score I’ll post that, but I think my best question was on Russia-Ukraine peace. No special insight, and I wasn’t even particularly further from the crowd than on other questions. I just kept the forecast updated better (20 vs my median 4 updates).
And I finished #1 in the 2025 Vox Future Perfect Contest (forecasts scored over the whole year)! I’m not sure what happened here. No especially high-scoring forecasts, but I only lost a few points on a few. I guess once in striking range you eventually get lucky on a bundle of 20-30 questions.
My essay 'After Orthogonality,' the culmination of 4 years of work on virtue-ethics, rationality, AI, decision-theory, and praxis is out on The Gradient
If someone loses their job to a more productive machine, total productivity goes up. It doesn’t automatically go up enough that everyone could be better off after redistribution. Why do people so often talk as though it does?
Question, Coverage, Score Will at least twice as many deportations by U.S. ICE occur in Fiscal Year 2025 compared with Fiscal Year 2024? 100.0% 148.800 On December 31, 2025, will Google, Meta, Amazon, Tesla, or X accept crypto as a payment? 100.0% 132.131 Will a new war or a substantial escalation to a previous war kill at least 5,000 people in 2025? 100.0% 111.363 Will there be a bilateral ceasefire in the Russo-Ukraine conflict before 2026? 100.0% 87.385 Will semaglutide be taken off FDA's drug shortage list in 2025? 100.0% 84.914
Will the Democrats be favored to win the 2028 US presidential election in the last week of 2025, according to Kalshi? 100.0% -3.855 Will Elon Musk cease to be an advisor to Donald Trump and face public criticism from Donald Trump before 2026? 100.0% -10.695 Will the 12-month percentage change in the US Consumer Price Index be lower in November 2025 than it was in November 2024? 100.0% -11.170 Will there be at least 1,000 deaths due to direct conflict between Israel and Iran in 2025? 100.0% -49.755 Will any rationalist, effective altruist, or AI safety researcher go on the Joe Rogan Experience before 2026? 100.0% -76.455
I finished #16 in the 2025 ACX Prediction Contest (spot forecasts Jan 31 2025). Better than I'd expected, not well enough to force ACX to plug my 7th most recent blog post. Best five questions and worst five questions. I don't think I got lucky, probably should have known better on some of these.
The idea that Davis is more believable for not being self serving was always odd to me. Setting aside how everything serves the “hero in grand struggle against blinded rationalist leaders” self-mythology, believing unflattering things about yourself is also an extremely common cognitive distortion!
Yeah, to me (in physics) sequential colormaps with monotonic lightness and changing hue are usually in the "heatmap" family where brighter is more (viridis using blue-green-yellow, inferno using black-orange-yellow). If you're just changing saturation then I think darker usually looks like more.
Emil O. W. Kirkegaard: "learning outcomes" are really just intelligence tests, which they call PISA etc.
He's not worried about bias, since it's "an extremely simple regression that it would be hard to fake." Meanwhile, the preprint's author seems to affirm in the comments that he believes "Which organ in a frog has a function similar to the function of lungs in a bird?" is really just an IQ question.
This sounds like a joke, but it's literally the argument. Here's Scott Alexander, who would like you to believe Lynn's numbers have been confirmed by a preprint that averages them with data on learning outcomes and finds the result correlates similarly with measures of national development.
I think I get what you’re saying, but this is also maybe how one ends up being Alexander thinking he can just sort the garbage out of his nrx-enriched ecosystem
Urgently want to, I put a hold on it at the library but was like 9th in line
That one put me in a weird headspace for a while. The summary doesn’t really get across how heavy it feels for such a short stylized piece.
For the maximalist endpoint of liminal and surreal you might like Solenoid?
It is very good, I am halfway between “wish it were two times longer” and “who am I (is anyone) to second guess Clarke”
...and now they are calling the NYC mayor race for this guy! www.youtube.com/watch?v=bEvV...
Rose/House by Arkady Martine?
I like Kagi. I don’t know if you’ll find it’s worth paying for but there’s a free trial. I still get spammy results for some searches but I don’t think it’s as bad as DDG or Google lately. The founder is into AI but the AI features are all opt-in.
Definitely agree about both effects of revenues being spread out. I have never really found the argument that competition will speed things up via “race dynamics” compelling, especially in a capital-intensive industry (although it does make a coordinated slowdown harder).
If I didn’t know better, I’d have taken “as newsrooms have diversified beyond the white, Western, and male culture…, this approach has been scrutinized. Should journalists reconsider the ethical merits of objectivity” to mean “we’ve become less biased, but maybe it’s not all that great?”
It reads like the authors and everyone they quote are using “objectivity” as a term of art for a particular thing taught in journalism school, not realizing it means something else to most people.
Prohibiting certain words, therefore, would not deprive egos of their expressive possibilities so much as change egos. Neurath explained that one eventually learns, as he did, to avoid dangerous words "half-consciously and without constraint" (1941a, p. 146). The result is not the same person with new linguistic habits but--literally--a changed person: "Building up a Universal Jargon needs a comprehensive training, which is connected with an alteration of our whole attitude.... What comes from an `experiment' with a modified scientific language will be analyzed by a man who is modified by this `experiment,' which is more than an experiment: it performs a kind of self-education" (1941b, p. 216; author's emphasis). Those who like to use words that come to be prohibited may feel frustrated at first, but their mental habits will change along with their vocabularies. With the success of the Unity of Science Movement, for instance, the divisions between the sciences--and the words used to delineate them--would start to seem anachronistic: "A new generation educated according to unified science will not understand the difference between the `mental' and the `physical' sciences, or between `philosophy of nature' and of `culture'" (Neurath 1933, p. 9). Similarly, devotees of metaphysics would miss their favorite concepts no more than scientists of today miss the vocabularies of phrenology or alchemy.
Screenshot of part of a list of potential prohibited metaphysical words with sources in parentheses where Neurath mentioned or discussed each word: explanation (h, k) external world / internal world (a) fact (a, k) forces (a) good/bad (a, d17) good/evil (c226, d17) idea (f) immanent (d8) interests (m, a16) interpretation (i) intuition (d8) judgment (i) justice (a, c23) material/immaterial (a) meaning (a, a18, b66, c218, e147)
go.gale.com/ps/i.do?id=G...
bsky.app/profile/nomi...
This one's also tricky because I find his claim to have been desperate for reasons to disbelieve for at least seven years further less credible for how easy it was to find multiple fatal problems with interpreting the survey as he does, but I can't assume the reader agrees with me on those problems.
I’ve quoted this entire passage since I want to emphasize very clearly that I’m not trying to use this as a bludgeon against Alexander for not reaching the same conclusions that I did about the survey. I’m pointing this out for the sake of anyone reading him: when he says he’s been freaking out, recognizes his biases, and wishes for reasons to disbelieve the study, be aware he was citing it 7 years prior as evidence for his opinions alongside his strategy to promote those opinions without publicly endorsing them.
Added a section. I didn't include this example to begin with because I was worried that it would trigger much more defensiveness than the other examples, particularly considering the effort Alexander puts into preempting attacks here. Maybe it will help to address that directly.