JoshYou

Karma: 688

JoshYou 3 Apr 2026 15:43 UTC
3 points
0 ∶ 0
in reply to: huw’s comment on: How did Leopold do? Evaluating Situational Awareness’s predictions
Chinese companies look a lot more competitive in AI than they did when Situational Awareness was published, even when accounting for the likely impacts of distillation, which I think counts somewhat towards his thesis. But there is little or no evidence of a scaling-pilled AI push from China. My impression is that overall capex/funding for Chinese AI is pretty tiny compared to for the US and state support/industrial policy for AI or semiconductors doesn’t look like a big deal though I don’t have the full numbers on hand. And if anything the Chinese government is getting in the way of Chinese AI companies by discouraging/delaying Nvidia H200 imports, treating AI chips as a normal industry where helping to build up the national champion Huawei is more important than racing to near-term AGI.

JoshYou 2 Mar 2026 5:36 UTC
11 points
0 ∶ 0
in reply to: Chris Leong’s comment on: casebash’s Shortform
Off hand, METR, Forethought, MIT Future Tech, AI Futures Project, AI Index, HAL, and Artificial Analysis all substantially overlap with us in our research focuses and other work, though no two orgs have exactly the same remit. The list of orgs and individual researchers whose work at least partially overlaps with us is far larger.
Personally I think there is a huge amount of descriptive and forecasting research to be done around AI, far more than any one organization can or should take on. I would welcome more “competitors” and I don’t want anyone who is interested in our research areas to feel like we have these topics “covered”. And I’m confident there are many good critiques to be made of our work and much better analyses to can be done on the questions we’ve tackled.

JoshYou 24 Feb 2026 0:56 UTC
1 point
0 ∶ 0
in reply to: JoshYou’s comment on: Front-Load Giving Because of Anthropic Donors?
Indeed, Anthropic is now running a staff share sale with $6B in buyers lined up.

JoshYou 18 Feb 2026 21:28 UTC
3 points
1 ∶ 0
on: Interactive replication of GiveWell’s cost-effectiveness model
insanely cool!
noticed one bug: when you click “reset to defaults” it resets the country choice, not just parameter values (unexpected to me) but more importantly it keeps the UI pointed to the country you chose. Below, Nigeria is highlighted but the parameter values were reset to Chad’s defaults.

JoshYou 8 Jan 2026 4:03 UTC
6 points
1 ∶ 3
in reply to: SiobhanBall’s comment on: SiobhanBall’s Quick takes
i doubt it will be very consequential for EA either way. I think what matters is the discursive impact (effect on prevailing social opinion) not total viewers. and people don’t care very much about Netflix shows. would be different if it was a movie that got traction.

JoshYou 5 Dec 2025 1:00 UTC
26 points
2 ∶ 0
on: Front-Load Giving Because of Anthropic Donors?
Note that Anthropic employees can liquidate significant amounts of equity even without an IPO if Anthropic decides to run a major secondary share sale. OpenAI employees recently sold $6.6 billion in shares.

JoshYou 2 Dec 2025 20:21 UTC
2 points
0 ∶ 0
on: Are there any EA hubs that are both warm and sunny?
Austin is something of a hub I think?

JoshYou 11 Aug 2025 4:26 UTC
3 points
0 ∶ 0
on: Notes and updates on GPT-5
GPT-5 is an enormous step up from GPT-4 in capabilities, to be sure, setting aside training compute.

JoshYou 24 Jun 2025 23:31 UTC
5 points
0 ∶ 0
on: Try o3-pro in ChatGPT for $1 (is AI a bubble?)
If they time the subsidized user push right their model of expected annual recurring revenue is $10B/y and $11B in 2025 is possible
OpenAI says they already hit $10B annual recurring revenue, for what it’s worth. They don’t provide a breakdown but they do say this excludes major one-time deals and the licensing fees Microsoft pays to use OpenAI models in its own products (this is a substantial source of revenue, but I’m guessing OpenAI excludes this to avoid being accused of using wash transactions to juice their numbers: they in turn pay Microsoft for the servers to train and run their models).
Based on OpenAI only having 3M Team+Enterprise+Edu subscribers in May, I don’t think this $10B/year rate was achieved via $1 Team trial subscriptions.
I think you are reading too much into the growth rate of free users. OpenAI has made a recent push into acquiring lots of new free users, e.g. by making signups easier and putting ChatGPT on WhatsApp, which makes their conversion rate look worse. But their revenue, which comes from paid subscribers and API usage, is still growing at a very healthy and relatively steady rate (3x from $3.4B last year, and 10x from $1B in August 2023) and my guess is that it will continue to grow rapidly.

JoshYou 4 Jun 2025 20:45 UTC
4 points
0 ∶ 0
on: Estimating the Substitutability between Compute and Cognitive Labor in AI Research
(comment originally posted on Twitter, Cheryl’s response here)
I’ll flag that estimating firm-level training compute with [Epoch AI’s] notable models dataset will produce big underestimates. E.g. with your methodology, OpenAI spent ~4e25 FLOP on training and 1.3e25 FLOP on research in 2023 and 2024. the latter would cost ~$30 million. but we know OpenAI spent at least $1 billion on research in 2024! (also note they spent $1 billion on research compute after amortizing this cost with an undisclosed schedule).
But I don’t have a great sense of how sensitive your results are to this issue.
(this raises other questions: what did OpenAI spend $3 billion in training compute on in 2024? that’s enough for 50 GPT-4 sized models. Maybe my cost accounting is quite different from OpenAI’s. A lot of that “training” compute might really be more experimental)

JoshYou 21 May 2025 17:32 UTC
13 points
4 ∶ 0
on: Is There Actually a Standard or Convincing Response to David Thorstad’s Criticisms of the Value of X-Risk Reduction and of Longtermism?
Note that Thorstad’s arguments apply more against strong longtermism, i.e. that future generations are overwhelmingly or astronomically more important than current generations, not merely that they are important or even much more important than current generations.

JoshYou 29 Apr 2025 14:26 UTC
2 points
0 ∶ 0
in reply to: Parker_Whitfill’s comment on: Parker_Whitfill’s Quick takes
They could exclusively deploy their best models internally, or limit the volume of inference that external users can do, if running AI researchers to do R&D is compute-intensive.
There are already present-day versions of this dilemma. OpenAI claims that DeepSeek used OpenAI model outputs to train its own models, and OpenAI doesn’t reveal their reasoning models’ full chains of thought to prevent competitors from using it as training data.

JoshYou 29 Mar 2025 22:21 UTC
3 points
2 ∶ 1
on: How We Might All Die in A Year
Kinda weird that the story contains an intelligence explosion that happens both incredibly fast and incredibly soon but glosses over how it happens in a single paragraph, in favor of descriptions of nanobots dematerializing people.

JoshYou 8 Mar 2025 0:55 UTC
6 points
0 ∶ 0
in reply to: huw’s comment on: Anthropic’s submission to the White House’s RFI on AI policy
This is worth considering, but FWIW, 50 GW would be around 10% of US electricity if it runs continuously (the US consumes at a rate of about 500 GW if you divide total consumption by one year). If the new capacity is as clean as the overall electric grid that would be about 2.5% of US emissions (25% of US emissions come from electricity) and 0.35% of global emissions (US emissions are ¹⁄₇ of global emissions).
I’m not going to do this math now but I think if the new capacity is 100% natural gas then that’s about as carbon-intense as the US electric grid as a whole, or maybe somewhat worse (the US has a lot of clean energy, but it also has coal plants which are >2x more carbon intense than gas). 100% natural gas would be the worst case, because there is no scenario where the US builds new coal plants (edit: it’s not the worst case, because increased power demand could cause the delay of coal plant retirements, but I don’t think this changes the conclusion all that much)

JoshYou 26 Jan 2025 0:58 UTC
10 points
0 ∶ 0
on: Is DeepSeek-R1 already better than o3 when inference costs are held constant?
R1 is probably not 6x cheaper than o1-mini and 30x cheaper than o1 in terms of the actual, underlying cost. (meaning that DeepSeek probably charges a much lower gross margin on its API than OpenAI does). R1 has 37B active parameters (though its 671B total parameters are also relevant). We don’t know how many parameters o1-mini or o1 have, but IMO they’re probably a lot less than ~200B and ~1T, respectively.

JoshYou 10 Jan 2025 16:33 UTC
4 points
0 ∶ 0
in reply to: Cullen 🔸’s comment on: Funding Diversification for Mid-Large EA Organizations is Nearly Impossible in the Short-Medium Term
Forbes estimates that the seven co-founders are now worth $1-2 billion each.

JoshYou 28 Aug 2024 19:43 UTC
1 point
0 ∶ 0
in reply to: Jacob Watts🔸’s comment on: How do you deal with the “meat eater” problem
I’m not proposing any sort of hard rule against concluding that some people’s lives are net negative/harmful. As a heuristic, you shouldn’t think it’s bad to save the lives of ordinary people who seem to be mostly reasonable, but who contribute to harmful animal agriculture.
The pluralism here is between human viewpoints in general. Very naively, if you think every human has equal insight into morality you should maximize the lifespan and resources that go to any and all humans without considering at all what they will do. That’s too much pluralism, of course, but I think refraining from cheaply saving human lives because they’ll eat meat is too far in the other direction.

JoshYou 23 Aug 2024 14:19 UTC
11 points
5 ∶ 3
on: How do you deal with the “meat eater” problem
I think if you put some weight on viewpoint pluralism you should mostly not conclude that other peoples’ lives aren’t valuable because those people will make the wrong moral choices.

JoshYou 3 Jan 2024 13:15 UTC
5 points
2 ∶ 0
on: Stop talking about p(doom)
Toby Ord’s existential risk estimates in The Precipice were for risk this century (by 2100) IIRC. That book was very influential in x-risk circles around the time it came out, so I have a vague sense that people were accepting his framing and giving their own numbers, though I’m not sure quite how common that was. But these days most people talking about p(doom) haven’t read The Precipice, given how mainstream that phrase has become.
Also, in some classic hard-takeoff + decisive-strategic-advantage scenarios, p(doom) in the few years after AGI would be close to p(doom) in general, so these distinctions don’t matter that much. But nowadays people are worried about a much greater diversity of threat models.

JoshYou 28 Dec 2022 16:10 UTC
10 points
0 ∶ 0
on: How many hours is your standard workweek? Why?
I’m probably “on the clock” about 45 hours per week—I try to do about 8 hours a day but I go over more often than not. But maybe only about 25-35 hours of that is focused work, using a relatively loose sense of “focused” (not doing something blatantly non-work, like reading Twitter or walking around outside). I think my work output is constrained by energy levels, not clock time, so I don’t really worry about working longer hours or trying to stay more focused, but I do try to optimize work tasks and non-work errands to reduce their mental burdens.