I’m a computational physicist, I generally donate to global health. I am skeptical of AI x-risk and of big R Rationalism, and I intend on explaining why in great detail.
titotal
Karma: 9,386
Does disaster frequency follow power laws? It’s complicated
Interesting analysis! I’m somewhat skeptical that space based cells will perform as well as you find in this model. This seems like a classic case where something that looks good on paper will run into problems in real world applications for something as complicated as putting massive datacenters in space. I don’t think your model is accounting for murphy’s law here.
Also, on a technical note, do you have the source for the starlink solar cell specifications used for the phrase “Back-calculated from Starlink V3 specifications (50,400 W from 554 kg array, 257 m² area)”? This seems like one of the critical figures of your model but I couldn’t find where you linked the source for it.
It is extremely difficult to determine base rates for something like sexual harassment, because it’s an offence that allows for ambiguity and plausible deniability, because there’s room for retaliation, etc, and it will strongly depend on how much people trust the bodies they are reporting to.
What we can do is look at the responses to the incidents that do get raised, and the experiences of victims, and judge whether or not they live up to the standards we want to see in a group that takes sexual harrasment seriously. I do not think the grades are very good on this front.
Only 3 months ago we had a writeup detailing a shockingly terrible response to sexual harrassment by one of the most prominent EA orgs out there. The response is far worse than anything I’ve ever seen at any organisation I’ve ever been in. This indicates to me that the environment is nowhere the high standards that should be aimed for.
Regardless of the actual base rates, the question that matters the most is whether there is room for improvement, and I think it’s blindingly obvious that the answer is yes.
LLM disclosure in general is just a good idea to do. The internet is absolutely flooded with LLM-written spam at the moment, so if people detect LLM writing with no context it’s natural to assume your post is spam as well. This is a shame when someone who is a non-native speaker has just used it for translation or whatnot.
Personally I’d recommend against using LLM-written text if you can help it, as in the age of spam the value of cultivating your own stylistic voice is increasing.
Analysing the extreme spread in existential risk survey estimates
I think you would benefit from re-reading the article in question. For example, they directly adress your point 1 by pointing out that consumer diffusion figures are often misleading by expressing figures in terms of “percentage of people that use chatbots on occasion”, rather than on frequency of use.
Point 3 is not even an argument, just a restatement of what they believe: yes, they think AI domination will take decades. They state the reasons they believe this very clearly in the section “Diffusion is limited by the speed of human, organizational, and institutional change”: if you disagree with this, you have to present actual arguments. From what I know, most economists would agree with them.
Point 5 is not an argument either: they are not to blame for how you interpret their “vibes”. If people interpret “AI will be akin to the internet” as anything other than “AI will be akin to the internet” that’s their fault, not the authors.
As for point 6, I’m confused as to what your position is here. Do you think that AI systems are merely cheating on every single benchmark? In the section “benchmarks do not mention real-world utility”, I took them as referring to benchmarks that are actually meaningful: saying that while they genuinely are good at taking law tests, even non-contaminated ones, that this doesn’t translate into being a good lawyer because of the aspects that are not easily measurable. I don’t see how this is a contradiction to any of their previous work?
I think “speck of dust in the eye” was a bad choice for the central example of this debate, because in some situations a speck in your eye can be literally zero painful, and in others it can be actually quite painful and distressing. I think this leads to miscommunications and poor intuitions.
My preferred alternative would be something like “lightly scratching your palm with your fingernail”. And while this is technically pain, I find a single light scratch to be so minor that it has literally zero effect on my levels of happiness: in fact I will sometimes do this to myself on purpose when I get sufficiently bored.
I therefore think that that premise 1: “mild pain is bad”, is wrong for sufficiently small definitions of “mild pain”. I think you need a threshold of badness for the argument to work. Furthermore, I think most people who would side with the “dust specks” also have some threshold where they would pick the torture: for example if it was “punching a billion people in the face vs torture one person”.
To be clear, I wasn’t saying that complexity itself was the cause of consciousness, just that some level of algorithmic complexity may be a requirement for consciousness. This seems like a common position: the prospect of present or future LLM sentience is a subject of debate, but it’s rare to see a similar debate about the sentience of a pocket calculator.
A brain and a digital simulation have some similarities, but they also have a lot of differences. One of those differences is that the brains are running on “laws of physics” algorithms that are overwhelmingly faster and more complex than that of digital simulations. They didn’t need to evolve these “algorithms”: it’s inherent to any biological process. Seth identifies several other differences as well: continuous operation, embodiment, etc. His position seems to be that at least one of these differences may result in a lack of consciousness.
disclaimer: I am not too well-versed on the philosophy here so I could be saying dumb things, feel free to correct:
From my computational physics experience I know that it is physically impossible to simulate the exact electrical properties of a system of a couple hundred atoms on a classical digital computer, due to a blowup in computational complexity.
The laws of physics could be described as an algorithm, but the algorithm in question is on a level of complexity that is impossible for digital simulations to match. I think it’s generally agreed that some degree of complexity is required for consciousness: it doesn’t seem insane to say that that complexity might lie past what is digitally simulatable in practice.
The question of digital consciousness seems to depend on whether simulated abstracted approximations to the physical process of thinking are close enough to produce the same effect.
I am somewhat concerned about data contamination here: Are you sure that the original Givewell writeup has at no point leaked into your model’s analysis? Ie: was any of givewell’s analysis online before the august 2025 knowledge cutoff for GPT, or did your agents look at the Givewell report as part of their research?
Yeah, the future described in this post isn’t particuarly “weird”, per se, it’s just using the assumption that every technology that has been hypothetically proposed for the future will be created by ASI soon after AGI arrives.
I think the future will be a lot more unpredictable than this. Analolgously, I can imagine someone from 1965 being very confused about a future where immensely powerful computers can fit in your pocket, but human spaceflight had gone no further than the moon. It’s very hard to predict in advance the constraints and shortcomings of future technology, or the practical and logistical factors that affect what is achieved.
Have you considered that the reason these policies are not increasing AI usage is that AI usage is not particularly useful for many applications? Particularly when it comes to something like animal advocacy, I’m struggling to think of many things you’d actually need a full model subscription for (rather than just asking the occasional question to a free model).
I think the original policies are fine: they let people evaluate and decide for themselves how useful AI models are, and adjust strategies accordingly. Trying to pressure people to use AI beyond this level is going to make your team less effective.
You correctly point out that “AI safety leaders” is a group that selects for high concern about AI, which means that the average is skewed towards high concern, relative to experts more generally.
I would like to add that the same is probably true (to a lesser extent) for AGI timeline estimates: People that think that AGI is very far away are less likely to think that AI safety is a pressing concern and are thus less motivated to become AI safety leaders. Also, people who are concerned about present-day AI risks, but don’t think AGI is imminent often call themselves “AI ethicists”, rather than AI safety people. These “AI ethicists” are unlikely to show up to a “summit on existential security”.
To be clear, I think it’s good to write this article, but we should always be mindful of selection effects when interpreting surveys.
Unfortunately, most estimates of LLM energy use are somewhat out of date due to the rise of reasoning models. A small amount of personal usage is probably still not that energy intensive, but I don’t think it’s negligible anymore.
The most up-to-date estimates I’ve seen of AI energy use is this paper here. I recommend you look at table 4. For the o3 reasoning model, which is probably the closest analogue to todays reasoning models, a short query costs something like 7 Wh, a medium query is 20 Wh, and a long query is 30 Wh. Using a non-reasoning model like GPT-4o was much less intensive at like 0.4 Wh for a small query, however in my experience the results tend to be a lot worse.
So if you end up using like 10 medium queries to a reasoning model over the course of a project, that would add up to 0.2 kWh: if you use 100 queries, that would be 2 kWh. The typical household energy use is something like 30 kWh per day. So the impact is small, but non-neglible: probably there are other things you can do that will have a bigger impact on energy use.
Personally, I would be worried about cognitive offloading: I think that an overreliance on AI can hamper your ability to learn things, if you offload mentally difficult tasks to the AI.
This interpretation is not true. Thiel was talking specifically about money going to Gates in the event of Musk dying:
That’s how Thiel said he persuaded Musk. He said he looked up actuarial tables and found the probability of Musk’s death in the coming year equated to giving $1.4 billion to Gates, who has long sparred with the Tesla CEO.
“What am I supposed to do—give it to my children?” Musk responded, in Thiel’s telling. “You know, it would be much worse to give it to Bill Gates.”
I think this would only make sense if Musk had specifically willed his pledge money to the gates foundation?
I think there is a good reason to focus more on novice uplift than expert uplift: there are significantly more novices out there than experts.
To use a dumb simple model, say that only 1 in a million people is insane enough to want to kill millions of people if given the opportunity. If there’s 300 million americans, but only 200 thousand biology PhDs, that means we expect there to be 300 crazy novices out there, but only 0.2 crazy biology PhD’s. The numerical superiority of the former group may outweigh the greater chance of success of the latter group.
I fully agree with this post.
I think this type of belief comes from a flattening understanding of the difficulty of doing things: it’s assumed that because doing well on a math olympiad is hard, and that curing death is hard, that if you can make an AI do the former it will soon be able to do the latter. But in fact, curing death is so much more difficult to do than a math olympiad that it breaks the scale.
You can also see this in the casual conflation of things like “cure cancer” and “cure death”. The latter is many, many, many orders of magnitude more difficult than the former: claiming that the latter would occur at the same time as the former is an incredibly extraordinary claim, and it requires commensurate evidence to back it up.
The chief argument in favour of this is “recursive self improvement”, but intelligence is not a magic spell that you can just dial up to infinity. There are limits in the forms of empirical knowledge, real-world resources and computational complexity. Certainly current day AI trends seems to be limited by scaling laws that would be impractical a pretty fucking long way from god-like intelligence.
This seems like the original article that is being quoted from. The quoted comments seem pretty bleak:
Thiel said he’s nudged a few to erase their signatures. “I’ve strongly discouraged people from signing it, and then I have gently encouraged them to unsign it,” Thiel said. Notably, in transcripts and audio lectures given by Thiel to Reuters last year, he recalled calling on the world’s richest man and soon-to-be first ever minted trillionaire Elon Musk to retract his pledge, warning the Tesla founder his wealth would go to “left-wing nonprofits that will be chosen by Bill Gates.”
Thiel said he’s had conversations with some signatories who have expressed uncertainty about their original decisions to commit. “Most of the ones I’ve talked to have at least expressed regret about signing it,” he said.
It seems like you’re not saving much time by doing the double degree, compared to two single degrees. Why not do a single degree in CS and then retrain if and only if the market goes south?
Also… EA aside, what do you want to do? If you really like dentistry and think you’ll do well in it, but coding makes you miserable, then dentistry is probably the right choice for you. For any of this “lifetime contribution” considerations to matter, the choice has to be one that you can sustain.
This has not been my experience from 9 years of academia in physics and material science. Opinions published in scientific papers must be backed up with reference to actual evidence, not merely opinion. When deferral happens behind the scenes, it’s usually justified by the person in question being an actual expert that knows their shit.
EA is far worse: I sometimes see people defer to random blog posters who have zero expertise in the subject they are talking about.