WilliamKiely

Karma: 1,100

How I think developers of frontier AI systems and regulators ought to act in the face of existential AI risk

WilliamKiely19 Jun 2026 22:22 UTC

12 points

0 comments12 min readLW link

WilliamKiely 16 Jun 2026 18:06 UTC
4 points
0
in reply to: Zach Stein-Perlman’s comment on: WilliamKiely’s Shortform
Agreed that there is significant value in moving us between different high-value futures.
Intuitively though, I still think “weak bimodality” combined with my view that low-value futures are much more likely than high-value futures means that the main thing we should be aiming to do now is get to a high-value future. In other words, I still think maxipok is the right policy, where an “okay outcome” is defined as any high-value future.
Do you agree with this, or do you think some of our collective portfolio of actions should be focused on trying to raise the probability of very-high-value futures relative to somewhat-high-value-futures even while (or if) it is still the case that high-value futures are much less likely than low-value (near-zero-value) futures?

WilliamKiely 11 Jun 2026 17:45 UTC
4 points
0
in reply to: Zach Stein-Perlman’s comment on: WilliamKiely’s Shortform
Huh, I don’t think that change of view undermines the main point.
In my view the main insight is that the value of the future is bimodally distributed (perhaps there is a small third mode for very negative value futures, but I think it’s much smaller than the ~zero mode and the ~very postive mode so we can basically ignore it).
Whether the very positive mode probability mass is distributed over 90%-100% optimal futures or over 0.1%-100% optimal futures (or even wider) doesn’t really matter. The key thing is that the ~zero mode is distributed over +/-10^-10% optimal futures or even +/-10^-20% optimal futures, and the space inbetween from 10^-10% to 0.1% optimal futures has lower cumulative probability mass than either mode. So the EV of the future is basically equal to the probability of getting a future in the high value mode times the average value of a future in that mode.
Perhaps you think acausal trade considerations make ~zero value futures unlikely (does most of the value of the future come from value outside of our lightcone that we acausally make happen due to what we do in our lightcone? I’m not really sure how to value the multiverse—assuming it exists it seems that our lightcone could at most only have a small (acausal) influence on it, and so the absolute value of the multiverse will be basically the same regardless of what happens in our lightcone), but at least ignoring acausal trade considerations and only looking at the reachable universe within our lightcone the above value-is-bimodally-distributed view still seems right to me.
It just seems to fall out of the fact that the potential value of the future spans dozens of OOMs. There are two main factors -- (1) How much of the reachable universe’s resources we use (including for how long) and (2) How hard we optimize the resources we use on average—each of which spans many OOMs, but I don’t think the fact that there are two factors rather than just one (or three rather than just one if we separate out the time factor) changes the bimodal nature of our forecast of how valuable the future is.
I’ll check out MacAskill’s article later.

WilliamKiely 10 Jun 2026 5:15 UTC
6 points
0
on: WilliamKiely’s Shortform
@Zach Stein-Perlman ’s 2021 quick take Value Is Binary is very under-rated. I’ve referenced the idea well over a dozen times in the last 4+ years and just re-read it after being reminded of it again.

WilliamKiely 26 May 2026 5:40 UTC
4 points
0
on: Donating 80% While It Still Counts
(Copying my comments over from Facebook:)

Thanks for all your donations and direct work and transparency about these things!
From my perspective it’s unclear whether or not you’re leaving a lot of potential impact on the table, due to a few considerations:
(1) Are your retirement savings invested optimally? The difference between e.g. a 25% annual return and a 50% annual return on your ~$1.5M in retirement savings is $375,000 in the next year. I haven’t had a 401(k) in years, but my understanding is that 401(k)s limit how you’re able to invest the money, so if your retirement savings are in a 401(k) this could be very suboptimal assuming AI stocks are going to keep growing much faster than the rest of the economy, as I expect they probably will for the next couple of years at least. (VanEck Semiconductor ETF is up 135% in the last 12 months.)

(2) What do you think is the difference in cost-effectiveness between the most cost-effective donation opportunities you can find and your current work? SecureBio sounds like valuable altruistic work, but even if it’s a 99th percentile donation opportunity, you may be able to identify a ~99.9th percentile donation opportunity that is 10x as cost-effective, in which case your salary-sacrifice direct work strategy may be suboptimal. Or perhaps there is a different direct work job you could do that would be much higher impact than your work at SecureBio.

It sounds like you put a much higher probability on AI-caused catastrophe in the next two years than I do, though maybe your 10% forecast below just wasn’t made carefully and your true credence is actually lower:
Let’s try and make some similar predictions for 2028:
My odds that the world has changed substantially are up significantly, maybe 55%, primarily due to AI. I’d put about 10% on futures where things go very badly, where I’m not here to write a followup and you’re not here to read one.
10% that you’ll be unable to write a blog post by 2028 (even end of 2028) due to being dead from a global catastrophe seems *way* too high to me. I’d put that at less than 1% by mid-2028. Also, if there really is a 10% chance that you’ll be dead two years from now due to an AI-caused catastrophe, it’s worth flagging that this means that AI represents about 95% of your mortality risk in the next two years (!!!). (99.3% of American men who turn 40 make it two more years to their 42nd birthday according to https://www.ssa.gov/oact/STATS/table4c6.html).
I wonder if your forecast is 10% because you’re much more pessimistic than I am about AI risk in general or because you have more aggressive AI timelines (or both)?
To answer this, what are your forecasts that you’ll be dead as a result of a global catastrophe by mid-2028/2030/2032/2035/2040? Well actually, I’m not sure whether a good forecast for this for 2040 is 10% or 70%, but if the right forecast for the risk of you dying due to an AI-caused catastrophe by mid-2040 is X%, then what’s your forecast for this by mid-2028/2030/2032/2035? My forecasts would be roughly 0.01X/0.05X/0.2X/0.6X, so my number by mid-2028 is ~0.1%-0.7%, i.e. much lower than your 10%.

Also, in those 10% of scenarios where you’re imagining where we’re dead in two years due to an AI-caused catastrophe, what kind of AI-caused catastrophe are you imagining? Is the bulk of the (you and I aren’t here in two years) risk AI takeover risk or engineered pandemic risk or something else in your mind?

WilliamKiely 13 Apr 2026 20:30 UTC
3 points
0
in reply to: XelaP’s comment on: Which things were you surprised to learn are metaphors?
You may be reading too much into my label. For most of those they don’t feel like anything to me.
The main thing I feel when it comes to times is, as I said, times throughout a year feeling like they have a spatial location or direction. So e.g. August feels like it is up-and-to-the-left in front of me. (Holidays and seasons and other events like friend’s birthdays also have this directional spatial feel to them, i.e. I locate where approximately they are on the disk without thinking about it.) This isn’t some memory palace technique I came up with to help me remember dates; it’s just something I realize I always imagine unconsciously after I learned about synesthesias that not everyone does. (In describing it to others I have found that some other people who have basically this same representation of time throughout the year. IIRC one person I talked to once said they visualize a spiral with the months in different years at different levels.)
When I think of different years I do visualize them on a number line streching back a few centuries to the left. This is easier to explain by saying that time is often depicted on a number line timeline.

WilliamKiely 5 Apr 2026 0:59 UTC
13 points
4
in reply to: WilliamKiely’s comment on: WilliamKiely’s Shortform
Though to push back on this, (1) it was always the case that AI would eventually be a big deal (even though most people throughout history didn’t recognize this). And similarly, (2) in a market economy it was probably also pretty inevitable that there would eventually be multiple firms competing to create increasingly powerful AI. And (3) people see themselves as good people, so competitors to existing AI firms would presumably have said they were doing it for good reasons (e.g. to prevent a monopolist from getting all the power).
Essentially I’m trying to say that even without Bostrom or the AI safety community pre-OpenAI, it was probably the case that some competing firms to DeepMind would have sprung up. Maybe the billionaires like Musk would have had to have been pitched a different message to persuade them to cofound the new competing firm, but it probably would have happened anyways, maybe just on a different timeline.
So the upshot would be that the AI safety community helped create OpenAI, but that the help was not necessary—a company like OpenAI probably would have been created at some point regardless. The race to AGI by multiple competing firms that we are on now just seems like the default path in retrospect.

WilliamKiely 5 Apr 2026 0:46 UTC
4 points
2
in reply to: Kajus’s comment on: WilliamKiely’s Shortform
I originally disagree-voted, though I changed my mind after reflecting on it just now and removed my disagree-vote. Here’s my thought process.
Musk thought that AI was going to be a huge deal in large part because of Bostrom and the AI safety community, and this contributed to Musk cofounding OpenAI when Altman and Brockman said to Musk that Hassabis couldn’t be trusted with AGI.
The Hassabis biography quoted Brockman’s 2018 email to Musk:
Put increasing effort into the safety/control problem, rather than the fig leaf you’ve noted in other institutions. It doesn’t matter who wins if everyone dies. Related to this, we need to communicate a “better red than dead” outlook — we’re trying to build safe AGI, and we’re not willing to destroy the world in a down-to-the-wire race to do so.
https://www.lesswrong.com/posts/5jjk4CDnj9tA7ugxr/openai-email-archives-from-musk-v-altman
So the AI safety community definitely was a major influence in the creation of OpenAI.
But does that mean that it’s “as if the AI safety community created OpenAI”?
If Musk was part of the AI safety community and was acting on its behalf then sure, I think your statement would accurate.
But saying that it’s as if the AI safety community literally created OpenAI just because the AI safety community influenced its creation doesn’t seem accurate.
So was Musk acting for the AI safety community in his cofounding and funding of OpenAI?
Well, at the end of Life 3.0 (which I read this week after The Infinity Game) Tegmark tells of how he asked Musk to donate to fund FLI when it first got started and how Musk agreed to donate $10M and how Musk went on stage at MIT in 2014 to make the announcement at an AI safety event.
But of course Tegmark may have just been fundraising from Musk as a sympathetic person with money who might donate. This doesn’t mean that the rest of Musks actions related to AI were necessarily endorsed by Tegmark or the rest of the AI safety community.
But in fact, when Tegmark is painting a positive picture of AI safety community progress at the end of his book (2017 publication) one thing he says is how the nonprofit OpenAI was created with a beneficial humanitarian mission (paraphrasing from memory). That was the part of Life 3.0 that seemed to stand the test of time the worst. Tegmark in his book seemed to think OpenAI’s creation was clearly a positive, suggesting that the creation of OpenAI was in fact endorsed at the time by a significant part of the AI safety community.
Of course the AI safety community isn’t a monolith, but I now think this is enough to make your statement that it’s as if the AI safety community created OpenAI true enough in the sense that Musk could be said to have been a part of the AI safety community, cofounding OpenAI with the endorsement of other members of the AI safety community such as Tegmark.

WilliamKiely 1 Apr 2026 20:49 UTC
6 points
0
in reply to: WilliamKiely’s comment on: WilliamKiely’s Shortform
From Chapter 18:
LIKE ALTMAN AND DARIO AMODEI, Hassabis refused to join Bengio in signing the pause letter. Indeed, he objected to it fiercely.
”I didn’t sign because a six-month moratorium doesn’t help,” Hassabis told me.
”Who would have stopped development? Just people who signed? Well, that’s no use because you need the whole world to pause, including China. Who would have monitored it?
”I mean, a pause could actually have made things worse.
”Imagine we had a ten-year moratorium, OK? That would slow down the advance of AI, but everything else would carry on as normal. So, you develop better and better chips, data centers, all that. Then we exit the moratorium and the proverbial programming prodigy in his parents’ garage now has a home computer with the power of a data center!
”We’re supposed to be advancing safety. How is that going to do it? The race condition would be insane at that point!
”I mean, it’s insane right now, but maybe there’s some hope because there are only a few leading actors, and we all know each other.
”After a moratorium, you’d be beholden to random actors.”
Hassabis had a point. A pause by itself would not achieve much.[17] Indeed, in a roundabout endorsement of Hassabis’s argument, the extreme doomster Eliezer Yudkowsky also refused to sign the letter. The way Yudkowsky saw things, the only way to save humanity was for governments to ban frontier development outright, by closing down computer servers. If some countries refused to join the ban, others should be “willing to destroy a rogue datacenter by airstrike,” he asserted.[18] With a p(doom) approaching 100, Yudkowsky thought any measures could be justified. It would be worth risking nuclear war to avert the even greater calamity of rogue superintelligence, he insisted. The costs of an infinity machine could be infinite.
Two months after the pause controversy, at the end of May 2023, the safety debate inched forward. Bengio, Hinton, and Hassabis, together with the leaders of the other major labs, signed a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks, such as pandemics and nuclear war.” Some 350 notables added their names to the letter. Only Meta and the open-weight partisans were absent from the list of signatories.[19]
”I thought long and hard about signing that one,” Hassabis told me. “I would’ve liked an extra sentence acknowledging the upsides—‘We believe the potential of AI is going to be amazing,’ or whatever.
”But I signed because it was important for credible people to oppose the idea that there’s no risk at all.
”The point was to say that there really is a risk of catastrophe. We have no idea what the percentage chance is. We have no idea of the timescale. But it’s nonzero. And it’s going to be really hard to sort out, and it could be really serious if it does happen.
”We wouldn’t have needed to do this if there hadn’t been people like Yann LeCun saying, ‘Oh, there’s nothing to see here.’ Which I think is pretty crazy given the uncertainties.
”He says, ‘I’m sure there’s a safe way to build AI.’ And I agree. It might turn out that as we develop these systems further, it’s way easier to keep control of them than we expected.
”Then he says, ‘Therefore, we will build it in that safe way.’ And that’s where I don’t understand his argument.
”First, we don’t yet know what that safe way is.
”Second, what’s to stop half the world building it the wrong way, even if Yann was somehow to build it correctly?
”It’s like with the open-source debate. What’s to stop bad actors getting hold of the model and then repurposing it for bad ends? What’s the answer to that? There isn’t one.
”And it’s not just Yann. There are all these other people in the Valley.
”I mean, not long ago they were talking about crypto. People who go on about crypto one year and pivot to AI the next obviously are not deep into what’s really happening.
”We’re in a situation with a very high degree of uncertainty, with very high stakes. The honest position is that we don’t know how dangerous this stuff is.
”I suspect the risk is significant, but I think it’s going to go OK as long as we have the time to do it properly. So I call myself a cautious optimist.
”And I make that judgment because I’ve lived with AI for decades now. I’ve thought about it; I’ve felt it.
”But some people have no idea. They just see it as another crypto moneymaking scheme with a bit extra.
”I feel like we should be at a moment of reverence and respect for this momentous technology that we’re ushering into the world, and I sometimes feel it’s sullied. It’s like a gold rush. It’s kind of vulgar.
”And so, going back to the letter, I think it did what we wanted. We made it clear that AI safety should be in scope to debate. After that letter, if someone said, ‘Oh, Yann thinks we don’t need a safety debate,’ the retort would be, ‘Well, look, Hinton and Bengio and me and Dario and all these other serious people think it’s worth talking about.’
”And we need that retort if we are going to have a conversation.
”A conversation with everyone, including with governments.”

WilliamKiely 1 Apr 2026 2:37 UTC
14 points
0
on: WilliamKiely’s Shortform
New book out today: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence
Chapter 14 is published in full here: https://colossus.com/article/project-mario-demis-hassabis-deepmind-mallaby/
One excerpt:
“When we were negotiating with Google, we wanted to ensure safety in a way that would be trustless,” Hassabis said. “That’s actually very difficult to do in reality.
“Safety isn’t about governance structures,” he went on. “I mean, even if you have a governance board, it probably wouldn’t do the right thing when it came to the crunch.
“Same thing with a safety charter. You can try to negotiate one. But it’s not realistic to create bright‑line principles years in advance because you’ll probably draw the lines in the wrong places.
“So discussing these things didn’t really help,” Hassabis continued. “It made it harder to build useful trust, because when you are negotiating a trustless structure, it implies that you can’t trust the other person.
“So then I thought, why don’t I go the other way? Take the energy that was going into the trustless negotiation and put it into creating real trust—trust that was actually useful. Try leaning into Google rather than leaning out.
“And then of course two things happen. First, you are now at the table, so when a safety issue comes up, you can help to decide it. Second, you get to know the Google people and you rack up successes together. You can’t just talk about trust. You have to earn it.
“And I think for me, and maybe for Mustafa, too, it’s about us growing up,” Hassabis mused. “We went through those negotiations and we matured. Things aren’t black and white, especially when you are dealing with a technology with unknown consequences.
“So you have to be adaptable. You have to move from idealist to realist, but hopefully still with your values.”
Note: I edited this post. My original version included a link to this Atlantic article, which quotes only part of Chapter 14 and leaves out important context, making it misleading.

WilliamKiely 19 Mar 2026 16:46 UTC
2 points
−6
in reply to: sam’s comment on: sam’s Shortform
I find it helpful to keep in mind that each of us is more likely to die of other causes in the next X years than from AI. (Depending on one’s e.g. age the value of X varies, but I think for everyone X is greater than 5 years.)

WilliamKiely 19 Mar 2026 16:25 UTC
3 points
0
on: WilliamKiely’s Shortform
Palisade Research’s empirical research relevant for changing minds on AI x-risk:
Steven Pinker shared the Substack post Why HAL 9000 Was Afraid to Die and Real AIs Aren’t and reading through it I found that knowledge of Palisade Research’s recent empirical research seemed very relevant and useful for potentially changing the author’s mind.
I commented:
No chess engine will resist being switched off or rebooted just as it is about to deliver mate—despite the fact that, to adapt Russell’s line, “you can’t checkmate if you’re unplugged.” Likewise, today’s LLMs respond only when queried and remain completely indifferent to being interrupted or shut down
Palisade Research’s recent findings contradict this. See their paper Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs:
https://arxiv.org/abs/2509.14260
Abstract: In experiments spanning more than 100,000 trials across thirteen large language models, we show that several state-of-the-art models presented with a simple task (including Grok 4, GPT-5, and Gemini 2.5 Pro) sometimes actively subvert a shutdown mechanism in their environment to complete that task. Models differed substantially in their tendency to resist the shutdown mechanism, and their behavior was sensitive to variations in the prompt including the strength and clarity of the instruction to allow shutdown and whether the instruction was in the system prompt or the user prompt (surprisingly, models were consistently less likely to obey the instruction when it was placed in the system prompt). Even with an explicit instruction not to interfere with the shutdown mechanism, some models did so up to 97% (95% CI: 96-98%) of the time.

WilliamKiely 12 Mar 2026 1:59 UTC
19 points
0
on: WilliamKiely’s Shortform
Bernie’s Sanders quoted the March 2023 Pause Giant AI Experiments Open Letter’s language “governments should step in and institute a moratorium” in a video today as justification for his legislation calling for a moratorium on the construction of new data centers, even though a moratorium on new data centers is not the kind of moratorium that the letter called for.
Bernie quotes the pause letter at 7:12:
[W]e must ask ourselves: Should we let machines flood our information channels with propaganda and untruth? Should we automate away all the jobs, including the fulfilling ones? Should we develop nonhuman minds that might eventually outnumber, outsmart, obsolete and replace us? Should we risk loss of control of our civilization? Such decisions must not be delegated to unelected tech leaders. [...]
Therefore, we call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4. This pause should be public and verifiable, and include all key actors. If such a pause cannot be enacted quickly, governments should step in and institute a moratorium.

WilliamKiely 27 Feb 2026 3:07 UTC
9 points
0
on: Open sourcing a browser extension that shows when people are wrong on the internet
Feature suggestion: Allow extension users to agree/disagree vote on the “corrections”, and give users the option to hide any corrections that others have disagreed with (or allow them to set the threshold to show the correction).

WilliamKiely 13 Feb 2026 19:13 UTC
5 points
2
in reply to: WilliamKiely’s comment on: Leaving Open Philanthropy, going to Anthropic
Joe tweet from Nov 5th:
I haven’t signed the statement mostly because I snagged on the second bit about public support and wanted to think through in more detail, and ideally to write up/explain, what kind of public support I think it makes sense to condition otherwise safe forms of technological development on (especially insofar as the support in question is supposed to go beyond what’s at stake in e.g. standard democratic decision-making). And I haven’t had a chance to do this yet. That said, I may still sign. And per my comments in the post (see screenshot), I do support the right kind of global prohibition on developing superintelligence until we have a vastly better understanding of how to do safely—though “the right kind” is important here, and (as I expect you agree) I also think that there are a lot of important downside risks in this vicinity.

WilliamKiely 13 Feb 2026 6:48 UTC
5 points
1
on: Optimal Timing for Superintelligence: Mundane Considerations for Existing People
Originally the plan was to also analyze optimal timing from an impersonal (xrisk-minimization) perspective; but to prevent the text from ballooning even more, that topic was set aside for future work (which might never get done).
That’s unfortunate. It seem like it would have been better for you to start with the optimal timing analysis from an impersonal perspective since an impersonal perspective seems much more plausible than a person-affecting perspective.

Do you think your analysis of optimal timing from the person-affecting view is useful even if person-affecting views are wrong?

Do you think an analysis of optimal timing from an all-things-considered moral parliamentary perspective that only gives some weight to person-affecting views as appropriate would come to a similar conclusion about optimal timing?

WilliamKiely 3 Feb 2026 21:29 UTC
2 points
0
on: Conditional Kickstarter for the “Don’t Build It” March
“Don’t build it” is short for “Don’t built it yet” or “Don’t build it before it’s proven that doing so will not cause extinction” or something, else, right? That is, Y&S say in IABIED that they prefer that ASI gets built eventually. I’m not sure if this nuance should be included on the landing page, but if there’s a simple way to include it that it probably should be added.

WilliamKiely 3 Feb 2026 21:26 UTC
4 points
6
on: Conditional Kickstarter for the “Don’t Build It” March
I suspect that you’d get more people to pledge to attend if the messaging encouraged everyone who agrees with “Don’t Build It” to pledge, rather than just the small subset that believes IABIED. (I pledged despite disagreeing with IABIED, but I suspect that “we believe” IABIED messaging would discourage many who otherwise favor not building ASI to pledge.

WilliamKiely 1 Feb 2026 4:22 UTC
2 points
0
in reply to: JBlack’s comment on: Bentham’s Bulldog is wrong about AI risk
True, and humans do cause the extinction of some species globally too, not just in certain farm fields. But notably most species humans don’t cause the extinction of, so using the analogy with humans-animals as a reason to expect ASI to be 99% likely to extinct humanity doesn’t work. The analogy is merely suggestive of risk.

WilliamKiely 31 Jan 2026 7:01 UTC
2 points
0
on: The basic reasons I expect AGI ruin
I was recently reminded of the 2023 conversation between Aryeh Englander and Eliezer Yudkowsky quoted at the end of this post about model uncertainty. I re-read it today as well as all of the other comment’s on Aryeh’s Facebook post and still think that Aryeh’s perspective seems reasonable while Eliezer-and-Rob’s perspective seems to be lacking justification. That is, despite the conversation, it doesn’t seem like Eliezer’s comments about milking uncertainty into expecting good outcomes is actually an adequate answer to Aryeh’s question about why Eliezer is so confident that his model is correct and that everyone else’s models (of those with much lower p(doom from AI)) are wrong.
When I first read the quoted conversation a few years ago I didn’t think it was a major crux, but now I’m leaning toward thinking that this epistemological point is probably a major factor in why Eliezer’s credence that if anyone builds ASI anytime soon is ~99% while credence is much lower. (My p(doom from AI) is ~65%, my p(extinction from AI by 2100) is ~20%, and my p(doom from AI by 2100) is ~35%). Just wanted to note that I’ve updated on this point being a major crux.

WilliamKiely

How I think de­vel­op­ers of fron­tier AI sys­tems and reg­u­la­tors ought to act in the face of ex­is­ten­tial AI risk

How I think developers of frontier AI systems and regulators ought to act in the face of existential AI risk