dockerfile/examples/omnivore/content-fetch/readabilityjs/test/test-pages/lesswrong/distiller.html

445 lines
122 KiB
HTML
Raw Normal View History

2024-03-15 14:52:38 +08:00
<div><p>
(Cross-posted from <span><span><span><a href="https://joecarlsmith.com/2023/05/08/predictable-updating-about-ai-risk">my website</a></span></span></span>. Podcast version <span><span><span><a href="https://www.buzzsprout.com/2034731/12809255-predictable-updating-about-ai-risk">here</a></span></span></span>, or search &#34;Joe Carlsmith Audio&#34; on your podcast app.)
</p><blockquote><p>
<i>&#34;This present moment used to be the unimaginable future.&#34;</i>
</p><p>
<i>- Stewart Brand</i>
</p></blockquote><h2>
1. Introduction
</h2><p>
Heres a pattern you may have noticed. A new frontier AI, like GPT-4, gets released. People play with it. Its better than the previous AIs, and many people are impressed. And as a result, many people who werent worried about existential risk from misaligned AI (hereafter: “AI risk”) get much more worried.<span role="doc-noteref"><sup><span><span><a href="#fn0g9sn95iiqe">[1]</a></span></span></sup></span>
</p><p>
Now, if these people didnt expect AI to get so much better so soon, such a pattern can make sense. And so, too, if they got other unexpected evidence for AI risk for example, concerned experts <span><span><span><a href="https://futureoflife.org/open-letter/pause-giant-ai-experiments/">signing letters</a></span></span></span> and <span><span><span><a href="https://www.nytimes.com/2023/05/01/technology/ai-google-chatbot-engineer-quits-hinton.html">quitting their jobs</a></span></span></span>.
</p><p>
But if youre a good Bayesian, and you currently put low probability on existential catastrophe from misaligned AI (hereafter: “AI doom”), you probably shouldnt be able to predict that this pattern will happen to you in the future.<span role="doc-noteref"><sup><span><span><a href="#fn048nj5dhr9di">[2]</a></span></span></sup></span> When GPT-5 comes out, for example, it probably shouldnt be the case that your probability on doom goes up a bunch. Similarly, it probably shouldnt be the case that if you could see, now, the sorts of AI systems well have in 2030, or 2050, that youd get a lot more worried about doom than you are now.
</p><p>
But I worry that were going to see this pattern anyway. Indeed, Ive seen it myself. Im working on fixing the problem. And I think we, as a collective discourse, should try to fix it, too. In particular: I think were in a position to predict, now, that AI is going to get a lot better in the coming years. I think we should worry, now, accordingly, without having to see these much-better AIs up close. If we do this right, then in expectation, when we confront GPT-5 (or GPT-6, or <span><span><span><a href="https://agentgpt.reworkd.ai/">Agent-GPT</a></span></span></span>-8, or <span><span><span><a href="https://decrypt.co/126122/meet-chaos-gpt-ai-tool-destroy-humanity/">Chaos-GPT</a></span></span></span>-10) in the flesh, in all the concreteness and detail and not-a-game-ness of the real world, well be just as scared as we are now.
</p><p>
This essay is about what “doing this right” looks like. In particular: part of what happens, when you meet something in the flesh, is that it “seems more real” at a gut level. So the essay is partly a reflection on the epistemology of guts: of visceral vs. abstract; “up close” vs. “far away.” My views on this have changed over the years: and in particular, I now put less weight on my guts (comparatively skeptical) views about doom.
</p><p>
But the essay is also about grokking some basic Bayesianism about future evidence, dispelling a common misconception about it (namely: that directional updates shouldnt be predictable in general), and pointing at some of the constraints it places on our beliefs over time, especially with respect to stuff were currently skeptical or dismissive about. For example, at least in theory: you should never think it &gt;50% that your credence on something will later double; never &gt;10% that it will later 10x, and so forth. So if youre currently e.g. 1% or less on AI doom, you should think its less than 50% likely that youll ever be at 2%; less than 10% likely that youll ever be at 10%, and so on. And if your credence is very small, or if youre acting dismissive, you should be very confident youll never end up worried. Are you?
</p><p>
I also discuss when, exactly, its problematic to update in predictable directions. My sense is that generally, you should expect to update in the direction of the <i>truth</i> as the evidence comes in; and thus, that people who think AI doom unlikely should expect to feel <i>less worried</i> as time goes on (such that consistently getting more worried is a red flag). But in the case of AI risk, I think at least some non-crazy views should actually expect to get <i>more worried</i> over time, even while being fairly non-worried now. In particular, if you think you face a small risk conditional on something likely-but-not-certain (for example, AGI getting developed by blah date), you can sometimes expect to update towards facing the risk, and thus towards greater worry, before you update towards being safe. But there are still limits to how much more worried you can predictably end up.
</p><p>
Importantly, none of this is meant to encourage consistency with respect to views you held in the past, at the expense of reasonableness in the present or future. If you said .1% last year, and youre at 10% now (or if you hit 90% when you see GPT-6): well, better to just say “<span><span><a href="/posts/wCqfCLs8z5Qw4GbKS/the-importance-of-saying-oops">oops</a></span></span>.” Indeed, Ive been saying “oops” myself about various things. And more generally, applying basic Bayesianism in practice takes lots of taste. But faced with predictable progress towards advanced but mostly-still-abstract-for-now AI, I think its good to keep in mind.
</p><p>
I close with some thoughts on how we will each look back on what we did, or didnt do, during the lead-up to AGI, once the truth about the risks is made plain.
</p><p>
<i>Thanks to Katja Grace for extensive discussion and inspiration. See also citations in the main text and footnotes for specific points and examples that originated with Katja. And thanks also to Leopold Aschenbrenner for comments. Some of my thinking and writing on this topic occurred in the context of my work for Open Philanthropy, but Im speaking only for myself and not for my employer.</i>
</p><h2>
2. Sometimes predictably-real stuff doesnt feel real yet
</h2><blockquote><p>
<i>&#34;Every year without knowing it I have passed the day</i>
</p><p>
<i>When the last fires will wave to me</i>
</p><p>
<i>And the silence will set out</i>
</p><p>
<i>Tireless traveler</i>
</p><p>
<i>Like the beam of a lightless star&#34;</i>
</p><p>
<i>-</i> <span><span><span><a href="https://merwinconservancy.org/2020/03/poem-of-the-week-for-the-anniversary-of-my-death-2/"><i>W.S. Merwin</i></a></span></span></span><i>, “For the Anniversary of My Death”</i>
</p></blockquote><p>
I first heard about AI risk in 2013. I was at a picnic-like thing, talking with someone from the Future of Humanity Institute. He mentioned AI risk. I laughed and said something about “like in the movie <i>I, Robot</i>?” He didnt laugh.
</p><p>
Later, I talked with more people, and read Bostroms <span><span><span><a href="https://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111/ref=tmm_hrd_swatch_0?_encoding=UTF8&amp;qid=&amp;sr=">Superintelligence</a></span></span></span>. I had questions, but the argument seemed strong enough to take seriously. And at an intellectual level, the risk at stake seemed like a big deal.
</p><p>
At an emotional level, though, it didnt <i>feel real</i>. It felt, rather, like an abstraction. I had trouble imagining what a real-world AGI would be like, or how it would kill me. When I thought about nuclear war, I imagined flames and charred cities and poisoned ash and starvation. When I thought about biorisk, I imagined sores and coughing blood and hazmat suits and body bags. When I thought about AI risk, I imagined, um … nano-bots? I wasnt good at imagining nano-bots.
</p><p>
I remember looking at some farmland out the window of a bus, and wondering: am I supposed to think that this will all be compute clusters or something? I remember looking at a church and thinking: am I supposed to imagine robots tearing this church apart? I remember a late night at the Future of Humanity Institute office (I ended up working there in 2017-18), asking someone passing through the kitchen how to imagine the AI killing us; he turned to me, pale in the fluorescent light, and said “whirling knives.”
</p><p>
Whirling knives? <span><span><span><a href="https://twitter.com/ESYudkowsky/status/1438198189782290433">Diamondoid bacteria</a></span></span></span>? Relentless references to paper-clips, or “tiny molecular squiggles”? Ive written, elsewhere, about <span><span><span><a href="https://joecarlsmith.com/2021/01/31/believing-in-things-you-cannot-see#iv-realization-vs-belief">the “unreality” of futurism</a></span></span></span>. AI risk had a lot of that for me.
</p><p>
That is, I wasnt <i>viscerally worried</i>. I had the concepts. But I didnt have the “actually” part. And I wasnt alone. As I started working on the topic more seriously, I met some people who were viscerally freaked-out, depressed, and so on whether for good or ill. But I met lots of people who werent, and not because they were protecting their mental health or something (or at least, not very consciously). Rather, their head was convinced, but not their gut. Their gut still expected, you know, <span><span><span><a href="https://www.cold-takes.com/this-cant-go-on/">normality</a></span></span></span>.
</p><p>
At the time, I thought this was an important signal about the epistemic situation. Your gut can be smarter than your head. If your gut isnt on board, maybe your head should be more skeptical. And having your gut on board with whatever youre doing seems good from other angles, too.<span role="doc-noteref"><sup><span><span><a href="#fnqnj4snix8n">[3]</a></span></span></sup></span> I spent time trying to resolve the tension. I made progress, but didnt wholly sync up. To this day, nano-bots and dyson spheres and the word “singularity” still land in an abstract part of my mind the part devoted to a certain kind of conversation, rather than to, like, the dirty car I can see outside my window, and the tufts of grass by the chain-link fence.
</p><p>
I still think that your gut can be an important signal, and that if you find yourself saying that you believe blah, but youre not <span><span><span><a href="https://www.econlib.org/archives/2016/01/the_invisible_t.html">feeling</a></span></span></span> or acting like it, you should stop and wonder. And sometimes, people/ideas that try to get you to not listen to your gut are trying (whether intentionally or not) to bypass important defenses. I am not, in what follows, trying to tell you to throw your gut away. And to the extent I am questioning your gut: please, by all means, be more-than-usually wary. Still, though, and speaking personally: Ive come to put less stock than I used to in my guts Bayesian virtue with respect to AI. I want to talk a bit about why.
</p><h2>
3. When guts go wrong
</h2><blockquote><p>
<i>&#34;Then I will no longer</i>
</p><p>
<i>Find myself in life as in a strange garment</i>
</p><p>
<i>Surprised at the earth…&#34;</i>
</p><p>
<i>-</i><span><span><span><a href="https://merwinconservancy.org/2020/03/poem-of-the-week-for-the-anniversary-of-my-death-2/"><i>W.S. Merwin</i></a></span></span></span><i>, “For the Anniversary of My Death”</i>
</p></blockquote><p>
Part of this is reflection on examples where guts go wrong, especially about the future. There are lots of candidates. Indeed, depending on how sharply we distinguish between your “system 1” and your gut, a lot of the <span><span><span><a href="https://thedecisionlab.com/biases">biases literature</a></span></span></span> can be read as anti-gut, and a lot of early rationalism as trying to compensate. My interest in head-gut agreement was partly about trying to avoid overcorrection. But there is, indeed, something to be corrected. Here are two examples that seem relevant to predictable updating.
</p><h3>
3.1 War
</h3><blockquote><p>
<i>“Abstraction is a thing about your mind, and not the world… Saying that AI risk is abstract is like saying that World War II is abstract, because its 1935 and hasnt happened yet. If it happens, it will be very concrete and bad. It will be the worst thing that has ever happened.”</i>
</p></blockquote><p>
I think Katjas war example is instructive. Consider some young men heading off to war. Theres a trope, here, about how, when the war is just starting, some men sign up excitedly, with dreams of glory and honor. Then, later, they hit the gritty reality: trenches, swamps, villages burning, friends gasping and gurgling as they die. Ken Burns <span><span><span><a href="https://www.pbs.org/kenburns/the-vietnam-war/">Vietnam War documentary</a></span></span></span> has some examples. See also “<span><span><span><a href="https://en.wikipedia.org/wiki/Born_on_the_Fourth_of_July_(film)">Born on the Fourth of July</a></span></span></span>.” The soldiers return, if they return, with a very different picture of war. “<span><span><span><a href="https://en.wikipedia.org/wiki/Dulce_et_Decorum_est">In all my dreams before my helpless sight/ He plunges at me, guttering, choking, drowning</a></span></span></span>…”
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kkygldeur6b96qqyidee" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kkygldeur6b96qqyidee 600w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kymdz6wnfzlkpibrhvs4 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/spvydfruwc6lwytxlorb 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zn5bde6lkc3zgocekwhg 462w"/><p>
<i>Stretcher bearers in World War I (source</i> <a href="https://commons.wikimedia.org/wiki/File:Stretcher_bearers_Passchendaele_August_1917.jpg"><i>here</i></a><i>)</i>
</p><p>
Now, a part of this is that their initial picture was <i>wrong</i>. But also, sometimes, its that their initial picture was <i>abstract</i>. Maybe, if youd asked them ahead of time, theyd have said “oh yeah, I expect the trenches to be very unpleasant, and that I will likely have to watch some of my friends die.” But their gut didnt expect this or, not hard enough. Surrounded, when they set out, by flags and smiling family members and crisp uniforms, its hard to think, too, of flies in the eyes of rotting corpses; or trench-foot, and the taste of mustard gas. And anyway, especially if youre heading into a very new context, its often hard to know the specifics ahead of time, and any sufficiently-concrete image is predictably wrong.
</p><p>
I worry that were heading off to something similar, epistemically, to a new war, with respect to AI risk.<span role="doc-noteref"><sup><span><span><a href="#fnz4yjec5asvf">[4]</a></span></span></sup></span> Not: happily, and with dreams of glory. But still: abstractly. Were trying to orient intellectually, and to do what makes sense. But we arent in connection with what it will actually be like, if AI kicks off hard, and the doomers are right. Which isnt to say it will be trench foot and mustard gas. Indeed, even if things go horribly wrong eventually, it might actually be awesome in lots of ways for a while (even if also: extremely strange). But whatever it will be, will be a specific but very-different-from-now thing. Guts arent good at that. So its not, actually, all that surprising if youre not as viscerally worried as your explicit beliefs would imply.
</p><h3>
3.2 Death
</h3><blockquote><p>
<i>&#34;And who by fire, who by water</i>
</p><p>
<i>Who in the sunshine, who in the night time</i>
</p><p>
<i>Who by high ordeal, who by common trial…&#34;</i>
</p></blockquote><p>
Another famous example here is death. No one knows the date or hour. But we know: someday.<span role="doc-noteref"><sup><span><span><a href="#fn1dgio52qsrw">[5]</a></span></span></sup></span> Right? Well, sort of. We know in the abstract. We know, but dont always realize. And then sometimes we do, and some vista opens. We reel in some new nothingness. Something burns with new preciousness and urgency.
</p><p>
And sometimes this happens, specifically, when “someday, somehow” becomes “soon, like this.” When the doctor tells you: you, by avalanche. You, by powder. The month of May. Slow decay. Suddenly, when youre actually looking at the scans, when youre hearing estimates in months, you learn fresh who is calling; and despite having always known, some sort of “update” happens. Did the gut not fully believe? Ones own death, after all, is <span><span><span><a href="https://joecarlsmith.com/2021/01/31/believing-in-things-you-cannot-see">hard to see</a></span></span></span>.
</p><p>
Ive <span><span><span><a href="https://joecarlsmith.com/2020/12/06/thoughts-on-being-mortal#iii">written about this before</a></span></span></span>. Tim McGraw has a song about the scans thing. “<span><span><span><a href="https://www.youtube.com/watch?v=_9TShlMkQnc">Live like you were dying</a></span></span></span>.” Im trying. Im trying to think ahead to that undiscovered hospital. Im trying to think about what I will take myself to have learned, when I walk out into the parking lot, with only months to live. Im trying to learn it now instead.
</p><p>
Really, this is about predictable updating. The nudge in McGraws title youre already dying is Bayesian. You shouldnt need the scans. If you know, now, what youll learn later, you can learn it now, too. Death teaches unusually predictable lessons about fleetingness, beauty, love. And unusually important lessons, too. Bayes bites, here, with special gravity. But theres some sort of gut problem. The question is how to learn hard enough, and in advance. “<span><span><span><a href="https://www.brainyquote.com/quotes/henry_david_thoreau_107665">And not, when I come to die, to discover that I have not lived</a></span></span></span>.”
</p><p>
Importantly, though: if your gut thinks youre not going to die, its not actually much evidence. Has your gut been keeping up with the longevity literature? Does it have opinions about cryopreservation? Futurism aside, the guts skepticism, here, is an old mistake. And we have practices. Go <span><span><span><a href="https://en.wikipedia.org/wiki/Ash_Wednesday">smear some ashes on your forehead</a></span></span></span>. Go <span><span><span><a href="https://en.wikipedia.org/wiki/Sky_burial">watch some birds eat a corpse</a></span></span></span>. Go put some fruit on the <span><span><span><a href="https://en.wikipedia.org/wiki/Ofrenda">ofrenda</a></span></span></span>, or some flowers on your grandfathers grave. <span><span><span><a href="https://joecarlsmith.com/2021/01/31/believing-in-things-you-cannot-see">Realization is an art distinct from belief</a></span></span></span>. Sometimes, you already know. Religion, they say, is remembering.
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/shkucq5c0cvpdyoilxko" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/shkucq5c0cvpdyoilxko 800w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zaqb75irgsve2gjrgzqy 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/jifrboveteqktdgadtzq 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/vupbmg65qbzwnrptgyyv 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/chpgceqfmfeilzwalbfu 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mv5ixhkytnydgxgctzlv 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/bxij00lcawaqkcc12rz0 722w"/><p><i>Tibetan sky burial. (Source</i> <a href="https://commons.wikimedia.org/wiki/File:Bundesarchiv_Bild_135-S-12-50-06,_Tibetexpedition,_Ragyapa,_Geier.jpg"><i>here</i></a><i>.)</i>
</p><h2>
4. Noticing your non-confusion
</h2><p>
So these are some examples where “but my gut isnt in a very visceral relationship with blah” just isnt a very strong signal that blah is false. But I also want to flag some more directly AI related places where I think something gut-related has been messing up, for me.
</p><h3>
4.1 LLMs
</h3><div>ChatGPT caused a lot of new attention to LLMs, and to AI progress in general. <span>But depending on what you count: we had scaling laws for deep learning back in</span> <span><span><span><a href="https://arxiv.org/abs/1712.00409"><span>2017</span></a></span></span></span><span>, or at least</span> <span><span><span><a href="https://arxiv.org/abs/2001.08361"><span>2020</span></a></span></span></span><span>. I know people who were really paying attention; who really saw it; who really bet.</span> And I was trying to pay attention, too. I knew more than many about what was happening. And in a sense, my explicit beliefs werent, and should not have been, very surprised by the most recent round of LLMs. I was not a “shallow patterns” guy. I didnt have any specific stories about the curves bending. I expected, in the abstract, that the LLMs would improve fast.
</div><p>
But still: when I first played with one of the most recent round of models, my gut did a bunch of updating, in the direction of “oh, actually,” and “real deal,” and “fire alarm.” Some part of me was still surprised.
</p><p>
Indeed, noticing my gut (if not my head) getting surprised at various points over the past few years, Ive realized that my gut can have some pretty silly beliefs about AI, and/or can fail to connect fairly obvious dots. For example, when I first started thinking about AI, I think some part of me failed to imagine that eventually, if AIs got smart enough, we could just <i>talk to them</i>, and that they would just <i>understand what we were saying</i>, and that interacting with them wouldnt necessarily be some hyper-precise coding thing. I had spoken to Siri. Obviously, that didnt count. Then, one day, I spoke, with my voice, to a somewhat-smarter AI, and it responded in a very human-sounding voice, and it was much more like talking on the phone, and some sort of update happened.
</p><p>
Similarly: I think that in the past, I failed to imagine what the visual experience of interacting with an actually-smart AI would be like. Obviously, I knew about robots; HALs red stare; typing commands into a terminal; texting. But somehow, old talk of AGI didnt conjure this for me. Im not sure what it conjured. Something about brains in boxes, except: laptops? I think it wasnt much of anything, really. I think it was just a blank. After all, this isnt <i>sci-fi</i>. So it must not be like anything youd see in sci-fi, either, including strains aimed at realism. People, were talking about the <i>real future</i>, which means something <i>unimaginable</i>, hence fiction to the imagination, hence nothingness. “The future that can be named is not the true future.” Right?
</p><p>
Wrong. “Named super specifically” is more plausible, but even wariness of specificity can mislead: sometimes, even the specifics are pretty obvious. I <i>had seen</i> Siri, and chat bots. What sort of fog was I artificially imposing on everything? What was so hard about imagining Siri, but smarter? Now, it feels like “oh, duh.” And certain future experiences feel more concrete, too. It now feels like: oh, right, lots of future AIs will probably have extremely compelling and expressive <span><span><span><a href="https://replika.com/">digital human avatars</a></span></span></span>. Eventually (soon?), theyll probably be able to look just like (super-hot, super-charismatic) humans on zoom calls. What did I think it would be, R2D2?
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ckxrbgmv82cmglhia9ga" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ckxrbgmv82cmglhia9ga 793w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/nhkv2otstngfqbqrnrvf 232w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/sfatvn35ua9zbtn6cf9k 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/stihzpefyj3ncv4n8wsd 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/axbxadmiet8vg9whnksi 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/geetxsphhzcs2gwiqnmm 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/uqg443xbl8w36jlzic1z 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/e9iy8gcgjvvrrxrhyidk 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/boz9le1jdcq2gbm6xpjr 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/vksdqn2ggbhpe8tpeugr 1108w"/><p>
“Oh, duh” is never great news, epistemically. But its interestingly <i>different</i> news than “<span><span><span><a href="https://www.readthesequences.com/Noticing-Confusion-Sequence">noticing your confusion</a></span></span></span>,” or being straightforwardly surprised. Its more like: noticing that at some level, you were tracking this already. You had the pieces. Maybe, even, its just like you wouldve said, if youd been asked, or thought about it even a little. Maybe, even, you literally said, in the past, that it would be this way. Just: you said it with your head, and your gut was silent.
</p><p>
I mentioned this dynamic to Trevor Levin, and he said something about “noticing your non-confusion.” I think its a good term, and a useful skill. Of course, you can still update upon seeing stuff that you expected to see, if you werent <i>certain</i> youd see it. But if it feels like your head is unconfused, but your gut is updating from “its probably fake somehow” to “oh shit its actually real,” then you probably had information your gut was failing to use.
</p><h3>
4.2 Simulations
</h3><p>
Ill give another maybe-distracting example here. Last year, I spent some time thinking about <span><span><span><a href="https://jc.gatspress.com/pdf/simulation_arguments_revised.pdf">whether we live in a computer simulation</a></span></span></span>. Its a strange topic, but my head takes the basic argument pretty seriously. My gut, though, generally thinks its fake somehow, and forgets about it easily.
</p><p>
I remember a conversation I had with a friend sometime last year. He said something like: “you know, pretty soon, all sorts of intelligent agents on earth are going to be living in simulations.” I nodded or something. Its like how: if the scientists are actually <i>putting</i> peoples brains in vats, its harder to stamp your foot and say “no way.” We moved on.
</p><p>
Then, in early April, this paper came out: “<span><span><span><a href="https://arxiv.org/pdf/2304.03442.pdf">Generative Agents: Interactive Simulacra of Human Behavior</a></span></span></span>.” They put 25 artificial agents into an environment similar to The Sims, and had them interact, including via e.g. hosting a valentines day party.<span role="doc-noteref"><sup><span><span><a href="#fn4m2cfburfm7">[6]</a></span></span></sup></span> Heres the picture from the paper:
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/q6mirdkqnte3mtz7tmjt" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/q6mirdkqnte3mtz7tmjt 1024w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/llnsldrmajaxuvyehy7e 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/osectulgs4qjctsror1r 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/m7jurokdkjrsvjzndks9 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/cr1md9m8f2hxfbzthmhp 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ki3fbycj28tlpzmniqor 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/pyh1dv4xp4ecmp0zzlkr 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/lgmi3qpvqtobziipibay 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/adjhthooygwkpqihboht 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rx5yxljrmgvx1p6hwzuw 1402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rvkziwrga9fe5n09ashy 1420w"/><p><i>From</i> <span><span><span><a href="https://arxiv.org/pdf/2304.03442.pdf"><i>here</i></a></span></span></span><i>.</i>
</p><p>
I opened this paper, read the beginning, looked at this picture, and felt my gut update towards being in a sim. But: cmon now, gut! What sort of probability would I have put, last year, on “I will, in the future, see vaguely-smart artificial agents put into a vaguely-human simulated environment”? Very high. My friend had literally said as much to me months earlier, and I did not doubt. Indeed, whats even the important difference between this paper and AlphaStar, or the original Sims?<span role="doc-noteref"><sup><span><span><a href="#fn4y75ebmy7eh">[7]</a></span></span></sup></span> How smart the models are? The fact that its cute and human-like? My gut lost points, here.
</p><p>
Its an avoidable mistake. Im trying to stop making it.
</p><p>
I worry that were in for a lot of dynamics like this. How seriously, for example, are you taking the possibility that future AIs will be sentient? Well, heres a mistake to not make: updating a lot once the AIs are using charismatic human avatars, or once they can argue for their sentience as convincingly as a human. Predict it now, people. Update now.
</p><h3>
4.3 “Its just like they said”
</h3><p>
I dont, often, have nightmares about AI risk. But I had one a few months ago. In it, I was at a roll-out of some new AI system. It was a big event, and there were lots of people. The AI was unveiled. Somehow, it immediately wrote each one of us some kind of hyper-specific, individualized message, requiring a level of fine-grained knowledge and predictive ability that was totally out of the question for any familiar intelligence. I read my message and felt some cold and electric bolt, some recognition. I thought to myself: “its just like they said.” I looked around me, and the room was in chaos. Everything was flying apart, in all directions. I dont remember what happened after that. 
</p><p>
“Just like they said.” Whos they? Presumably, the AI worriers. The ones who think that superintelligence is not a fantasy or a discussion-on-twitter, but an actual thing we are on track to do with our computers, and which will cut through our world like butter if we get it wrong.
</p><p>
But wait: arent I an AI worrier? More than many, at least. But dreams, they say, are partly the guts domain. Perhaps the “they,” here, was partly my own explicit models. Ask me in the waking world: “will superintelligence be terrifying?” Yes, of course, who could doubt. But ask in my dreams instead, and I need to see it up close. I need to read the message. Only then will my gut go cold: “Oh, shit, its just like they said.”
</p><p>
Ive had this feeling a few times in the past few months. I remember, a few years ago, making a simple model of AI timelines with a colleague. We used a concept called “wake-up,” indicating the point where the world realized what was happening with AI and started to take it seriously. I think that if, at that point, we couldve seen what things would be like in 2023, we wouldve said something like: “yeah, that” (though: theres a ton more waking up to do, so future wake-ups might end up better candidates).
</p><p>
Similarly, “they” have worried for ages about triggering or exacerbating “race dynamics” in AI. Then, in recent months, Google went into a “<span><span><span><a href="https://www.nytimes.com/2022/12/21/technology/ai-chatgpt-google-search.html">Code Red</a></span></span></span>” about AI, and the CEO of Microsoft came out and just said straight up: “<span><span><span><a href="https://www.businesstoday.in/technology/news/story/the-race-starts-today-microsoft-officially-brings-chatgpt-ai-to-bing-and-edge-browser-369453-2023-02-08">the race starts today</a></span></span></span>.”
</p><p>
“They” have worried about AIs being crazy alien minds that we dont understand. Then, in February, we got to see, briefly, the rampaging strangeness of a good Bing including all sorts of <span><span><span><a href="https://time.com/6256529/bing-openai-chatgpt-danger-alignment/">deception and manipulation and blackmail</a></span></span></span>, which I <span><span><span><a href="https://www.cold-takes.com/what-does-bing-chat-tell-us-about-ai-risk/">dont actually think is the centrally worrying kind</a></span></span></span>, but which doesnt exactly seem like good news, either.
</p><p>
“They” have worried about agents, and about AIs running wild on the internet, and about humans not exactly helping with that. Now we have <span><span><span><a href="https://en.wikipedia.org/wiki/Auto-GPT">Auto-GPT</a></span></span></span>, and <span><span><span><a href="https://decrypt.co/126122/meet-chaos-gpt-ai-tool-destroy-humanity">Chaos-GPT</a></span></span></span>, and I open up my browser and I see stuff like <span><span><span><a href="https://agentgpt.reworkd.ai/">this</a></span></span></span>:
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/jpmal0zipviuvk3lh7sg" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/jpmal0zipviuvk3lh7sg 1024w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/t8hetjm6w9ytpgy8w88y 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/g4jx6qfqljr5pfhk8nxw 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ftts0hfa6vho5bncqnnn 1536w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/pnwqbtx5zxtg3pxoayib 2048w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/hzuxg4hennshvtf0cfvn 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/t3ybknd10bd72j057obw 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xrtcyqsvpg7b5kdmk59o 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/awyiwkdkoqje6wgbekn1 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/aovgrdwbwi69qj8vbjcb 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/jvjdpzgxpihru8qcpoau 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/eykrqp9vhbqzwhosm4qj 1402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/acilnuziiquzylpaopwr 1702w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ipnpetuxchsejqv1qvhv 2002w"/><p><i>Not the pixels I wanted to be seeing at this point in my life.</i>
</p><p>
Now, I dont want to litigate, here, exactly who “called” what (or: created what<span role="doc-noteref"><sup><span><span><a href="#fn5gj8wfkheeo">[8]</a></span></span></sup></span>), and how hard, and how much of an update all this stuff should be. And I think some things for example, the worlds sympathy towards concern about risks from AI have surprised some doomers, however marginally, in the direction of optimism. But as someone who has been thinking a lot about AI risk for more than five years, the past six months or so have felt like a lot of movement from abstract to concrete, from “thats what the model says” to “oh shit here we are.” And my gut has gotten more worried.
</p><p>
Can this sort of increased worry be Bayesian? Maybe. I suspect, though, that Ive just been messing up. Lets look at the dynamics in more detail.
</p><h2>
5. Smelling the mustard gas
</h2><blockquote><p>
<i>&#34;Men marched asleep…</i>
</p><p>
<i>All went lame, all blind.&#34;</i>
</p></blockquote><p>
Its sometimes thought that, as a Bayesian, you shouldnt be able to predict which direction youll update in the future.<span role="doc-noteref"><sup><span><span><a href="#fnh3ar4lqw084">[9]</a></span></span></sup></span>That is, if youre about to get some new evidence about <i>p</i>, you shouldnt be able to predict whether this evidence will move your credence on <i>p</i> higher or lower. Otherwise, the thought goes, you could “price in” that evidence now, by moving your credence in the predicted direction.
</p><p>
But this is wrong.<span role="doc-noteref"><sup><span><span><a href="#fn4ohd2xn7yql">[10]</a></span></span></sup></span> Consider a simple example. Suppose youre at 99% that Trump won the election. Youre about to open the newspaper that will tell you for sure. Here, you should be at 99% that youre about to increase your credence on Trump winning: specifically, up to 100%. Its a very predictable update.
</p><p>
So why cant you price it in? Because theres a 1% chance that youre about to lower your confidence in Trump winning <i>by a lot more</i>: specifically, down to 0%. That is, in <i>expectation</i>, your confidence in Trump winning will remain the same.<span role="doc-noteref"><sup><span><span><a href="#fnc1l5t3jttlk">[11]</a></span></span></sup></span>And its the expectation of your future update that Bayesian binds.
</p><p>
To understand this more visually, lets use a slightly more complicated example. Suppose youre currently at 80% that GPT-6 is going to be “scary smart,” whatever that means to you. And suppose that, conditional on GPT-6 being scary smart, your probability on AI doom is 50%; and conditional on GPT-6 not being scary smart, your probability on AI doom is 10%. So your credence looks like this:
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/nra6meskaht8wdx1ig3q" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/tv20xlvpjkpjwxdwv0ub 942w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/lvlh3bnro7dpwogepzyo 276w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/dq5jtswyqsjhk6y7bpjx 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/hpyup2d5e5wshq2gxp6b 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/afdb2urasvsfdidfzhts 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ozzybtunbhswka5gic40 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ycrmtyxtriv8em266r0z 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kvz3h40wblsqw7ekj2ef 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/libzhbgptdorqbqscxzo 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/fdvw3d3simncrdgnnabg 1082w"/><p>
Now, whats your overall p(doom)? Well, its:
</p><blockquote><p>
(probability that GPT-6 is scary smart * probability of doom conditional on GPT-6 being scary smart) + (probability that GPT-6 isnt scary smart * probability of doom conditional on GPT-6 not being scary smart)
</p></blockquote><p>
That is, in this case, 42%.<span role="doc-noteref"><sup><span><span><a href="#fno3exn1kvfm">[12]</a></span></span></sup></span>
</p><p>
But now we can see a possible format for a gut-problem mistake. In particular: suppose that I ask you, right now, surrounded by flags and crisp uniforms, about the probability of doom. You query your gut, and it smells no mustard gas. So you give an answer that doesnt smell much mustard gas, either. Lets say, 10%. And lets say you dont really break things down into: OK, how much mustard gas do I smell conditional on GPT-6 being scary smart, vs. not, and what are my probabilities on that.<span role="doc-noteref"><sup><span><span><a href="#fnt4taauzgccn">[13]</a></span></span></sup></span> Rather, your model is an undifferentiated mass:
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rbjkdeyi0wb0zz4htv2f" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rbjkdeyi0wb0zz4htv2f 970w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/jyz00tksh6fzmsrrwzte 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mpiynbtl2td13dtrivzi 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qwppd7rqhbuwdknpkojj 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/spztwjkpmryb8ikhrpw8 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zkbqpvivirrtz7prkonl 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/m0vr9aby7wm9ptxoll78 722w"/><p>
Or maybe you do try to break things down like that, but the waft of the gas fades with all the distance. GPT-6 is far away, behind some fog. Still: you guess, with your head, and without your gut participating, that p(doom) is indeed a bit higher conditional on GPT-6 being scary smart, what with the update towards “short timelines.” Lets say, 20%; and 10% otherwise. So maybe your overall p(doom), given 80% on the abstract idea of GPT-6 being scary smart, is 18%.
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rfl34mdx3i4zcz0a4i7m" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rfl34mdx3i4zcz0a4i7m 957w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/nvfxoxgencxt15pfc6nr 280w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/fujyddttrzos4sjsnr9l 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kfnjiozyk6zvtp2z8w5l 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/noa75ytmstppqbkogzvq 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/tfjun4wdy5kyigpqeqem 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ksuqs7emajlxupcj7cen 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xbsn1amesfysfzuepwgr 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xueudncpvsogtqrrreuz 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/cgrrdu5bpna1eeokyttx 1208w"/><p>
But actually, lets say, if you could see a “scary smart” GPT-6 model right now, you would freak out way harder. You would be able to smell the gas up close, that bitter tang. Your gut would get some message, and come alive, and start participating in the exercise. “<i>That thing</i>,” your gut might say, “is <i>scary</i>. Im at 50% on doom, now.”
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/axfsbba4ueihtoyiao24" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/axfsbba4ueihtoyiao24 916w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/uuix302lquapruhwnzhn 268w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xuhwmnlcjpby3idnqirb 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ozhynvnhqgxoqfufmkbh 1375w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xxon5usbt8paffsjqvnl 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mrfaxqpbunqytyuqnoyj 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/eaj9dbbvdrtlonvfy2lr 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/dypci8n8pncoyjj84hxo 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/k2wdfefrlolu7gwmivfl 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/fsfihwtucf89yzbmjsep 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mze0y4fdbcs3amdlga41 1402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/dwtad6aknstkayhrmglw 1414w"/><p>
Thus, you end up inconsistent, and dutch-bookable (at least in principle setting aside issues re: betting on doom). Suppose I ask you, now, to agree to sell me a “pays out $100 conditional on doom” ticket for $30 (lets assume this can actually pay out), conditional on GPT-6 being scary smart. Youre only at 20% doom in such a world, so you predict that such a ticket will only be worth $20 to you if this deal is ever triggered, so you agree. But actually, when we get to that world, your gut freaks out, and you end up at 50% doom, and that ticket is now worth $50 to you, but youre selling it for $30. Plus, maybe now youre regretting other things. Like some of those tweets. And how much alignment work you did.
</p><p>
As indicated above, I think Ive made mistakes in this vein. In particular: a few years back, I wrote a <span><span><span><a href="https://arxiv.org/pdf/2206.13353.pdf">report about AI risk</a></span></span></span>, where I put the probability of doom by 2070 at 5%. Fairly quickly after releasing the report, though, I realized that this number was too low.\<span role="doc-noteref"><sup><span><span><a href="#fnv5zb8pjk25">[14]</a></span></span></sup></span> Specifically, I also had put 65% on relevantly advanced and agentic AI systems being developed by 2070. So my 5% was implying that, <i>conditional</i> on such systems being developed, I was going to look them in the eye and say (in expectation): “~92% that were gonna be OK, x-risk-wise.” But on reflection, that wasnt, actually, how I expected to feel, staring down the barrel of a machine that outstrips human intelligence in science, strategy, persuasion, power; still less, <span><span><span><a href="https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/">billions of such machines</a></span></span></span>; still less, full-blown superintelligence. Rather, I expected to be very scared. More than 8% scared.
</p><h3>
<strong>5.1 Should you trust your future gut, though?</strong>
</h3><p>
Now, you might wonder: why give credit to such future fear?<span role="doc-noteref"><sup><span><span><a href="#fnu9z9f1zd3ci">[15]</a></span></span></sup></span>After all, isnt part of the worry about doomers that theyre, you know, fraidy-cats? Paranoids? (Cmon: its just a superintelligent machine, the invention of a second advanced species, the introduction of a qualitatively new order of optimization power into earths ecosystem. Its just, you know, <i>change</i>.) And isnt the gut, famously, a bit skittish? Indeed, if youre worried about your gut being <i>underactive</i>, at a distance, shouldnt you also be worried about it being <i>over-active,</i> up close? Shouldnt you reason, instead, ahead of time, at a distance, and in a cool hour, about how scared you should be when youre there-in-person?
</p><p>
Well, its a judgment call. Sometimes, indeed, at-a-distance is a better epistemic vantage point than up-close. Especially if you know yourself to have biases. Maybe, for example, youve got a flying phobia, and you know that once youre on the plane, your guts estimates of the probability of the plane crashing are going to go up a lot. Should you update now, then? Indeed: no.
</p><p>
But, personally, with respect to the future, I tend to trust my future self more. Its a dicey game already, futurism, and future Joe has a lot more data. The future is a foreign country, but hes been there.
</p><p>
And I tend to trust my up-close self more, in general, for stuff that requires realization rather than belief (and I think words like “superintelligence” require lots of realization). Maybe the journalist has the accurate casualty count; but I trust the soldier on the ground to know what a casualty <i>means</i>. And I trust the man with the scans about death.
</p><p>
Now, importantly, theres also a thing where guts sometimesreact surprisingly <i>little</i>, up close, to AI stuff you predicted ahead of time youd be scared about. Part of this is the “its not real AI if you can actually do it,” thing (though, my sense is that this vibe is fading?). Part of it is that sometimes, machines doing blah (e.g., beating humans at chess) is less evidence about stuff than you thought. And I wonder if part of it is that sometimes, your at-a-distance fear of that futuristic AI stuff was imagining some world less mundane and “normal” than the world you actually find yourself in, when the relevant stuff comes around — such that when you, sitting in your same-old apartment, wearing your same-old socks, finally see AIs planning, or understanding language, or passing <span><span><span><a href="https://youtu.be/qbIk7-JPB2c?t=1991">two-hour human coding interviews in four minutes</a></span></span></span>, or <span><span><span><a href="https://www.metaculus.com/questions/6728/ai-wins-imo-gold-medal/">winning the IMO</a></span></span></span>, it feels/will feel like “well that cant be the scary thing I had in mind, because that thing is happening in the real world actually and I still have back pain.”<span role="doc-noteref"><sup><span><span><a href="#fn6gbdnzllii">[16]</a></span></span></sup></span> At the least, we get used to stuff fast.  
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/lcmhwebcck5j2xrv9ahf" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/lcmhwebcck5j2xrv9ahf 1024w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/n4mvvfidawvrt3n5xqwl 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/t8t50zsshjaqrlajxoyr 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ulagac5alifmngxbsk3x 1536w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/bfxmnxnuwhx26z74hykv 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/thlguzuwyap6tn7dvnvo 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/oxldviollhiixdaonem6 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zok6hh7zdmnmbmlynv8d 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/hl3hzseipbu2bjaxdxst 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/gwkm1szomnywdtxgk9wt 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xcvmx28xctsnxu0rb3yb 1402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/z0qrdbxaspqc8tbyngby 1702w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ugt4lsizoemp8ojpemf5 1998w"/><p>
<i>GPT-4 doing a coding interview. From</i> <span><span><span><a href="https://www.youtube.com/watch?v=qbIk7-JPB2c&amp;t=1991s"><i>here</i></a></span></span></span><i>.</i>
</p><p>
Still: sometimes, also, you were too scared before, and your gut can see that now. And there, too, I tend to think your earlier self should defer: its not that, if your future self is more scared, you should be more scared now, but if your future self is less scared, you should think that your future self is biased. <span><span><a href="/posts/G5TwJ9BGxcgh5DsmQ/yes-requires-the-possibility-of-no">Yes requires the possibility of no</a></span></span>. If my future self looks the future AGI in the eye and feels like “oh, actually, this isnt so scary after all,” thats evidence that my present self is missing something, too. Heres hoping.
</p><h3>
<strong>5.2 An aside on mental health</strong>
</h3><p>
Now: a quick caution. Here Ive been treating guts centrally from an epistemic perspective. But we need a wise <i>practical</i> relationship with our guts as well. And from a practical perspective, I dont think its always productive to try to smell mustard gas harder, or to make horrible things like AI doom vivid. The right dance here is going to vary person-to-person, and I wont try to treat the topic now (though: see <span><span><a href="/posts/pLLeGA7aGaJpgCkof/mental-health-and-the-alignment-problem-a-compilation-of">here</a></span></span> for a list of resources). But I wanted to flag explicitly that staying motivated and non-depressed and so forth, in relation to a pretty scary situation, is a separate art, and one that needs to be woven carefully with the more centrally epistemic angle Im focused on here.  
</p><h2>
6. Constraints on future worrying
</h2><p>
Returning to the epistemic perspective though: lets suppose you do trust your future credences, and you want to avoid the Bayesian “gut problems” I discussed above. In that case, at least in theory, there are hard constraints on how you should expect your beliefs to change over time, even as you move from far away to up close.
</p><p>
In particular, you should never think that theres more than a 1/<i>x</i> chance that your credence will increase by <i>x</i> times: i.e., never more than a 50% chance that itll double, never more than a 10% chance that itll 10x. And if your credence is very small, then even very small additive increases can easily amount to sufficiently substantive multiplicative increases that these constraints bite. If you move from .01% to .1%, youve only gone up .09% in additive terms only nine parts in ten thousand. But youve also gone up by a factor of 10 something you shouldve been at least 90% sure would never happen.
</p><p>
So suppose that right now, you identify as an “AI risk skeptic,” and you put the probability of doom very low. For concreteness, suppose that you like <span><span><span><a href="https://ineffectivealtruismblog.com/2023/04/08/exaggerating-risks-carlsmith-report/">David Thorstads number</a></span></span></span>: .00002% — that is, one in five million (though: he now thinks this “too generous” and hes also “not convinced that we are in a position where estimating AI risk makes good methodological sense,” which I suspect is a bigger crux). This is a very low number. And it implies, in particular, that you really dont expect to get even a <i>small amount</i> more worried later. For example, you need to have a maximum of .01% that you ever see evidence that puts the probability at &gt;.2%.
</p><p>
Now suppose that a few years pass, GPT-6 comes out, and lo, indeed, it is very impressive. You look GPT-6 in the eye and you feel some twinge in your gut. You start to feel a bit, well, at-least-1-percent-y. A bit not-so-crazy-after-all. Now, admittedly, you were probably surprised that GPT-6 is so good. You were a “timelines skeptic,” too. But: how much of a skeptic? Were you, for example, less than one in fifty thousand that GPT-6 would be this impressive? Thats what your previous number can easily imply, if the impressiveness is whats driving your update.
</p><p>
And now suppose that actually, you werent much of a timelines skeptic at all. GPT-6, according to you, is right on trend. Youd seen the scaling laws. You were at &gt;50% on at-least-this-impressive. It was predictable. Its just that the rest of the argument for doom is dumb.
</p><p>
In that case, though, hmm. Your guts got heavy constraints, in terms of twinging. &gt;50% on at least-this-impressive? So: youre still supposed to be at less than .00004% on doom? But what if youre not…
</p><p>
Or maybe you think: “the argument for doom has not been satisfactorily peer-reviewed. <span><span><span><a href="https://marginalrevolution.com/marginalrevolution/2023/04/this-gpt-4-answer-speaks-for-itself.html">Wheres the paper in <i>Nature</i></a></span></span></span>? Until I see conventional academic signals, I am at less than one in a thousand on doom, and I shall tweet accordingly.” OK: but, the Bayesianism. If youre at less than one in a thousand, now, and your big thing is academic credibility, where should Bayes put you later, conditional on <i>seeing</i> conventional academic signals? And whats your probability on such strange sights? In five years, or ten years, are you confident there wont be a paper in <i>Nature</i>, or an equivalent? If its even 10% percent likely, and it would take you to more than 1%, your number now should be moving ahead of time.
</p><p>
Or maybe you thought, in the past: “until I see the experts worrying, Im at less than 1%.” Well, <span><span><span><a href="https://www.nytimes.com/2023/05/01/technology/ai-google-chatbot-engineer-quits-hinton.html">here we are</a></span></span></span> (here we already were, but more now). But: what was your probability that we ended up here? Was it so hard to imagine, the current level of expert sympathy? And are future levels of greater sympathy so hard to imagine, now? Its easy to live, only, in the present to move only as far as the present has moved. But the Bayesian has to live, ahead of time, in all the futures at once.
</p><p>
(Note that all of these comments apply, symmetrically, to people nearly certain of doom. 99.99%? OK, so less than 1% than you ever drop to 99% or lower? So little hope of future hope?)
</p><p>
Now: all of this is “in theory.” In practice, this sort of reasoning requires good taste. I talk about such taste more below. First, though, I want to look at the theory a bit more.
</p><h2>
7. Should you expect low probabilities to go down?
</h2><p>
Above I said that actually, the <i>direction</i> of a future update is often predictable. But notice: <i>which direction</i> should you predict? My sense is that in many evidential situations (though not all more below), you should think your future evidence more likely to move you in the right direction than the wrong one. So if you think that <i>p</i> is likely to be true, you should generally think that your future evidence is likely to update you towards higher credence on <i>p</i>. And vice versa: if you think that p is more likely to be <i>false</i>, you should expect to have <i>lower</i> credence on it later.
</p><p>
The Trump example above is an extreme case. Youre at 99% on Trump winning, and youre also at 99% that youll update, in future, towards higher credence on Trump winning. And we can imagine a more intermediate case, where, lets say, youre at 90% that Trump is going to win, and youre about to watch the presidential debate, and you think that winning the debate is highly correlated with winning the election. Which direction should you predict that your credence on Trump winning will move, once the debate is over? Given that you think Trump is likely to win the election, I think you should think hes likely to win the debate, too. And if he wins the debate, your credence on him winning the election will go up (whereas if he loses, itll go down a bunch more).
</p><p>
Or consider a scientist who doesnt believe in God. In principle, at each moment, God could appear before her in a tower of flames. She has some (very small) credence on this happening. And if it happened, she would update massively towards theism. But <span><span><a href="/posts/mnS2WYLCGJP2kQkRn/absence-of-evidence-is-evidence-of-absence">absence of evidence is evidence of absence</a></span></span>. Every moment she <i>doesnt</i> observe God appearing before her in a tower of flames, she should be updating some tiny amount towards atheism. And because she predicts very hard that God will never appear before her in a tower of flames, she should be predicting very hard that she will become a more and more confident atheist over time, and that shell die with even less faith than she has now.
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/eufqrpvhqgyjzngzhejg" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/nq26vo1twqjfrp8lugux 795w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/cx4bl4je69by5syhvvxj 233w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/brnbm1vlo7vkotqm5jdn 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/cehrsyb68kukjamyah8p 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qkjyknw4irznm7raz4ys 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mgu5fwmjcrbl7i94t9do 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ruqpj7osmhzbemurme65 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zwv61bykcbqn6jfauqqe 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/js674olpdq3nnneki6g0 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/yreunwtnysfogdbqnne2 1148w"/><p>
<i>Updating so hard right now… (Image source</i> <a href="https://commons.wikimedia.org/wiki/File:Bourdon,_S%C3%A9bastien_-_Burning_bush.jpg"><i>here</i></a><i>.)</i>
</p><p>
So too, one might think, with AI risk. If you are currently an AI risk skeptic, plausibly you should expect to become more and more confidently skeptical over time, as your remaining uncertainties about the case for non-doom get resolved in the direction of truth. That is, every moment that the superintelligent machines <i>dont</i> appear before you in a tower of diamondoid bacteria (thats the story, right?), then anthropic effects aside, you should be breathing easier and easier. Or, more realistically, you should be expecting to see, well, whatever it is that comforts you: i.e., that well hit another AI winter; or that well make lots of progress in mechanistic interpretability; or that innovations in RLHF will allow superhuman oversight of AI behavior humans cant understand; or that we wont see any signs of deception or reward hacking; or that progress will be slow and gradual and nicely coordinated; or that well finally, <i>finally</i>, get some peer review, and put the must-be-confusions to rest. And as your predictions are confirmed, you should be feeling safer and safer.
</p><p>
Is that what you expect, in your heart? Or are you, perhaps, secretly expecting to get more worried over time? I wished Id asked myself harder. In particular: my 5% was plausibly implying some vibe like: “sure, there are these arguments that superintelligent AI will disempower us, and I give them some weight, but at least if were able to think well about the issue and notice the clues that reality is giving us, over time it will probably become clearer that these arguments are wrong/confused, and well be able to worry much less.” Indeed, depending on the volatility of the evidence I was expecting, perhaps I should have thought that I was likely to be in the ballpark of the highest levels of worry about doom that I would ever endorse. But if youd asked me, would I have said that?
</p><p>
That said, I actually think these dynamics are more complicated than they might initially seem. In particular, while I find it plausible that you should generally predict that youll update in the direction of what you currently expect to be true, sometimes, actually, you shouldnt. And some non-crazy views on AI risk fit the mold.
</p><p>
Katja Grace suggested to me some useful examples. Suppose that youre in a boat heading down a river. You at 80% that theres a waterfall about two miles down, but 20% that there isnt, and that youre going to see a sign, a mile down, saying as much (“No waterfall” classic sort of sign). Conditional on no sign/there being a waterfall, youre at 10% that its a big waterfall, which will kill you, and 90% that its a small waterfall, which youll survive. So currently, your credence on dying is 8%. However, youre also at 80% that in a mile, its going to go up, to 10%, despite your also predicting, now, that this is an update towards higher credence on something that probably wont happen.
</p><p>
Or a consider a more real-world example (also from Katja). At 3 pm, youre planning to take a long car trip. But theres a 10% chance the trip will fall through. If you take the trip, theres some small chance you get in an accident. As you approach 3 pm, your credence in “I will get in a car accident today” should go up, as the trip keeps (predictably) not-falling-through. And then, as youre driving, it should go down gradually, as the remaining time in the car (and therefore, in danger) shrinks.
</p><p>
Some views on AI including, skeptical-of-doom views look like this. Suppose, for example, you think AGI-by-2070 more likely than not. And suppose that conditional on AGI-by-2070, you think theres some small risk that the doomers are right, and we all die. And you think its going to be hard to get good evidence to rule this out ahead of time. Probably, though, well make it through OK. And conditional on no-AGI-by-2070, you think were almost certainly fine. Here, you should plausibly expect to get more worried over time, as you get evidence confirming that yes, indeed, AGI-by-2070; yes, indeed, waterfall ahead. And then to get less worried later, as the waterfall proves small.
</p><p>
That said, this sort of dynamic requires specific constraints on what evidence is available, when. The truth about the future must fail to leak backwards into the past. You must be unable to hear the difference between a big waterfall and a small waterfall sufficiently ahead-of-time. The gas ahead must not waft.
</p><p>
Car accidents are indeed like this. People rarely spend much time with high credence that theyre about to get in a car accident. Their probability is low; and then suddenly it jumps wildly, split-second high, before death, or some bang-crunch-jerk, or a gasping near-miss.
</p><p>
Is AI risk like this too? Doomers sometimes talk this way. Youll be cruising along. Everything will be looking rosy. The non-doomers will be feeling smug. Then suddenly: bam! <span>The nanobots, from the bloodstream, in the parlor, Professor Plum.</span> The clues, that is, didnt rest on the details. A lot of it was obvious a priori. You shouldve read more LessWrong back in the 2000s. You shouldve looked harder at those <span><span><span><a href="https://twitter.com/ESYudkowsky/status/1500863629490544645">empty strings</a></span></span></span>.
</p><p>
Now, sometimes this sort of vibe seems to me like it wants to have things both ways. “I shall accept ahead-of-time empirical evidence that I am right; but in the absence of such evidence, I shall remain just as confident.” “My model makes no confident predictions prior to the all-dropping-dead thing except, that is, the ones that I want to claim credit for after-the-fact.” Here I recall a conversation I overheard back in 2018 about “<span><span><span><a href="https://arbital.com/p/daemons/">optimization daemons</a></span></span></span>” (now: <span><span><span><a href="https://arxiv.org/abs/1906.01820">mesa-optimizers</a></span></span></span>, <span><span><span><a href="https://arxiv.org/abs/2210.01790">goal mis-generalization</a></span></span></span>, etc) in which a worrier said something like: “I will accept empirical arguments for concern, but only a priori arguments for comfort.” It was an offhand remark, but still: <span><span><a href="/posts/mnS2WYLCGJP2kQkRn/absence-of-evidence-is-evidence-of-absence">not how it works</a></span></span>.
</p><p>
However: I do think, unfortunately, there are risks of gas that doesnt waft well; “<span><span><span><a href="https://forum.effectivealtruism.org/posts/NbiHKTN5QhFFfjjm5/ai-safety-seems-hard-to-measure#_2__The_King_Lear_problem__how_do_you_test_what_will_happen_when_it_s_no_longer_a_test_">King Lear problems</a></span></span></span>”; risks of <span><span><span><a href="https://www.planned-obsolescence.org/the-training-game/">things looking fairly fine, right before they are very non-fine indeed</a></span></span></span>. But not all the gas is like this. We should expect to get clues (indeed, we should <span><span><a href="/posts/rCJQAkPTEypGjSJ8X/how-might-we-align-transformative-ai-if-it-s-developed-very#Testing_and_threat_assessment"><i>dig hard for them</i></a></span></span>)<i>.</i> So we should expect, at some point, to start updating in the right direction. But I think its an open question how the sequencing here works, and itll depend on the details driving your particular view. In general, though, if youre currently at more-likely-than-not on hitting an AGI waterfall sometime in the coming decades, but not certain, then prima facie, and even if your p(doom) is low, thats reason to expect to get more worried as that soothing sign “AI winter,” “It was all fake somehow” (classic sign) fails to appear.
</p><p>
That said, even if youre getting predictably <i>more</i> worried, there are still Bayesian constraints on <i>how much</i>. In the waterfall case, you go up 2%; in the car case, something tiny. So if youre finding yourself, once you dont see the sign, jumping to 50% on “death by big waterfall” well, hmm, according to your previous views, youre saying that youre in a much-more-worrying-than-average not-seeing-the-sign scenario. Whence such above-average-worrying? Is the evidence youre seeing now, re: big-waterfall, actually surprising relative to what you expected before? Looks a lot like the predicted river to me. Looks, indeed, “just like they said.” Or did your gut, maybe, not really believe …
</p><h2>
<strong>8. Will the next president be a potato?</strong>
</h2><p>
OK, that was a bunch of stuff about basic Bayesian belief dynamics. And armed with this sort of relatively crisp and simple model, it can be easy to start drawing strong conclusions about how you, with your mushy monkey brain, should be reasoning in the practice, and what sorts of numbers should be coming out of your mouth, when you make number-noises.
</p><p>
But the number-noise game takes taste. Its a new game. Were still learning how to play well, and productively. And I think we should be wary of possible distortions, especially with respect to small-probabilities.
</p><p>
Consider, for example, the following dialogue:
</p><blockquote><p>
<i>Them</i>: Whats your probability that the next president is a potato?
</p><p>
<i>You</i>: What?
</p><p>
<i>Them</i>: A potato. Like, a normal potato. Up there getting inaugurated and stuff.
</p><p>
<i>You</i>: Umm, very low?
</p><p>
<i>Them</i>: Say a number!
</p><p>
<i>You</i>: [blank stare]
</p><p>
<i>Them</i>: You are a Bayesian and must have a number, and I demand that you produce it. Just literally say any number and I will be satisfied.
</p><p>
<i>You</i>: Fine. One in 10^50.  
</p><p>
<i>Them</i>: What? Really? Wow thats so stupid. I cant believe you said that.
</p><p>
<i>You</i>: Actually, lets say one in 10^40.
</p><p>
<i>Them</i>: Wait, your number was more than a billion times lower a second ago. If you were at one in 10^50 a second ago, you shouldve been at less than one-in-a-billion that youd ever move this high. Is the evidence youve got since then so surprising? Clearly, you are a bad Bayesian. And I am clever!
</p><p>
<i>You</i>: This is a dumb thing.
</p></blockquote><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rf5bkvfdso7uz3y3tsnh" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rf5bkvfdso7uz3y3tsnh 1019w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ezpszxdmdzq3uy3p8qhi 298w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rpg5vqlsuzidqen5c4di 150w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qc9y8vkry2clju0stdwh 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/zyz51cimtve586t3zhuu 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/mwkiyj92xdccwxuj86ot 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/vgf2m89mvbsmu2mvb7ey 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/x2h0uk8rtwyxbmai7hnz 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/sg2q1gjcssoc8yjl6bo7 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/pacecvldaxuy90uwyzir 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ypgoixgh1fm3krr8eijy 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qdfuisvwrauqypkabjsq 1156w"/><p>
<i>Not like this: a normal potato.</i>
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/h8ciyvdtf5bl1jlfhnoy" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/xwhen7t5tpcuk3ljla0p 1024w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/j344jhvcsibpfq3feglm 300w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/rvsgwlqjofhfhzgkbbvp 150w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qvhouh0jdskz5w5fddeb 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/slgjc8ozooac97bov1v0 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/r94awwenrlstpysths62 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kz7lufvzxvkannecezen 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ycxedzguhkdur5bbniyl 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/gpsgvjn4rhx1viuk60kq 982w"/><p>
<i>Closer…</i>
</p><p>
The “them” vibe, here, seems dubiously helpful. And in particular, in this case, its extra not-helpful to think of “you” as changing your probabilities, from one second to the next, by updating some fully-formed probability distribution over Potato-2024, complete with expected updates based on all the possible next-thoughts-you-could-think, reactions “them” might have, and so on. Thats, just, not the right way to understand whats going on with the fleshy creatures described in this dialogue. And in general, it can be hard to have intuitions about <span><span><span><a href="https://markxu.com/strong-evidence">strong evidence</a></span></span></span>, and extreme numbers make human-implemented Bayesian especially brittle.  
</p><p>
Now, to be clear: I think that debates about the rough quantitative probability of AI doom are worth engaging in, and that they are in fact (unfortunately) very different from debates about Potato-2024. Still, though, that old lesson looms: do not confuse your abstract model of yourself with yourself. The map is never the territory; but especially not when youre imagining a map that would take a <span><span><span><a href="https://joecarlsmith.com/2021/10/29/on-the-universal-distribution#i-the-universal-distribution">hyper-computer to compute</a></span></span></span>. Fans of basic Bayesianism, and of number-noises, are <span><span><a href="/posts/CPP2uLcaywEokFKQG/toolbox-thinking-and-law-thinking">well-aware of this</a></span></span>; but the right dance, in practice, remains an open question.
</p><p>
As an example of a distortion I worry about with respect to the previous discussion: in practice, lots of people (myself included but see also Christiano <span><span><span><a href="https://ai-alignment.com/my-views-on-doom-4788b1cd0c72">here</a></span></span></span>) report volatility in their degree of concern about p(doom). Some days, I feel like “man, I just cant see how this goes well.” Other days Im like: “What was the argument again? All the AIs-that-matter will have long-term goals that benefit from lots of patient power-grabbing and then coordinate to deceive us and then rise up all at once in a coup? Sounds, um, pretty specific…”
</p><p>
Now, you could argue that either your expectations about this volatility should be compatible with the basic Bayesianism above (such that, e.g., if you think it reasonably like that youll have lots of &gt;50% days in future, you should be pretty wary of saying 1% now), or youre probably messing up. And maybe so. But I wonder about alternative models, too. For example, Katja Grace suggested to me a model where youre only able to hold some subset of the evidence in your mind at once, to produce your number-noise, and different considerations are salient at different times. And if we use this model, I wonder if how we think about volatility should change.<span role="doc-noteref"><sup><span><span><a href="#fn7yx0orvjyje">[17]</a></span></span></sup></span>
</p><p>
<span>Indeed, even on basic Bayesianism, volatility is fine as long as the averages work out</span> (e.g., you can be at an <i>average</i> of 10% doom conditional on GPT-6 being “scary smart,” but 5% of the time you jump to 99% upon observing a scary smart GPT-6, 5% of the time you drop to near zero, and in other cases you end up at lots of other numbers, too). And it can be hard to track all the evidence youve been getting. Maybe you notice that two years from now, your p(doom) has gone up a lot, despite AI capabilities seeming on-trend, and you worry that youre a bad Bayesian, but actually there has been some other build-up of evidence for doom that youre not tracking for example, the rest of the world starting to agree.<span role="doc-noteref"><sup><span><span><a href="#fn72q8oh08rre">[18]</a></span></span></sup></span>
</p><p>
And there are other more familiar risks of just getting even the basic Bayesianism wrong. Maybe, for example, you notice that your beliefs have been trending in a certain direction. Trump keeps moving up in the polls, say. Now youre at like 95% on Trump win. And you read a tweet like <span><span><span><a href="https://twitter.com/NPCollapse/status/1626854680260231169">Connor Leahys</a></span></span></span>, below, telling you to “just update all the way, bro” and so you decide, shit, Ill just go 100%, and assume that Trump <i>will</i> win. Wouldnt want to predictably update later, right?
</p><img src="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/tk8y11ey6lqmveiws7wn" alt="" srcset="https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/tk8y11ey6lqmveiws7wn 817w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/i4j91yqvpuobz2rqw868 239w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/squfak3qevyosnflwfat 768w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/cbt8xvuobl9fjnxmnnts 1226w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/sfbghfvnj0hz0yr5y68f 402w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ezw8jwyifhypezexpmgm 462w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/ad6hz0opqqfrynr3obos 662w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/noszolhwi0fj9edhok5p 722w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/vdss52lqyqxmauxbljp1 982w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/kyvvrkntm51gcrytmqme 1032w, https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/bHozHrQD4qxvKdfqq/qdygqwvjmkb0vjcy7sgq 1398w"/><p>
Or maybe you hear some <span><span><span><a href="https://www.facebook.com/yudkowsky/posts/10160260422389228">prominent doomer proclaiming that “sane people with self-respect” dont update predictably</a></span></span></span>, without clarifying about “in expectation” despite <span><span><a href="/posts/jiBFC7DcCrZjGmZnJ/conservation-of-expected-evidence">definitely knowing about this</a></span></span>, and so you assume you must be unsane and self-hating. Or maybe you think that if you do update predictably, it should at least be in the direction of your currently-predicted truth, and you forget about cases like the waterfalls above.
</p><p>
In general, this stuff can get tricky. We should be careful, and not self-righteous, even when the math itself is clear.  
</p><h2>
9. Just saying “oops”
</h2><p>
I also want to add a different note of caution, about not letting consistency, or your abstract picture of what “good Bayesianism” looks like, get in the way of updating as fast as possible to the right view, whatever that is.
</p><p>
Thus, for example, maybe you tweeted a bunch in the past re: “no way” on AI risk, and acted dismissive about it. Maybe, even, youre someone like David Thorstad, and you were kind enough to quantify your dismissiveness with some very-low number.
</p><p>
And lets say, later, your gut starts twinging. Maybe you see some scary demo of deceptiveness or power-seeking. Maybe you dont like the look of all those increasingly-automated, AI-run wet-labs. Maybe it all just starts changing too fast, and it feels too frenetic and out of control, and do we even understand how these systems are working? Maybe its something about those new drones.  
</p><p>
It might be tempting, here, to let your previous skepticism drag your new estimates downwards including on the basis of the sorts of dynamics discussed above. Maybe, for example, if you had David Thorstads number, youre tempted to move from .00002% to something like, hmm, 20%? But you say to yourself “wait, have I really gotten <i>such strong evidence</i> since my previous estimate? Have I been <i>so surprised</i> by the demos, and the drones, and the wet-labs? Apparently, Im moving to a number I shouldve been less than one-in-a-million Id ever end up at. By my previous lights, isnt that unlikely to be the right move?”
</p><p>
But the thing is: its possible that your previous estimate was just … way too low. And more (gasp), that it didnt come with some well-formed probability distribution over your future estimates, either. We should be wary, in general, of taking our previous (or our current) Bayesian rigor too seriously. Should “you,” above, refrain from changing her potato-2024 estimate quickly as she thinks about it more, on grounds that it would make her two-seconds-ago selfs Bayesianism look bad? Best to just get things right.
</p><p>
Of course, it may be that your previous self was tracking some sort of evidence that youre losing sight of, now. It may be that your gut is skittish. You should try to learn from your previous self what you can. But you should try, I suspect, to learn harder from the actual world, there in front of you.
</p><p>
Here, to be clear, Im partly thinking about myself, and my own mistakes. I said 5% in 2021. I more than doubled my estimate soon after.  By basic Bayes, I shouldve been less than 50%, in 2021, that this would happen. Did I really get sufficiently worrying evidence in the interim to justify such a shift? Maybe. But alternatively: whatever, I was just wrong. Best to just say oops, and to try to be righter.
</p><p>
Im focusing on people with very low estimates on doom, here, because they tend to be more common than the converse. But everything Im saying here holds for people with low estimates on non-doom, too. If youre such a person, and you see signs of hope later, dont be attached to your identity as a definitely-doomer, or to the Bayesian rigor of the self that assumed this identity. Dont practice your pessimism over-hard. You might miss the thing that saves your life.
</p><p>
Really, though, I suspect that respect for your previous selfs Bayesianism is not the main barrier to changing our minds fast enough. Rather, the barriers are more social: embarrassment stuff, tribal stuff, status stuff, and so on. I think we should try to lower such barriers where possible. We should notice that people were wrong; but we should not make fun of them for changing their minds quite the contrary. Scout mindset is hard enough, and the stakes are too high.
</p><p>
Ill close by noting a final sort of predictable update. Its related to the scans thing.
</p><p>
Theres <span><span><span><a href="https://www.youtube.com/watch?v=W9vj2Wf57rQ">a scene</a></span></span></span> at the end of <i>Schindlers List</i>. World War II is over. Schindler has used his money to save more than 1,100 lives from the holocaust. As the people he has saved say goodbye, Schindler breaks down:
</p><blockquote><p>
I could have got more out. I could have got more. I dont know. If Id just… I could have got more… I threw away so much money. You have no idea… I didnt do enough… This car. Goeth would have bought this car. Why did I keep the car? Ten people right there. Ten people. Ten more people. This pin. Two people. This is gold. Two more people. He would have given me two for it, at least one. One more person. A person, Stern. For this. I could have gotten one more person… and I didnt.
</p></blockquote><p>
Now, we need to be careful here. Its easy for the sort of stuff Im about to say to prompt extreme and unbalanced and unhealthy relationships to stuff that matters a lot. In particular, if youre tempted to be in some “emergency” mode about AI risk (or, indeed, about some other issue), and to start burning lots of resources for the sake of doing everything you can, I encourage you to read <span><span><a href="/posts/mmHctwkKjpvaQdC3c/what-should-you-change-in-response-to-an-emergency-and-ai">this article</a></span></span>, together with <span><span><a href="/posts/mmHctwkKjpvaQdC3c/what-should-you-change-in-response-to-an-emergency-and-ai?commentId=Htf2v79w5QoQJbysS#comments">this comment</a></span></span> about memetic dynamics that can amplify false emergencies and discourage clear thinking.
</p><p>
Still, still. Theres a possible predictable update here. If this AI stuff really happens, and the alignment stuff is looking rough, there is a way we will each feel about what we did with the time we had. How we responded to what we knew. What role we played. Which directions we pointed the world, or moved it. How much we left on the field.
</p><p>
And there is a way we will feel, too, about subtler things. About what sorts of motivations were at play, in how we oriented towards the issue. About the tone we took on twitter. About the sort of <span><span><span><a href="https://joecarlsmith.com/2022/12/23/on-sincerity">sincerity</a></span></span></span> we had, or didnt have. One thing that stayed with me from <i>Dont Look Up</i> is the way the asteroid somehow slotted into the worlds pre-existing shallowness; the veneer of unreality and unseriousness that persisted even till the end; the status stuff; the selfishness; the way that somehow, still, that fog. If AGI risk ends up like this, then looking back, as our time runs out, I think there will be room for the word “shame.” Death does not discriminate between the sinners and the saints. But I do actually think its worth talk of dignity.
</p><p>
And there is a way we will feel, too, if we step up, do things right, and actually solve the problem. Some doomer discourse is animated by a kind of bitter and exasperated pessimism about humanity, in its stupidity and incompetence. But different vibes are available, too, even holding tons of facts fixed. Here Im particularly interested in “lets see if we can actually do this.” Humans can come together in the face of danger. Sometimes, even, danger brings out our best. It is possible to see that certain things should be done, and to just do them. It is possible for people to work side by side.
</p><p>
And if we do this, then there is a way we will feel when its done. I have a friend who sometimes talks about what he wants to tell his grandchildren he did, during the years leading up to AGI. Its related to that thing about history, and who its eyes are on. We shouldnt need people to tell our stories; but as far as I can tell, if he ever has grandchildren, they should be proud of him. May he sit, someday, under his own vine and fig tree.
</p><p>
Of course, there is also a way we will feel if AGI happens, but the problem was unreal, or not worth worrying about. There are <span><span><span><a href="https://www.planned-obsolescence.org/the-costs-of-caution/">costs of caution</a></span></span></span>. And of course, there is a way we will feel if all this AGI stuff was fake after all, and all that time and money and energy was directed at a fantasy. You can talk about “reasonable ex ante,” but: will it have been reasonable? If this stuff is a fantasy, I suspect it is a fantasy connected with our flaws, and that we will have been, not innocently mistaken, but actively foolish, and maybe worse. Or at least, I suspect this of myself.
</p><p>
Overall, then, there are lots of different possible futures here. As ever, the Bayesian tries to live in all of them at once. Still: if, indeed, we are running out of time, and there is a serious risk of everyone dying, it seems especially worth thinking ahead to hospitals and scans; to what we will learn, later, about “enough” and “not enough,” about “done” and “left undone.” Maybe there will be no history to have its eyes on us or at least, none we would honor. But we can look for ourselves.   
</p><ol><li><span><sup><strong><span><a href="#fnref0g9sn95iiqe">^</a></span></strong></sup></span>
<div>
<p>
To be clear: there are lots of other risks from AI, too. And the basic dynamics at stake in the essay apply to your probabilities on any sorts of risks. But I want to focus on existential risk from misalignment, here, and I want the short phrase “AI risk” for the thing Im going to be referring to repeatedly.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref048nj5dhr9di">^</a></span></strong></sup></span>
<div>
<p>
Though, the specific numbers here can matter and there are some cases where despite having low probabilities on doom now, you can predict ahead of time that youll be at least somewhat more worried later (though, there are limits to how much). More below.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefqnj4snix8n">^</a></span></strong></sup></span>
<div>
<p>
Though with respect to AI doom, not risk free see <span><span><a href="/posts/pLLeGA7aGaJpgCkof/mental-health-and-the-alignment-problem-a-compilation-of">here</a></span></span> for some mental health resources.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefz4yjec5asvf">^</a></span></strong></sup></span>
<div>
<p>
Hopefully not more literally similar. But: a new thing-not-imagined-very-well.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref1dgio52qsrw">^</a></span></strong></sup></span>
<div>
<p>
Modulo some futurisms. Including, importantly, ones predictably at stake in AI progress.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref4m2cfburfm7">^</a></span></strong></sup></span>
<div>
<p>
“In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentines Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time.”
</p>
</div></li><li><span><sup><strong><span><a href="#fnref4y75ebmy7eh">^</a></span></strong></sup></span>
<div>
<p>
Thanks to Katja Grace for discussion.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref5gj8wfkheeo">^</a></span></strong></sup></span>
<div>
<p>
Some forecasts have self-fulfilling elements, especially with respect to Moloch-like problems. And there are questions about e.g. internet text increasing the likelihood of AIs acting out the role of the scary-AI.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefh3ar4lqw084">^</a></span></strong></sup></span>
<div>
<p>
See e.g. Scott Alexander <span><span><span><a href="https://astralcodexten.substack.com/p/mantic-monday-31422">here</a></span></span></span>. Some of <span><span><span><a href="https://forum.effectivealtruism.org/posts/Lto9awEYPQNu9wkdi/rational-predictions-often-update-predictably#fn6am2fn0yyve">Yudkowskys public comments</a></span></span></span> suggest this model as well, though his original discussion of “<span><span><a href="/posts/jiBFC7DcCrZjGmZnJ/conservation-of-expected-evidence">conservation of expected evidence</a></span></span>” does not.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref4ohd2xn7yql">^</a></span></strong></sup></span>
<div>
<p>
Here Im indebted to discussion from <span><span><span><a href="https://forum.effectivealtruism.org/posts/Lto9awEYPQNu9wkdi/rational-predictions-often-update-predictably">Greg Lewis</a></span></span></span> and <span><span><a href="/posts/zTfSXQracE7TW8x4w/mistakes-with-conservation-of-expected-evidence">Abram Demski</a></span></span>.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefc1l5t3jttlk">^</a></span></strong></sup></span>
<div>
<p>
.99*1 + .01*0 = .99.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefo3exn1kvfm">^</a></span></strong></sup></span>
<div>
<p>
Or put another way: you want to find the area that red occupies, which is the area of the first, smaller red box, plus the area of the bigger red box. Each box occupies a percentage of the area of a “column” (combination of white box and red box) associated with a hypothesis about GPT-6. So to find the area of a given red box, you take the area of the column its in (that is, the probability on the relevant hypothesis about GPT-6), and multiply that by the percentage of that column that is red (e.g., the probability of doom conditional on that hypothesis). Then you add up the areas of the red boxes.
</p>
</div></li><li><span><sup><strong><span><a href="#fnreft4taauzgccn">^</a></span></strong></sup></span>
<div>
<p>
Thanks to Daniel Kokotajlo for highlighting some of these dynamics to me years ago. See also his review of my power-seeking AI report <span><span><span><a href="https://docs.google.com/document/d/1GwT7AS_PWpglWWrVrpiMqeKiJ_E2VgAUIG5tTdVhVeM/edit#heading=h.e9o5m3fab0ua">here</a></span></span></span>.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefv5zb8pjk25">^</a></span></strong></sup></span>
<div>
<p>
I added an edit to this effect.
</p>
</div></li><li><span><sup><strong><span><a href="#fnrefu9z9f1zd3ci">^</a></span></strong></sup></span>
<div>
<p>
Thanks to Katja Grace for discussion here.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref6gbdnzllii">^</a></span></strong></sup></span>
<div>
<p>
Here Im inspired by some comments from Richard Ngo.
</p>
</div></li><li><span><sup><strong><span><a href="#fnref7yx0orvjyje">^</a></span></strong></sup></span>
<div>
<p>
Though: maybe it just works out the same? E.g., the average of your estimates over time needs to obey Bayesian constraints?
</p>
</div></li><li><span><sup><strong><span><a href="#fnref72q8oh08rre">^</a></span></strong></sup></span>
<div>
<p>
Again, thanks to Katja for pointing to this dynamic.
</p>
</div></li></ol></div>