From facea5a2792336c52f71d8970297161ca20a434b Mon Sep 17 00:00:00 2001 From: "M. Taylor Saotome-Westlake" Date: Wed, 18 Jan 2023 20:59:39 -0800 Subject: [PATCH] memoir poke: AlphaGo moments --- ...xhibit-generally-rationalist-principles.md | 23 +++++++++++-------- notes/memoir-sections.md | 5 +++- 2 files changed, 18 insertions(+), 10 deletions(-) diff --git a/content/drafts/agreeing-with-stalin-in-ways-that-exhibit-generally-rationalist-principles.md b/content/drafts/agreeing-with-stalin-in-ways-that-exhibit-generally-rationalist-principles.md index 687a2a4..86988cb 100644 --- a/content/drafts/agreeing-with-stalin-in-ways-that-exhibit-generally-rationalist-principles.md +++ b/content/drafts/agreeing-with-stalin-in-ways-that-exhibit-generally-rationalist-principles.md @@ -516,19 +516,23 @@ I notice that this advice leaves out a possibility: that the "seems to believe" So far, I've been writing from the perspective of caring (and expecting Yudkowsky to care) about human rationality as a cause in its own right—about wanting to _make sense_, and wanting to live in a Society that made sense, for its own sake, and not as a convergently instrumental subgoal of saving the world. -That's pretty much always where I've been at. I _never_ wanted to save the world. I got sucked in to this robot cult because Yudkowsky's philsophy-of-science writing was just that good. I did do a little bit of work for the Singularity Institute back in the day (an informal internship in 'aught-nine, some data-entry-like work manually adding Previous/Next links to the Sequences, designing several PowerPoint presentations for Anna, writing some Python scripts to organize their donor database), but that was because it was my social tribe and I had connections. To the extent that I took at all seriously the whole save/destroy/take-over the world part (about how we needed to encode all of human morality into a recursively self-improving artificial intelligence to determine our entire future light cone until the end of time), I was scared rather than enthusiastic. +That's pretty much always where I've been at. I _never_ wanted to save the world. I got sucked in to this robot cult because Yudkowsky's philsophy-of-science blogging was just that good. I did do a little bit of work for the Singularity Institute back in the day (an informal internship in 'aught-nine, some data-entry-like work manually adding Previous/Next links to the Sequences, designing several PowerPoint presentations for Anna, writing some Python scripts to organize their donor database), but that was because it was my social tribe and I had connections. To the extent that I took at all seriously the whole save/destroy/take-over the world part (about how we needed to encode all of human morality into a recursively self-improving artificial intelligence to determine our entire future light cone until the end of time), I was scared rather than enthusiastic. Okay, being scared was entirely appropriate, but what I mean is that I was scared, and concluded that shaping the Singularity was _not my problem_, as contrasted to being scared, then facing up to the responsibility anyway. After a 2013 sleep-deprivation-induced psychotic episode which [featured](http://zackmdavis.net/blog/2013/03/religious/) [futurist](http://zackmdavis.net/blog/2013/04/prodrome/)-[themed](http://zackmdavis.net/blog/2013/05/relativity/) [delusions](http://zackmdavis.net/blog/2013/05/relevance/), I wrote to Anna, Michael, and some MIRI employees who had been in my contacts for occasional contract work, that "my current plan [was] to just try to forget about _Less Wrong_/MIRI for a long while, maybe at least a year, not because it isn't technically the most important thing in the world, but because I'm not emotionally stable enough think about this stuff anymore" (Subject: "to whom it may concern"). When I got a real programming job and established an income for myself, I [donated to CfAR rather than MIRI](http://zackmdavis.net/blog/2016/12/philanthropy-scorecard-through-2016/), because public rationality was something I could be unambiguously enthusiastic about, and doing anything about AI was not. -At the time, it seemed fine for the altruistically-focused fraction of my efforts to focus on rationality, and to leave the save/destroy/take-over the world stuff to other, more emotionally stable people, in accordance with the principle of comparative advantage. Yudkowsky had written his Sequences as a dependency for explaining [the need for friendly AI](https://www.lesswrong.com/posts/GNnHHmm8EzePmKzPk/value-is-fragile), ["gambl[ing] only upon the portion of the activism that would flow to [his] own cause"](https://www.lesswrong.com/posts/9jF4zbZqz6DydJ5En/the-end-of-sequences), but rationality was supposed to be the [common interest of many causes](https://www.lesswrong.com/posts/4PPE6D635iBcGPGRy/rationality-common-interest-of-many-causes). Even if I wasn't working or donating to MIRI, I was still _helping_, a good citizen according to the morality of my tribe. +At the time, it seemed fine for the altruistically-focused fraction of my efforts to focus on rationality, and to leave the save/destroy/take-over the world stuff to other, more emotionally-stable people, in accordance with the principle of comparative advantage. Yudkowsky had written his Sequences as a dependency for explaining [the need for friendly AI](https://www.lesswrong.com/posts/GNnHHmm8EzePmKzPk/value-is-fragile), ["gambl[ing] only upon the portion of the activism that would flow to [his] own cause"](https://www.lesswrong.com/posts/9jF4zbZqz6DydJ5En/the-end-of-sequences), but rationality was supposed to be the [common interest of many causes](https://www.lesswrong.com/posts/4PPE6D635iBcGPGRy/rationality-common-interest-of-many-causes). Even if I wasn't working or donating to MIRI specifically, I was still _helping_, a good citizen according to the morality of my tribe. But fighting for public epistemology is a long battle; it makes more sense if you have _time_ for it to pay off. Back in the late 'aughts and early 'tens, it looked like we had time. We had these abstract philosophical arguments for worrying about AI, but no one really talked about _timelines_. I believed the Singularity was going to happen in the 21st century, but it felt like something to expect in the _second_ half of the 21st century. -Now it looks like we have—less time? Not just tautologically because time has passed (the 21st century is one-fifth over—closer to a quarter over), but because of new information from the visible results of the deep learning revolution during that time. Yudkowsky seemed particularly spooked by AlphaGo and AlphaZero in 2016–2017. +Now it looks like we have—less time? Not just tautologically because time has passed (the 21st century is one-fifth over—closer to a quarter over), but because of new information from the visible results of the deep learning revolution.[^second-half] Yudkowsky seemed particularly [spooked by AlphaGo](https://www.lesswrong.com/posts/7MCqRnZzvszsxgtJi/christiano-cotra-and-yudkowsky-on-ai-progress?commentId=gQzA8a989ZyGvhWv2) [and AlphaZero](https://intelligence.org/2017/10/20/alphago/) in 2016–2017. -[TODO: specifically, AlphaGo seemed "deeper" than minimax search so you shouldn't dimiss it as "meh, games", the way it rocketed past human level from self-play https://twitter.com/zackmdavis/status/1536364192441040896] +[^second-half]: In an unfinished slice-of-life short story I started writing _circa_ 2010, my protagonist (a supermarket employee resenting his job while thinking high-minded thoughts about rationality and the universe) speculates about "a threshold of economic efficiency beyond which nothing human could survive" being a tighter bound on future history than physical limits (like the heat death of the universe), and comments that "it imposes a sense of urgency to suddenly be faced with the fabric of your existence coming apart in ninety years rather than 1090." -My AlphaGo moment was 5 January 2021, when OpenAI released [DALL-E](https://openai.com/blog/dall-e/) (by far the most significant news story of that week in January 2021). Previous AI milestones, like GANs for a _fixed_ image class, were easier to dismiss as clever statistical tricks. If you have thousands and thousands of photographs of people's faces, I didn't feel surprised that some clever algorithm could "learn the distribution" and spit out another sample; I don't know the _details_, but it doesn't seem like scary "understanding." DALL-E's ability to _combine_ concepts—responding to "an armchair in the shape of an avacado" as a novel text prompt, rather than already having thousands of avacado-chairs and just spitting out another one of those—viscerally seemed more like "real" creativity to me, something qualitatively new and scary. + But if ninety years is urgent, what about ... nine? Looking at what deep learning can do in 2023, the idea of Singularity 2032 doesn't seem self-evidently _absurd_ in the way that Singularity 2019 seemed absurd in 2010 (correctly, as it turned out). + +My AlphaGo moment was 5 January 2021, when OpenAI released [DALL-E](https://openai.com/blog/dall-e/) (by far the most significant news story of that week in January 2021). Previous AI milestones, like GANs for a _fixed_ image class, were easier to dismiss as clever statistical tricks. If you have thousands of photographs of people's faces, I didn't feel surprised that some clever algorithm could "learn the distribution" and spit out another sample; I don't know the _details_, but it doesn't seem like scary "understanding." DALL-E's ability to _combine_ concepts—responding to "an armchair in the shape of an avacado" as a novel text prompt, rather than already having thousands of examples of avacado-chairs and just spitting out another one of those—viscerally seemed more like "real" creativity to me, something qualitatively new and scary.[^qualitatively-new] + +[^qualitatively-new]: By mid-2022, DALL-E 2 and Midjourney and Stable Diffusion were generating much better pictures, but that wasn't surprising. Seeing AI being able to do a thing _at all_ is the model update; AI being able to do the thing much better 18 months later feels "priced in." [As recently as 2020, I had been daydreaming about](/2020/Aug/memento-mori/#if-we-even-have-enough-time) working at an embryo selection company (if they needed programmers—but everyone needs programmers, these days), and having that be my altruistic[^eugenics-altruism] contribution to the great common task. Existing companies working on embryo selection [boringly](https://archive.is/tXNbU) [market](https://archive.is/HwokV) their services as being about promoting health, but [polygenic scores should work as well for maximizing IQ as they do for minimizing cancer risk](https://www.gwern.net/Embryo-selection).[^polygenic-score] Making smarter people would be a transhumanist good in its own right, and [having smarter biological humans around at the time of our civilization's AI transition](https://www.lesswrong.com/posts/2KNN9WPcyto7QH9pi/this-failing-earth) would give us a better shot at having it go well.[^ai-transition-go-well] @@ -538,13 +542,14 @@ My AlphaGo moment was 5 January 2021, when OpenAI released [DALL-E](https://open [^ai-transition-go-well]: Natural selection eventually developed intelligent creatures, but evolution didn't know what it was doing and was not foresightfully steering the outcome in any particular direction. The more humans know what we're doing, the more our will determines the fate of the cosmos; the less we know what we're doing, the more our civilization is just another primordial soup for the next evolutionary transition. -But pushing on embryo selection only makes sense as an intervention for optimizing the future if AI timelines are sufficiently long, and the breathtaking pace (or too-fast-to-even-take-a-breath pace) of the deep learning revolution is so much faster than the pace of human generations, that it's starting to look unlikely that we'll get that much time. If our genetically uplifted children would need at least twenty years to grow up to be productive alignment researchers, but unaligned AI is on track to end the world in twenty years, we would need to start having those children _now_ in order for them to make any difference at all. +But pushing on embryo selection only makes sense as an intervention for optimizing the future if AI timelines are sufficiently long, and the breathtaking pace (or too-fast-to-even-take-a-breath pace) of the deep learning revolution is so much faster than the pace of human generations, that it's starting to look unlikely that we'll get that much time. If our genetically uplifted children would need at least twenty years to grow up to be productive alignment researchers, but unaligned AI is [on track to end the world in twenty years](https://www.lesswrong.com/posts/AfH2oPHCApdKicM4m/two-year-update-on-my-personal-ai-timelines), we would need to start having those children _now_ in order for them to make any difference at all. + +[It's ironic that "longtermism" got traction as the word for the "EA" cause area of benefitting the far future](https://applieddivinitystudies.com/longtermism-irony/), because the decision-relevant beliefs of most of the people who think about the far future, end up working out to extreme _short_-termism. Common-sense longtermism—a longtermism that assumed there's still going to be a world of recognizable humans in 2123—_would_ care about eugenics, and would be willing to absorb political costs today in order to fight for a saner future. The story of humanity would not have gone _better_ if Galileo had declined to publish his works for fear of the Inquisition. -[It's ironic that "longtermism" got some traction as the word for the "EA" cause of benefitting the far future](https://applieddivinitystudies.com/longtermism-irony/), because the decision-relevant beliefs of most of the people who think about the far future work out to extreme short-termism. +But if you think the only hope for there _being_ a future flows through maintaining influence over what large tech companies are doing as they build transformative AI, declining to contradict the state religion makes more sense—if you don't have _time_ to win a culture war, because you need to grab hold of the Singularity (or perform a [pivotal act](https://arbital.com/p/pivotal/) to prevent it) _now_. -Common-sense longtermism—a longtermism that assumed there's still going to be a world of recognizable humans in 2123—_would_ care about eugenics, and would be willing to absorb political costs today in order to fight for a saner future. The story of humanity would not have gone _better_ if Galileo had declined to publish his works for fear of the Inquisition. +(I remarked to "Wilhelm" in June 2022 that DeepMind changing its Twitter avatar to a rainbow variant of their logo for Pride month was a bad sign.) -But if you think the only hope for there _being_ a future flows through maintaining influence over what big state-backed corporations are doing, declining to contradict the state religion makes more sense—if you don't have _time_ to win a culture war, because you need to grab hold of the Singularity (or perform a [pivotal act](https://arbital.com/p/pivotal/) to prevent it) _now_. [...] diff --git a/notes/memoir-sections.md b/notes/memoir-sections.md index 9780033..bf727cf 100644 --- a/notes/memoir-sections.md +++ b/notes/memoir-sections.md @@ -1,7 +1,7 @@ marked TODO blocks— +✓ AlphaGo seemed deeper [pt. 5] - "Agreeing with Stalin" recap intro [pt. 5] _ social justice and defying threats [pt. 5] -_ AlphaGo seemed deeper [pt. 5] _ scuffle on "Yes Requires the Possibility" [pt. 4] _ confronting Olivia [pt. 2] _ "Lesswrong.com is dead to me" [pt. 4] @@ -50,6 +50,9 @@ New (bad) time estimate: With internet available— +_ DeepMind June 2022 Twitter archive? +_ January 2021 Capitol riot +_ No Fire Alarm _ "not hard to find": link to more /r/itsafetish-like anecdotes _ Keltham's masochism insight _ stats of SIAI vs. SingInst hits (for ^siai footnote) -- 2.17.1