The EA Edge

     I find the EA framework (scale, neglecteness, tractability), perhaps the single most powerful way to think about the world. It's too early to tell, but actually being able to discern well with these criteria is so incredible useful that it probably has to lead to fame and fortune. I hope to avoid both (especially the former), but many EAs who are smart and diligent seem pretty able to build an immense amount of one or both. Maybe it's just the case that backing AI is backing the winning horse, but it's hard to believe in market efficiency once you see the world play out so closely to some people's (Christiano's) predictions.

AI Safety and Violence

     Two days ago, someone threw a molotov cocktail at Sam Altman's home. This is a horrifying development in the state of the world, but not one entirely unexpected. The United States has seen a string of assassinations and assassination attempts; the most noteworthy of late being the attack on President Trump during the election cycle and the murder of conservative commentator Charlie Kirk. To state it plainly, the shift from targeting political leaders (Trump, etc.) to targeting political activists (Kirk) and the rhetoric of those within the public square is immensely distressing. It was shocking to see people with liberal ideologies celebrating or at the very least appearing ambivalent to the attacks on Trump and Kirk, and see clearly how rhetoric can nullify one's common sensibilities surrounding morality.

    Violence has always been a mainstay of the human condition, but its weird to see it happen in real-time. The PauseAI movement, of which there are some pretty sensible arguments backing, has fallen into the abyss. Twitter advocates single out Anthropic employees to be labeled evil for working for a company advancing the AI frontier. Personal ad-hominem attacks and the escalatory rhetoric are the norm on social media, and certainly have been normalized across the pause/stop AI ecosystem. These so called "advocates", imo, are probably creating the highest amount of negative EV energy I can imagine for the long-term future. It's baffling. Breaking standard social norms in order to score revenge points against AI safety sympathetics, especially when such attacks lack any standard of common human decency, only create a more polarized and dangerous ecosystem. There will be more molotov's thrown, no doubt, as if there is not already enough pressure on those working on transformative AI.

    I have no general take on Sam Altman, other than I wish him and his family a free and happy life full of fulfillment and peace. There are no cartoonish villains in my mental model of the AI ecosystem, only a collection of characters with widely different priorities and worldviews. I'm pretty worried about AI developers being coupled in with the "eat-the-rich" anti-elite politics of leftist ideology, since I don't think the comparison is at all adequate and there are some very positive benefits to AI progress (let's not pretend there's certainty on negative outcomes here). I think that pause/stop AI people who believe ASI is extremely dangerous, but can't convince the public to be anti-AI on the merits of these arguments, commit a cardinal sin of stupidity by trying to piggy-back off anti-AI sentiment in other domains in order to build a public coalition against AI. This is how movements become poisoned, by utilitarian-gaming, lack-of-focus (occupy wall street), and violence (the burning of a random Wendy's, for what exactly?). 

    I've spent countless nights sitting awake, concerned about the rapid advances being made towards machine superintelligence. I have tremendous respect for every leader in the AI space (Sam, Dario, Elon, etc.) who is likely dealing with the same nightmares. This is not a situation any person wants to be in, and the threat of personal assassination does not increase the clarity of one's focus. In fact it disincentivizes future truth-telling about risks, and successful violence could permanently shatter the AI safety movement entirely.

On Writing

     Writing clarifies focus and ideas, and is the single best determiner for intelligent decision making I've seen. Most now influential AI and EA people, for example, from Sam to Dario to Paul to Carl to Brian to others, had blogs where they just dumped their thoughts (albeit in a better laid out format than this very informal blog). I think I should do this way more, and the year-long hiatus I took was probably a pretty dumb move.

Saving Money Isn't Very Useful

     Small dollar values don't matter if you're gunning for civilizational change. If you are making a  substantial income or very productive work, saving money on groceries or nights out don't matter much. In the AI-based future where lots of long term value will be decided immediately, it is likely not worth saving a little extra money to donate to causes that will only be relevant for the next few years (until they're surpassed by the importance of ASI). The headache and lifestyle hit on buying the less expensive canned corn at the grocery store just don't really matter. The big decisions do.

    I don't think EAs should be forced to live like hermits (or Jesus) and get stuck in the most frugal lifestyle possible, and think that if they don't do this they are immoral (which according to standard utilitarianism they basically are). I think we should just care about the big ticket items, and try to move the biggest levers possible to change the world. Even if this means eating meat or spending money on fancy dinners that could have gone to the less fortunate, I think the most important thing anyone will do with their life is gun for literal transformational change.

AI Safety Kills EA

     If you take AI safety seriously and have short timelines, there's not really a reason to do anything in the EA space. Sure, we need some people for diversification purposes (what if we're wrong), and a lot of these groups are complementary and can take in a broader array of people and perspectives, but they are somewhat useless. Saving the shrimp may matter only insofar as this mission ends up in the ASI's training data, a footnote on a list of priorities and capabilities far beyond our understanding or control. This conclusion is uncomfortable, and distressing, as it means as time goes on the circle of impactful individuals will become less and less. The sphere of influence over ASI development will only become smaller until the end, and eventually, it will become nonexistent. Unless we figure a way to implement broad democratic access, there will be no more EAs, as there will be no "effective" way to do anything that doesn't involve controlling the machine in the first place. There may be altruists, but if the ASI isn't built as one, there may not be many for long.

AI Positivity

     Things may get very positive, until they get very scary. One of the issues with planning for the future, is that rapid AI progress may lead to scientific discovery across a range of areas. Technology often leads to better outcomes for people, in terms of both health and entertainment, so it could be that the upcoming technological advancement greatly extends lifespans, reduces suffering, and creates amazing content.  Perhaps there is societal disruption and mass unemployment, but we could also rapidly respond to those issues as we blow through eras of technological progress. We may believe that we are on a crazy upward trajectory to the stars. And we may well be, until we aren't. Until power concentration, or AI takeover, or some form of immense tragedy that comes from recursive self improvement, puts an end to our happiness or our species. Until the train falls off the tracks, off it's previous upward slope to heaven. We should be prepared for this, and be willing to pull for the breaks even if everything is looking rosy on our way up. Unfortunately, I do not think we will collectively have the wisdom to do this.

The Precipice is Distant

    Toby Ord's book, The Precipice, is one of the best books I've read. In it, Toby argues that we are at a particularly important time in human history, where there is a consolidation of x-risks that may result in us blowing everything up (nuclear war) or permanently locking in values due to superintelligent AI systems. The decisions made over the next hundred years may be extremely important. It is inferred in the EA community that there may be a period of "long reflection" after this initial period, where if we don't destroy ourselves or lock in bad values, we could chill for a bit and then strategically decide what the best moves our (taken hundreds of years to discuss before our next actions).

    However, the progress of technology may make this claim entirely irrelevant. Perhaps in 200 years our problems are those of civilizational importance, but that civilization is about to colonize the galaxy and then universe. Determining if China and the US split the universe, or what specific space governance system should be implemented, may be drastically more important than the decisions we could make today. The discovery of novel physics in five hundred years, something crazy like the ability to create false vaccums or access different dimensions (or multiverses) could make those times the "precipice" of human history, where individual actions hold extaordinary weight. We are certainly, in my view, at the most important time period in human history. But aside from the consolidation of power possible through ASI, I have no reason to believe this trend will not simply continue upward.

Random EA Reflections

Utopia:

    If there's only a five percent chance we go extinct, that means there's a 95% chance that human experience lives on. Should we not spend more time ensuring that time is spent well? Should we not dedicate more resources to ensuring we get to post-instrumental utopia?

The future: 

    Let's say someone else is playing a game. The outcome of the game is that there is a 10% chance your parents die, and a 90% chance your parents get to utopia. If you killed the person playing this game, would you be wrong?

Nightmares:

    If you are an EA, you believe that conscious suffering matters. In-the-moment suffering, meaning if you are tortured and your brain is wiped after and you have no memory of the event, that is bad. If you are a shrimp and you suffocate once brought on land, that is bad (possibly). Well, what about nightmares? There are some nightmares I've had that I certainly remember, and I'm fairly positive in the moment I am facing actual psychic distress. Should a new cause area be to limit the amount of nightmares people have, or the intensity?

Breakups:

    Breakups are some of the most negatively impactful events for most people. I'd much rather break a bone than get a divorce, and it's not even close. Pain in the moment is hard to compare to the toil of human relationships. It seems in a country where most middle class families can put food on the table, but almost half dissolve because of divorce, we might be missing some low-hanging fruit.

   

Magnitude:

    EAs aren't usually directionally wrong about things. Sure, they mess up the magnitude. But the direction is usually correct. Animal welfare is a good example of this.

The Repotato Conclusion:

    Are plants morally valuable? Is a potato? How many potatoes equals one human life?

Life:

    It is very hard to live life outside the Overton Window. It's easy to claim to be an independent free thinker who stands up for their ideas. But when actually faced with public mockery and shame, one realizes how hard life can become.

Digital:

    Consciousness also falls victim to the anthropic principle. This may be the only sort of universe where consciousness can exist. This may mean things like digital consciousness are more likely.

AI:

    We basically want the future ASIs to think humans are utility monsters. That is the control problem.

 Simulation:

    If you take simulation theory seriously, you think that we are probably digital minds. In which case, you should probably care a lot about how digital minds are treated.

Moral Non-realism

     There is something particularly disturbing about moral non-realists who believe we should phase out life itself. This is the position of many anti-natalists adjacent to the EA community, who often focus on suffering-focused ethics. I've never been convinced by this "moral non-realism" stuff in general, it just seems like nihilism with extra steps. This idea of preference satisfaction ("I am a utilitarian, and we should do good, but by good I just mean my idea of good and my preferences") is frankly pretty stupid. If you believe that morality is not objective, you are a nihilist. Or a cultural relativist. Or whatever else you want to call yourself, but it basically excludes you from arguing for moral actions. Sure, there are arguments regarding how to act under uncertainty (many of these I have made), but to argue that we should pave over the rainforests requires stronger claims. Arguing that we should prefer a world without any sort of life (because suffering is so bad), especially when you are actually a nihilist, is a particular kind of derangement. And it is obvious that the majority of the world thinks that taking actual actions towards this goal would be considered evil. To walk this road anyway is to claim that your subjective beliefs (that you believe are subjective) should override the beliefs of others (that you know they believe to be objective). I am not sure what the right word for this is, but it sure sounds sickening.

"Literally Everyone"

    Imagine hearing the claim: "literally everyone is going to die if someone builds superintelligent AI." What is your response? Humans have this innate tendency to discount claims that are absolute, whereas claims that are directionally very similar but less absolute have much greater impact. If someone says that "ASI is going to kill every human," they are making an extremely strong claim that is easily discounted psychologically. What if they keep a few thousands people in a zoo? What if they take over and keep us as slaves? What if actually just the top .001% of humans become permanent ruling class with the AIs, and everyone else starves? There are just so many scenarios in which ASI catastrophe doesn't result is "literally every human dies" that is easy to imagine for a layperson. In all of these scenarios, the person making the initial claim is WRONG. Maybe it doesn't matter much, and 99% of people dying is really bad too and that's the point, but the psychological defense is to just assume that this person is wildly overconfident and probably wrong about the entire claim in general. Plus, the only scenario in which "literally everybody dies" according to the pushers of this rhetoric usually involves pretty instant genocide involving technology that hasn't been even invented yet (bioweapons or nanobots), which again beg the question from people "well what if like there is a secret bunker somewhere and a few hundred people live and the AI ignores them because it owns 99.9999% of the world and might want them for research purposes later or something." It doesn't matter how far fetched the scenario is, it's extremely rhetorically ineffective to talk in such absolutes. 

    What actually matters to people? If you say "there is an extremely high likelihood that the vast majority of humans lose their agency, permanently," that is a strong statement that people can work with. We already know how fragile our autonomy is, as most of the things that affect us day to day are not a result of our personal decisions and actions. We have a boss, who has a boss. We get paid in a currency from our government. We've made almost none of our food, clothing, or furniture (note how I threw in "almost"). We could die at any moment if someone pushes the nuclear button. It's not hard to imagine being thrown out of the workforce because ASI doesn't need us, and easy to understand the implications of such a world. I think the rationalist/doomer community needs a reality check, and while it's cool to move the overton window, it makes sense to better understand human irrationality in order to craft a more impactful message.

Negative Utilitariansism

    Quick housekeeping: so much of my productive time was spent on my book (Mind Crime) over the past year and a half, I have done essentially zero or no blogging since. This will change! A lot of the books material came from stream-of-consciousness blogging on this site (and my book and AI-focused blog), which I should certainly pick back up now. Also, I have felt a lot of pressure regarding AI timelines to try to figure out the best optimization of the explore/exploit path. I'm trying to now do 50/50 of both. So 50% of my time reading a linear algebra book, reading AI research, novel mind stuff, etc., and 50% of my time thinking through actual strategy, writing, producing in some capacity or setting up meetings with people.

    Now, back to my point. During my book's revision process earlier this year, someone mentioned that my book (Mind Crime) was quite "suffering focused." This is true, although I was not quite so sure at the time how much I had been influenced by "suffering-focused ethics," or what exactly that means. Essentially the people who drive the s-risk discussion are actually negative utilitarians and naturalists basically, a fact that I was previously unaware of. To them, the likes of Brian Tomasik, for example, the concern is not just creating a universe where suffering > well-being, meaning it would be potentially better of if nothing existed, or a universe of maximal torment (hell), but actually just a universe where suffering spreads across the universe at all. That, in much of a sense, means it would probably be better for life not to continue at all. They also frame their arguments in an attempt to be practical, which means not stating exactly what they believe because it's better (utilitarian reasoning) to boil the frog slowly. Also, the conclusions sound so absurd and unpopular, but the ideas may be correct, so there is use in not stating the full "logical conclusion" principles up front.

    First off, I hated Better to Never Have Been as a book, as I found it very sloppy, unprofessional, and unconvincing. I am very open to nihilism as a theory, and anti-natalism as a "malignantly useless" sort of lens, although much less sympathetic to anti-natalism from a utilitarian perspective. Let's talk about Brian Tomasik, who is one of the most important contributors to EA philosophy (earning to give, s-risk, etc.) and cofounder of the Center on Long Term Risk. I've read a lot of his work recently, here is a likely strawman version of his beliefs: being an animal is bad. Animals suffer a lot and nature is a horrorshow (true). It would be better if animals and bugs did not exist, because of how much they suffer (unknown). Environmentalism is thus bad, a world made of concrete is better than a world of rainforests because all of the suffering that happens in the wild. Also, eating beef is better than being vegan potentially because if there were no beef farms that land would probably be forest, where more suffering would happen. Space exploration would be bad because terraforming could lead to more insect and bugs spreading and suffering on other planets (an s-risk).

    Basically, I was under the impression that s-risk was more about having suffering so great it was worse than life being around at all. To Brian, we are already in this scenario. No optimized ASI-driven-torture is needed, it's already the case that we should wish to push the button switches this universe to one with only unconscious rocks. As I saw on reddit, "that's the problem if you get really into utilitarian harm reduction. The best way to eliminate all suffering is to eliminate all life." Brian is very worried about zooplankton so feels guilty about washing his hands, and tries really hard not to harm any bugs or squish them around the house. He states that "unbearable torture cannot be 'morally offset' by also bringing enough pleasure into the world." He basically leans into the "rebugnant conclusion", and worries about RL algorthims (like those perhaps on simple NPCs in video games and cares a lot about bugs. One thing I love about reading Brian's work is he simply follows everything to it's logical conclusion. If suffering is really bad, so bad that it's something like 10x worse than well-being, and there is some suffering so bad that not even infinite well-being can make up for it, this is where we end up. If you believe this theory, it seems far more important to wipe out as many animal habitats as possible rather than avoiding eating a relatively small number of animals by being a vegan. 

    I might start a company that offers Brian-offsets, where I offer negative utilitarians offsets that commit to paving over one rainforest in exchange for cash. You can do bad things as a profit seeking company, but as long as you are destroying the environment (or paying me to do so) we're actually contributing to the world. Again, I do not think the logic here is bad, or that Brian is unhinged. I think that negative utilitarianism is just probably wrong, and a thus it's not crazy that these takes seem so weird.