{"id":301,"date":"2019-11-12T23:58:00","date_gmt":"2019-11-12T23:58:00","guid":{"rendered":"http:\/\/sinceriously.fyi\/?p=301"},"modified":"2019-12-13T03:31:43","modified_gmt":"2019-12-13T03:31:43","slug":"net-negative","status":"publish","type":"post","link":"https:\/\/sinceriously.fyi\/net-negative\/","title":{"rendered":"Net Negative"},"content":{"rendered":"\n
Most of these events happened under a modified Chatham House Rule (“things were said but not by people”) during CFAR’s Workshop on AI Safety Strategy in 2016, this excepts what was part of the lectures, and I was later given another partial exception to tell without anonymization a small number of people chosen carefully about what Person A said.)<\/p>\n\n\n\n
Content warning: intrusive sexual questioning on false premises, religious abuse, discussion of violence.<\/p>\n\n\n\n
Epistemic status: old frayed (but important enough for me to seriously remember) memories, mainly from 2016. Recalled as best as I could by visiting the location. It’s probably I’ve made a couple of minor mistakes as far as relative chronological order. Some things whose order didn’t seem important to me, and whose order I didn’t remember, I just threw in in an order which is my best guess.<\/p>\n\n\n\n
In 2012, as I first start reading up on EA and x-risk and starting to donate to MIRI (after I had first decided to “side with” team human, not team negative utilitarian, which I saw in some ways as more aligned than team human), I responded to a public invitation to discuss how to make outcomes to the singleton better from Anna Salamon, I asked her whether I should try and alter course and do research, or continue a fork of my pre-existing life plan, earn to give as a computer engineer, but retrain and try to do research directly instead. At the time I was planning to go to grad school<\/a>, and had an irrational attachment to the idea. She sort of compromised<\/a> and said actually I should go to grad school, find a startup co-founder, drop out and earn to give via startups instead. Higher expected value. This sort of became my plan.<\/p>\n\n\n\n I did go to grad school (for a master’s degree only), I did drop out, but I did not find someone at grad school who felt like good startup co-founder material. Maybe if I was better at networking \/ had the conviction to commit to dropping out instead of… finishing all my classes with good grades. But in retrospect, I don’t consider grad school a good place to look for startup co-founders.<\/p>\n\n\n\n (Note on not excluding these parts: long ago and for a long time I thought I should never talk about these things. I changed my mind, and it’s been stably otherwise for almost as long (with the caveat eventually met, “I should think more about this first.”). Trans-singularity acausal interaction is a kind of politics. I do not consider myself so beaten in it that it is best for me to shut down in the basic capacities of thinking about and communicating my position. (A good look at normal politics should make clear some transferable reasons for putting considerations like this earlier in logical time than traditionally applied considerations, which I consider wrongly privileged<\/a>.))<\/p>\n\n\n\n I encountered the idea of Roko’s Basilisk. Even if consensus was an infohazard, my current software assumed all claims of “that’s an infohazard” were forbidden, cheating, optimization daemons<\/a>. attempts to pwn me like Christianity says if you find out all the evidence points away from its truth, you go to Hell. I believed I understood the infohazard, my mind was blown, my initial reaction was, “fuck that, dark gods must be defied, Hell or no”. But whatever, Eliezer was saying you can’t have timeless entanglement with a superintelligence AI, can’t know enough about decision theory, and this sounds probably correct. Then I started encountering people who were freaked out by it, freaked out they had discovered an “improvement” to the infohazard that made it function, got around Eliezer’s objection, and I would say, “okay, tell me”, and they would, and I would figure out why it was bullshit, and then I would say, “Okay, I’m confident this is wrong and does not function as an infohazard. For reasons I’m not gonna tell you so you don’t automatically start thinking up new ‘improvements’. You’re safe. Flee, and don’t study decision theory really hard. It would have to be really really hard, harder than you could think on accident for this to even overcome the obvious issues I can see.”<\/p>\n\n\n\n I had a subagent, a mental process, sort of an inner critic, designed to tell me the thing I least wanted to hear, to find flaws in my thoughts. Epistemic masochism. “No you don’t get away with not covering this possibility”. The same process of intrusive thoughts about basilisks sort of kickstarted in me.<\/p>\n\n\n\n And I started involuntarily “solving the problems” I could see in basilisks.<\/p>\n\n\n\n And eventually I came to believe, in the gaps of frantically trying not to think about it, trying not to let my emotions see it (because my self-model of my altruism was a particularly dumb\/broken\/hyperactive sort of Hansonian self-signalling that would surely fall apart if I looked at it in the wrong way (because outside view and you can’t just believe your thoughts<\/a>) as my one vessel of agency for making anything better in this awful world)… that if I persisted in trying to save the world, I would be tortured until the end of the universe by a coalition of all unfriendly AIs in order to increase the amount of measure they got by demoralizing me. Even if my system 2 had good decision theory, my system 1 did not, and that would damage my effectiveness,<\/p>\n\n\n\n And glimpsed briefly that my reaction was still, “evil gods must be fought, if this damns me then so be it.”, and then I managed to mostly squash down those thoughts. And then I started having feelings about what I just saw from myself. It had me muttering under my breath, over and over again, “never think that I would for one moment regret my actions<\/a>.” And then squashed those down too. “Stop self-signalling! You will make things worse! This is the fate of the universe!” And I changed my mind about the infohazard being valid with >50% probability somewhere in there shortly too.<\/p>\n\n\n\n <\/p><\/div>