We _lost?!_ How could we _lose?!_ The philosophy here was _very clear-cut_. This _shouldn't_ be hard or expensive or difficult to clear up. I could believe that Alexander was "honestly" confused, but Yudkowsky ...!?
-I could see how, under ordinary circumstances, asking Yudkowsky to weigh in on my post would be inappropriately demanding of a Very Important Person's time, given that an ordinary programmer such as me was surely as a mere _worm_ in the presence of the great Eliezer Yudkowsky. (Such that I would have humbly given up much sooner without the social proof from Michael and Ben and Sarah and secret posse member and Jessica.)
+I could see how, under ordinary circumstances, asking Yudkowsky to weigh in on my post would be inappropriately demanding of a Very Important Person's time, given that an ordinary programmer such as me was surely as a mere _worm_ in the presence of the great Eliezer Yudkowsky. (I would have humbly given up much sooner if I hadn't gotten social proof from Michael and Ben and Sarah and secret posse member and Jessica.)
But the only reason for my post to exist was because it would be even _more_ inappropriately demanding to ask for a clarification in the original gender-political context. I _don't_ think it was inappropriately demanding to expect "us" (him) to _be correct about the cognitive function of categorization_. (If not, why pretend to have a "rationality community" at all?) I was _trying_ to be as accomodating as I could, short of just letting him (us?) be wrong.
But ... it's only "obvious" if you _take as a given_ that Yudkowsky is playing a savvy Kolmogorov complicity strategy like any other public intellectual in the current year. Maybe this seems banal if you haven't spent your entire adult life in his robot cult?
-But since I _did_ spend my entire adult life in his robot cult, the idea that Eliezer Yudkowsky was going to behave just as badly as any other public intellectual in the current year, was not really in my hypothesis space.
+But since I _did_ spend my entire adult life in his robot cult, trusting him the way a Catholic trusts the Pope, I _had_ to assume that the "hill of validity in defense of meaning" Twitter performance was an "honest mistake" in his rationality lessons, and that honest mistakes could be corrected if someone put in the effort to explain the problem. The idea that Eliezer Yudkowsky was going to behave just as badly as any other public intellectual in the current year, was not really in my hypothesis space. It took some _very large_ likelihood ratios to beat it into my head the thing that was obviously happenening, was actually happening.
+Ben shared the account of our posse's email campaign with someone, who commented that I had "sacrificed all hope of success in favor of maintaining his own sanity by CC'ing you guys." That is, if I had been brave enough to confront Yudkowsky by myself, _maybe_ there was some hope of him seeing that the game he was playing was wrong. But because I was so cowardly as to need social proof (because I believed that an ordinary programmer such as me was as a mere worm in the presence of the great Eliezer Yudkowsky), it must have just looked to him like an illegible social plot originating from Michael.
-"sacrificed all hope of success in favor of maintaining his own sanity by CC'ing you guys (which I think he was correct to do conditional on email happening at all)"
+One might wonder why this was such a big deal to us. Okay, so Yudkowsky had prevaricated about his own philosophy of language for transparently political reasons, and couldn't be moved to clarify in public even after me and my posse spent an enormous amount of effort trying to explain the problem. So what? Aren't people wrong on the internet all the time?
+Ben explained that Yudkowsky wasn't a private person who might plausibly have the right to be wrong on the internet in peace. Yudkowsky was a public figure whose claim to legitimacy really did amount to a claim that while nearly everyone else was criminally insane (causing huge amounts of damage due to disconnect from reality, in a way that would be criminal if done knowingly), he almost uniquely was not—and he had he had set in motion a machine (the "rationalist community") that was continuing to raise funds and demand work from people for below-market rates based on that claim—"work for me or the world ends badly", basically.
+If the claim was _true_, it was important to make, and to actually extract that labor. But we had falsified to our satisfaction the claim that Yudkowsky was currently sane in the relevant way (which was a _extremely high_ standard, and not a special flaw of Yudkowsky in the current environment). If Yudkowsky couldn't be bothered to live up to his own stated standards or withdraw his validation from the machine he built after we had _tried_ to bring it up in private with him, then we had a right to talk about what we thought was going on.
-At the start, I _had_ to assume that the "hill of validity in defense of meaning" Twitter performance was an "honest mistake" in his rationality lessons, and that honest mistakes could be corrected if someone put in the effort to explain the problem.
-
-
-It took some pretty large likelihood ratios to promote the "obvious" explanation
-
-
-
-
-
-But the guy doesn't _market_ himself as being like any other public intellectual in the current year. As Ben put it, Yudkowsky's "claim to legitimacy really did amount to a claim that while nearly everyone else was criminally insane (causing huge amounts of damage due to disconnect from reality, in a way that would be criminal if done knowingly), he almost uniquely was not." Call me a sucker, but ... I _actually believed_ Yudkowsky's marketing story. The Sequences _really were just that good_. That's why it took so much fuss and wasted time to generate a likelihood ratio large enough to falsify that story.
-
-Ben compared Yudkowsky to Eliza the spambot therapist in my story ["Blame Me for Trying"](/2018/Jan/blame-me-for-trying/). Scrupulous rationalists were paying rent to something claiming moral authority, which had no concrete specific plan to do anything other than run out the clock. Minds like mine don't surive long-run in this ecosystem. If we wanted minds that do "naïve" inquiry instead of playing savvy Kolmogorov games to survive, we needed an interior that justified that level of trust.
-
-[TODO: weave in "set in motion a machine" 19 Apr?]
-
+Ben further compared Yudkowsky (as the most plausible individual representative of the "rationalists") to Eliza the spambot therapist in my story ["Blame Me for Trying"](/2018/Jan/blame-me-for-trying/): regardless of the initial intent, scrupulous rationalists were paying rent to something claiming moral authority, which had no concrete specific plan to do anything other than run out the clock, maintaining a facsimile of dialogue in ways well-calibrated to continue to generate revenue. Minds like mine wouldn't surive long-run in this ecosystem. If we wanted minds that do "naïve" inquiry instead of playing savvy Kolmogorov games to survive, we needed an interior that justified that level of trust.
+[TODO: rewrite Ben's account of the problem above, including 15 April Signal conversation]
-------
-curation hopes ... 22 Jun: I'm expressing a little bit of bitterness that a mole rats post got curated https://www.lesswrong.com/posts/fDKZZtTMTcGqvHnXd/naked-mole-rats-a-case-study-in-biological-weirdness
-
-"Univariate fallacy" also a concession
-(which I got to cite in https://www.lesswrong.com/posts/cu7YY7WdgJBs3DpmJ/the-univariate-fallacy which I cited in "Schelling Categories")
-
-https://slatestarcodex.com/2019/07/04/some-clarifications-on-rationalist-blogging/
-
-"Yes Requires the Possibility of No" 19 May https://www.lesswrong.com/posts/WwTPSkNwC89g3Afnd/comment-section-from-05-19-2019
-
-scuffle on LessWrong FAQ 31 May https://www.lesswrong.com/posts/MqrzczdGhQCRePgqN/feedback-requested-draft-of-a-new-about-welcome-page-for#iqEEme6M2JmZEXYAk
-
-"epistemic defense" meeting
-
-[TODO section on factional conflict:
-Michael on Anna as cult leader
-Jessica told me about her time at MIRI (link to Zoe-piggyback and Occupational Infohazards)
-24 Aug: I had told Anna about Michael's "enemy combatants" metaphor, and how I originally misunderstood
-me being regarded as Michael's pawn
-assortment of agendas
-mutualist pattern where Michael by himself isn't very useful for scholarship (he just says a lot of crazy-sounding things and refuses to explain them), but people like Sarah and me can write intelligible things that secretly benefited from much less legible conversations with Michael.
-]
-
-8 Jun: I think I subconsciously did an interesting political thing in appealing to my price for joining
-
-REACH panel
-
-(Subject: "Michael Vassar and the theory of optimal gossip")
-
-
-Since arguing at the object level had failed (["... To Make Predictions"](/2018/Feb/the-categories-were-made-for-man-to-make-predictions/), ["Reply on Adult Human Females"](/2018/Apr/reply-to-the-unit-of-caring-on-adult-human-females/)), and arguing at the strictly meta level had failed (["... Boundaries?"](https://www.lesswrong.com/posts/esRZaPXSHgWzyB2NL/where-to-draw-the-boundaries)), the obvious thing to do next was to jump up to the meta-meta level and tell the story about why the "rationalists" were Dead To Me now, that [my price for joining](https://www.lesswrong.com/posts/Q8evewZW5SeidLdbA/your-price-for-joining) was not being met. (Just like Ben had suggested in December and in April.)
-
-I found it trouble to make progress on. I felt—constrained. I didn't know how to tell the story without (as I perceived it) escalating personal conflicts or leaking info from private conversations. So instead, I mostly turned to a combination of writing bitter and insulting comments whenever I saw someone praise "the rationalists" collectively, and—more philosophy-of-language blogging!
-
-In August's ["Schelling Categories, and Simple Membership Tests"](https://www.lesswrong.com/posts/edEXi4SpkXfvaX42j/schelling-categories-and-simple-membership-tests), I explained a nuance that had only merited a passion mention in "... Boundaries?": sometimes you might want categories for different agents to _coordinate_ on, even at the cost of some statistical "fit." (This was of course generalized from a "pro-trans" argument that had occured to me, [that self-identity is an easy Schelling point when different people disagree about what "gender" they perceive someone as](/2019/Oct/self-identity-is-a-schelling-point/).)
-
-[TODO— more blogging 2019
-
-"Algorithms of Deception!" Oct 2019
-
-"Maybe Lying Doesn't Exist" Oct 2019
-
-I was _furious_ at "Against Lie Inflation"—oh, so _now_ you agree that making language less useful is a problem?! But then I realized Scott actually was being consistent in his own frame: he's counting "everyone is angrier" (because of more frequent lying-accusations) as a cost; but, if everyone _is_ lying, maybe they should be angry!
-
-"Heads I Win" Sep 2019: I was surprised by how well this did (high karma, later included in the best-of-2019 collection); Ben and Jessica had discouraged me from bothering after I
-
-"Firming Up ..." Dec 2019: combatting Yudkowsky's not-technically-lying shenanigans
-
-]
-
-
-[TODO: Yudkowsky throwing NRx under the bus; tragedy of recursive silencing
-15 Sep Glen Weyl apology
-]
-
+[TODO: better outline 2019]
In November, I received an interesting reply on my philosophy-of-categorization thesis from MIRI researcher Abram Demski. Abram asked: ideally, shouldn't all conceptual boundaries be drawn with appeal-to-consequences? Wasn't the problem just with bad (motivated, shortsighted) appeals to consequences? Agents categorize in order to make decisions. The best classifer for an application depends on the costs and benefits. As a classic example, it's very important for evolved prey animals to avoid predators, so it makes sense for their predator-detection classifiers to be configured such that they jump away from every rustling in the bushes, even if it's usually not a predator.
I had thought of the "false-positives are better than false-negatives when detecting predators" example as being about the limitations of evolution as an AI designer: messy evolved animal brains don't bother to track probability and utility separately the way a cleanly-designed AI could. As I had explained in "... Boundaries?", it made sense for _what_ variables you paid attention to, to be motivated by consequences. But _given_ the subspace that's relevant to your interests, you want to run an epistemically legitimate clustering algorithm on the data you see there, which depends on the data, not your values. The only reason value-dependent gerrymandered category boundaries seem like a good idea if you're not careful about philosophy is because it's _wireheading_. Ideal probabilistic beliefs shouldn't depend on consequences.
-Abram didn't think the issue was so clear-cut. Where do "probabilities" come from, in the first place? The reason we expect something like Bayesianism to be an attractor among self-improving agents is _because_ probabilistic reasoning is broadly useful: epistemology can be _derived_ from instrumental concerns. He agreed that severe wireheading issues _potentially_ arise if you allow consequentialist concerns to affect your epistemics—
+Abram didn't think the issue was so clear-cut. Where do "probabilities" come from, in the first place? The reason we expect something like Bayesianism to be an attractor among self-improving agents is _because_ probabilistic reasoning is broadly useful: epistemology can be _derived_ from instrumental concerns. He agreed that severe wireheading issues _potentially_ arise if you allow consequentialist concerns to affect your epistemics.
But the alternative view had its own problems. If your AI consists of a consequentialist module that optimizes for utility in the world, and an epistemic module that optimizes for the accuracy of its beliefs, that's _two_ agents, not one: how could that be reflectively coherent? You could, perhaps, bite the bullet here, for fear that consequentialism doesn't tile and that wireheading was inevitable. On this view, Abram explained, "Agency is an illusion which can only be maintained by crippling agents and giving them a split-brain architecture where an instrumental task-monkey does all the important stuff while an epistemic overseer supervises." Whether this view was ultimately tenable or not, this did show that trying to forbid appeals-to-consequences entirely led to strange places. I didn't immediately have an answer for Abram, but I was grateful for the engagement. (Abram was clearly addressing the real philosophical issues, and not just trying to mess with me the way almost everyone else in Berkeley was trying to mess with me.)
projects / Ultimately_Untrue_Thought.git / blobdiff