= 27d567bc2d48d9f40e9ccc20ea370b15)
Eliezer_Yudkowsky comments on Harry Potter and the Methods of Rationality discussion thread, part 19, chapter 88-89 - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
Comments (957)
It's a large space, not a binary yes-or-no, so successful predictions are impressive even given a large base. Also I could be prejudiced but MoR is supposed to be solvable god damn it.
Someone was criticized. S/he was right, the critics were wrong. The neural net updating algorithm calls for a nudge in the appropriate direction of "Beware of dismissing those who speak with what you think is too much confidence."
I update in favor of "user 75th is more experienced in the tropes of enigma fiction." Indeed I would not be at all surprised were I to discover that user 75th writes such fiction him or herself. It similarly wouldn't surprise me if user 75th had gone to the library and checked out and read some of the same 15 books Elizier checked out and read before writing HPMoR.
For example, before reading the author's notes on HPMoR I was not familiar with Chekhov's Gun. Now that I am, I am much more likely to catch such a device when it appears in other fiction. I now suspect user 75th is quite familiar with Chekhov's Gun and other standard tricks of this sort of story. 75th picked up on one such trope (one I'm still not familiar with) that signaled that Hermione was heading for death.
If there's a general update to be had here, it may go something like this:
Before dismissing those who speak with what I think is too much confidence, I need to consider the possibility that their confidence is based on facts or experience I am not aware of. I should probably take five minutes to ask them why they are so confident before dismissing them.
Ha, you've got me all wrong. I am woefully under-read, particularly in fiction. I get a very small percentage of the references Eliezer makes in Methods; most of the time, I find out that he's borrowed something months (or, let's face it, years) after I read it, only by seeing someone else explicitly point out the reference. I have had my life ruined by TV Tropes, but most of what I'm familiar with there is video games, and not too awfully many of those.
But it's not a matter of picking up on specific tropes, exactly. It's more a matter of getting into the author's head. Of constantly asking "If this were foreshadowing or a setup or a clue, what would be the most effective payoff?" I read Chapter 84, and then, put together with many other quotes from my many rereads of HPMoR ("Nothing really bad ever happens at Hogwarts", "Her life was officially over", etc.), I answered that question with "Hermione will die horribly," then posted how I felt about it.
It's the same deal with my prediction — which I'm far more certain of than I was that Hermione would die horribly — that Nzryvn Obarf xvyyrq Anepvffn Znysbl. I got into an argument with someone on Reddit once, who was bringing up all sorts of random canon characters who haven't even been introduced in Methods and coming up with categories in which to score them 1 to 5, then giving them wildly inappropriate scores and adding them up. She said things like "We're in a state of abysmally low evidence" and "At least I appreciate what Eliezer is trying to do," claiming to be following the spirit of Bayes better than me, when all I was doing was pointing out three successively more blatant clues that Eliezer put into the text and saying that my certainty had increased as they were successively pointed out to me.
I see a lot of that, it seems. A lot of people thinking the whole Bayes thing is about immediately reducing a situation to a bunch of numbers and seeing what the numbers say, and fie upon anyone who acts sure of something without attaching their spreadsheet. But it seems to me the numbers themselves are not the most important part of Bayesian thinking. Perhaps the least important part.
Re the rot13 bit, I called it that Qhzoyrqber xvyyrq ure based on text evidence before that was revealed, so the idea that it's Obarf has always seemed wrong to me. They can't both have done it, you know?
Arf, didn't mean to start this again, but here's my usual litany:
Gur bayl rivqrapr jr unir gung Qhzoyrqber xvyyrq Anepvffn vf gung Yhpvhf fnlf Qhzoyrqber gbyq uvz fb. Jr qba'g xabj gur rknpg jbeqf Qhzoyrqber hfrq, naq oheavat fbzrbar nyvir ernyyl qbrfa'g frrz yvxr Qhzoyrqber'f fglyr (nygubhtu V jvyy fnl gung Puncgre 89 vf gur svefg gvzr V'ir gubhtug gur Qhzoyrqber-vf-rivy pebjq zvtug npghnyyl unir fbzrguvat fhofgnagvir gb jbex jvgu). Zrnajuvyr:
Jura V bayl xarj nobhg #1, V jebgr vg bss nf n cbffvoyr pbvapvqrapr. Ohg gura crqnagreevsvp cbvagrq bhg #2 gb zr, naq V fgebatyl hctenqrq gur ulcbgurfvf'f cebonovyvgl. Gura yngre #3 unccrarq, naq V orpnzr nf pregnva nf V nz abj.
And if that's not as close as you can actually come to a Bayesian updating process when reading a fiction book, where the only experiment you can perform is "Wait for more chapters and then read them", I would love to learn what's legitimately closer.
1) Gryyvat fbzrbar gung lbh qvq fbzrguvat frrzf yvxr sne fgebatre rivqrapr gb zr guna hfvat n fvzvyne jbeq bapr.
2) Erzrzore, gur oheavat unccrarq evtug nsgre Noresbegu jnf xvyyrq, fb Qhzoyrqber'f zbgvir vf nyzbfg pregnvayl fgebatre guna Nzryvn'f.
3) Nyy guvf erdhverf vf na nyyl pybfr rabhtu gb xabj jung lbh'er guvaxvat naq gb trg lbh gb onpx qbja. V qba'g qbhog gung Nzryvn xabjf nobhg vg, V whfg qba'g guvax fur jnf gur bar jub crefbanyyl qvq vg. Tvira gung Obarf naq ZpTbantnyy ner(V guvax) gur bayl BBGC zrzoref va gur ebbz ng gur gvzr, vg'f irel jrnx rivqrapr - gur cebonovyvgl gung fur'f fcrnxvat hc nf na nyyl vf nyzbfg nf uvtu nf gur cebonovyvgl gung fur'f fcrnxvat hc nf gur thvygl cnegl.
Lbhe gurbel vf abg penml, ohg vg'f yrff cebonoyr guna gur Qhzoyrqber gurbel fb sne nf V pna gryy.
I think you miss that this is a work of fiction that has an author. Think about why the author was motivated to make certain choices instead of thinking why the characters were motivated.
That is actually the other reason that I believe as I do. It seems like a much more interesting storytelling decision for my theory to be correct than for the competitor theory to be.
How is it more interesting storytelling for the guy everyone thinks did it to have done it, as opposed to "Here's a puzzle and the clues to figure it out"?
Because it was a puzzle for the first what, 80 chapters? We're getting close to the end of the story as of about two arcs ago, it's time for puzzles to get solved. it's not an interesting enough puzzle to justify a double twist, IMO.
Amelia Bones isn't a member of the Order of the Phoenix.
One wonders why she would even know about it at all, if she had nothing to do with it.
Well, I think Lucius probably made sure a long time ago that everyone knew what Dumbledore (supposedly) said to him. I didn't get the feeling from that scene in the Wizengamot that Dumbledore-killing-Narcissa was any kind of a secret idea that people were just then finding out about.
This does rather change my view of some of the peripheral details, though. Previously, one possibility I pictured was Dumbledore restraining Amelia from her vengeance until Aberforth died, then relenting. I knew Amelia Bones wasn't in the OotP, and I knew she felt distaste at Dumbledore's softness, but somehow I never completely drew the conclusion that she wouldn't care one whit about what Dumbledore said or thought, and therefore probably wouldn't have cared if he had tried to restrain her.
Perhaps more likely, then, is the other way I pictured it: that Amelia couldn't get to Narcissa by herself, and after Aberforth's death, Dumbledore actively approached Amelia and said "Okay, I'm ready to help. I'll be the ward-breaker, you do the deed."
Whoops, looks like you're right, the accusation was public knowledge:
Which is why I think it a significant possibility that Dumbledore helped in some way.
Any of those defenses might be sufficient for a single clue, but you have to take the clues together. Three successive clues (plus her character) pointing to Amelia, and only some words spoken to Lucius that we never saw pointing to Albus (and everything else we know of his character pointing away from him, though I know some would argue that), increase the probability of Amelia's guilt quite a bit more than linearly.
I heard that it was na nppvqragny frys-xvyy ol Anepvffn hfvat Svraqsler ntnvafg fbzr rarzl.
I prefer the theory that qhzoyrqber hfrq svraqsler gb qrfgebl gur qvnel ubepehk jura ur gubhtug gur ubhfr jnf rzcgl, naq gura pynvzrq perqvg sbe anepvffn'f qrngu fb gung ure fnpevsvpr jbhyq abg or zrnavatyrff
Really? I remember being about eight years old, watching an episode of Power Rangers, and seeing some random no-name tourist drop a camera and the shot lingered on it for about half a second. My instant thought was "Oh, the bad guy is going to be built from a camera", and sure enough it was. They'd never have put that in the episode otherwise, and that was obvious to me at that age.
http://xkcd.com/1053/
No, it doesn't, not out of thousands of predictions of which you selected one post hoc. If I may quote you, our minds do not run on floating point beliefs.
If 75th's predictions were just luck, wouldn't it be likely for there to be other people who got a smaller number of the predictions right?
Yes. If you look through the threads or the PredictionBook entries, there are plenty of people blowing predictions. (A particularly good example was the Wizengamot trial: how would Harry rescue Hermione? The actual solution was 1 of the top 2 or 3 suggestions, but that still implies a lot of people favoriting wrong solutions.)
Even if 75ths predictions aren't just luck, you don't have enough information to meaningfully update across such a broad reference class. If it's got to overcome the weight of everyone I think is speaking with too much confidence on the other end of the lever, it's not going to move far enough to be noticeable.