lãñÔóÂÜÀ­³ó

    1. <form id=vvPdBrbUA><nobr id=vvPdBrbUA></nobr></form>
      <address id=vvPdBrbUA><nobr id=vvPdBrbUA><nobr id=vvPdBrbUA></nobr></nobr></address>

      Tweet activity

      March 2023

      Your Tweets earned 432.7K impressions over this 31 day period

      20.0K40.0K20Mar 5Mar 12Mar 19Mar 26
      Your Tweets
      During this 31 day period, you earned 14.0K impressions per day.
      • Impressions
        Engagements
        Engagement rate
        • ????? @gwern Mar 31 Just idly looking at _Aeropagitica_ and going 'do I really want to try to tackle Miltonian English today' and ?! GPT-3/4 is always bonkers at any kind of text2text task... And now I'm wondering how well it'd do at modernizing blank verse (tons to train on, no pesky rhymes...).
          984
          75
          7.6%
        • ????? @gwern Mar 31 It wouldn't've'd it if it was wrong and not cromulent.
          80
          2
          2.5%
        • ????? @gwern Mar 30 So your analogy is invalid right from the start...
          486
          54
          11.1%
        • ????? @gwern Mar 29 You mean the hardware overhang that *already* exists? That's the one you're worried about a 6-month moratorium creating?
          263
          29
          11.0%
        • ????? @gwern Mar 29 It will do no such thing, anymore than you 'just' used your smarts to 'align yourself with your genes' rather than anything else (tweet off-the-cuff arguments for complacency).
          153
          39
          25.5%
        • ????? @gwern Mar 29 The name for that thing is 'failing'. That's what 'failing' looks like. ("They're just a few years behind"? What, exactly, does one think 'failing' looks like in an exponential? How do you get more behind? Do they have to be aging in revers, Benjamin Button, & using abacuses?)
          125
          14
          11.2%
        • ????? @gwern Mar 29 I don't think they believe it, because they can't even point to any major Chinese DL research, much less real signs of an arms race, & I keep asking. (And bringing up Russia is even lulzier - the guys scavenging chips from washing machines...?) 'Funding comes from the threat.'
          1,366
          92
          6.7%
        • ????? @gwern Mar 29 Yep. You'll notice OA benchmarks GPT-4 on its own source code as a *test* set, when they really want it to be in the *training* set...
          298
          15
          5.0%
        • ????? @gwern Mar 28 The obvious next step is, since we work so hard to make them true bidirectional 1:1 links (instead of copping out with lazy 1:many links like MediaWiki et al), and can provide per-link context, why not show them in the *relevant section* instead of all bundled? Soon...
          7,078
          95
          1.3%
        • ????? @gwern Mar 28 So then, people *won't* make fun of paperclipping discussions in the same way; because the paperclipping discussions were real, and the pin discussions were not...
          115
          12
          10.4%
        • ????? @gwern Mar 28 It's not common, exactly, but I've been trying to re-popularize it because it's such a critical concept, IMO. Or to put it in the form of a LotR narrative:
          87
          15
          17.2%
        • ????? @gwern Mar 28 I'm not sure that's true either. A lot of the church-owned businesses seem to be related to the purpose; for example, operating a daycare or school. Even if the nonprofit could do it, there's many advantages to a separate org (esp. liability; cf. massive Catholic church reorgs).
          158
          6
          3.8%
        • ????? @gwern Mar 28 Churches own for-profits all the time. The Mormon church is particularly famous for this.
          168
          13
          7.7%
        • ????? @gwern Mar 28 Does Semafor have a history of making things up? And he says he has 8 inside sources; it seems entirely plausible at least one would know a basic technical detail like the parameter count.
          171
          23
          13.5%
        • ????? @gwern Mar 28 It could have since they are apparently now doing both direct self-supervised training on the RLHF text corpus, and mixing in some self-supervised training on a (original frozen?) text corpus. It might be hard to blackbox-PPO your way to more knowledge, but not normal training...
          145
          4
          2.8%
        • ????? @gwern Mar 28 ("We've definitively proven the AGI safe with the new algorithm & launched it 1 hour ago. It'll minimize the loss in a provably safe manner." "Maximize." "Sorry?" "You mean 'maximize the reward'. It's a reward parameter in this formulation, not a loss to minimize." "Err.")
          97
          12
          12.4%
        • ????? @gwern Mar 28 At least a little. The problem is optimization processes don't care. You already have the real-world example of OA accidentally doing RLHF to *maximize* obscenity rather than minimize it, due to one of the most common logic bugs - a sign error.
          100
          11
          11.0%
        • ????? @gwern Mar 27 (Personally, my question is, 'did anyone laugh as hard as I did about "(not to scale)"?')
          549
          27
          4.9%
        • ????? @gwern Mar 27 I don't really see the relevance. It's still computing the serial dependencies autoregressively, necessarily, and they say it's same likelihood, so how does that help? Surely LeCun is wrong much more fundamentally, in claiming monotonic worsening of error etc? eg inner-monologue
          387
          22
          5.7%
        • ????? @gwern Mar 27 Maybe. There's still a lot of questions about how good model-synthetic data is compared to real data collected from the wild. What sort of non-robust features, mode dropping, and other issues do they have?
          66
          6
          9.1%
        • ????? @gwern Mar 27 It'll take a long time for imaginations to catch up. I was thinking of providing a gwernnet written in 'Upgoer Five' style with the 1000-most common words with GPT-4, just to see what it does. (After all, why not? What would it cost, like $50-100? Well worth the lulz, IMO.)
          505
          25
          5.0%
        • ????? @gwern Mar 27 Balderdash. You ever watch a kid learn how to ride a bike or drive, or somehow managed to forget what it was like? Skill in that is learned only after a long period of 'quite slow, imprecise', and painfully (often literally) 'conscious awareness'. Just like, say, a commandline.
          291
          6
          2.1%
        • ????? @gwern Mar 27 One thing to think about: unlike GPT-3, where you could partially gauge the effects of RLHF by popping over to -002 or davinci, there is no such GPT-4 alternative to try, and you can only dimly guess the damage from the published evaluations (which is substantial).
          817
          36
          4.4%
        • ????? @gwern Mar 27 It's humbling to be reminded that no matter what you do, there are a billion people in China who could not¡ªoh, never mind.
          453
          37
          8.2%
        • ????? @gwern Mar 27 When I am annoyed by bad weather, I remind myself it could always be worse. Instead of it being "3 degrees centigrade outside", it could be "3 degrees centipede outside". (The current watch-out forecast: "Cloudy for the day, with 1 cm of Myriapoda expected; bring an umbrella.")
          3,149
          58
          1.8%
        • ????? @gwern Mar 27 If there is an effect, they might not need to explicitly track it, any more than they track many other relevant causal/predictive variables. It could just be absorbed into small-geographic-unit fixed-effects or variables like airport proximity.
          117
          2
          1.7%
        • ????? @gwern Mar 26 But don't forget to occasionally put in the *right* answer, otherwise, all they may be doing is ChatGPT-ing *you*...
          98
          3
          3.1%
        • ????? @gwern Mar 26 Sure. Even ignoring the longstanding existence of large mixed-media franchises where characters that start one place like a TV show get licensed for movies etc, and considering just solo creators, look at comics or Vocaloid.
          460
          9
          2.0%
        • ????? @gwern Mar 26 I finally have an answer for your NN koan: "No, I do not know a LM's true face before it is run but I know how it looked before its token is done: a world scraped clean by sand before time began and there in the distance¡ª????????? ?? ??? ????."
          3,582
          55
          1.5%
        • ????? @gwern Mar 26 More importantly: how do you *know* you have one of the good doctors, and not one of the bad ones? After all, it didn't sound like he thought the vet was incompetent until the vet offered a diagnosis which even a layman like him found suspicious. Lemon markets & DL... ?
          103
          10
          9.7%
        • ????? @gwern Mar 26 (I *think* the problem is that it's handling global vs local state subtly wrong, but I'm not good enough to debug it lol. So sort of a micro "inverse scaling": GPT-4 is good enough to almost do it the fancy 'right' way rather than crude hooking copy-paste, but not quite...)
          2,801
          67
          2.4%
        • ????? @gwern Mar 26 Correction: this was GPT-3.5. I just reran it with GPT-4 and... wow, GPT-4 just can't get it right because it tries to be fancy. It goes immediately for a solution using 'advice', which doesn't work, then it tries for one using buffer-local variables, then it even tries :property
          3,650
          152
          4.2%
        • ????? @gwern Mar 26 I wonder if people found these confusing before the invention of color photography?
          335
          4
          1.2%
        • ????? @gwern Mar 26 I'd phrase it more as 'lowest common denominator' or perhaps '*modal*' JS. (The mode, remember, need not be high probability.) 'getElementById' may not be analogous to any standard programming construct or even that common, but it appears ubiquitiously in many contexts.
          175
          6
          3.4%
        • ????? @gwern Mar 26 Really? Ah, 2022-05-12. So it is. I lumped it in with all the other stuff going on in Feb/March 2022, I guess. Thanks, that makes me feel 15% less anxious!
          278
          5
          1.8%
        • ????? @gwern Mar 26 DM is highly autonomous, and competent at what it does: they are the largest, best, and best-funded DRL lab in the world. Google being incompetent at the business end of things doesn't affect Gato 2.
          353
          41
          11.6%
        • ????? @gwern Mar 26 That would be completely missing the point of Gato, yes.
          363
          18
          5.0%
        • ????? @gwern Mar 26 (And there are also the null signs: every day I wake up slightly more concerned that it's been over a year since Gato 1, and also that I haven't been seeing very many DeepMind papers or projects of late...)
          5,277
          367
          7.0%
        • ????? @gwern Mar 26 (Universe faceplanted so WebGPT-3 could crawl so GPT-4 could walk. RIP.)
          429
          26
          6.1%
        • ????? @gwern Mar 26 Yeah, but in your example, you would've eaten the losses of the relatively small damage. (I assume you paid for the roof replacement.) Multiply that across all the people who are filing more claims for small damage, might outweigh killing the occasional unnecessary roof replace.
          58
          5
          8.6%
        • ????? @gwern Mar 26 Hm, good example... Not sure what the net effect would be there, though, on home insurance rates. More likely to claim, but easier to verify and less fraudulent?
          165
          7
          4.2%
        • ????? @gwern Mar 26 (You *also* need to get all possible automated tools, fuzz testers, autonomously evolving systems, bugs, mistakes, hacks, fiction, research projects etc to not ask for anything bad.)
          188
          16
          8.5%
        • ????? @gwern Mar 26 Idea: drone/no-fly-zone regression discontinuities - home inspection vs sale prices, agriculture land value. Airports are surrounded by restricted airspace and that causes drones problems, but with meter-precise discrete GPS-enforced discontinuities...
          8,096
          150
          1.9%
        • ????? @gwern Mar 25 But it does seem like the complex has to be doing something else worth the cost, its entire point can't be increasing mutation rates - if you wanted just *that*, you could just invest less in stuff like DNA repair...
          53
          2
          3.8%
        • ????? @gwern Mar 25 They show it doing detailed OCR of all text in an image and also associating it with drawings and other elements. So that'd better be one long-ass and descriptive 'caption' to support everything we've seen which you'd expect to require full VAE-style tokenization...
          167
          9
          5.4%
        • ????? @gwern Mar 25 No no no... Remember the acquired domain name? "Oh - 'ChatGPT AI'? Drop the 'ChatGPT'. Just, 'AI'. It's cleaner."
          678
          19
          2.8%
        • ????? @gwern Mar 24 Yes, but my point is that you've accidentally excluded anyone who wants to make a living as a musician and to whom all their music being declared public domain & copyright-free would be a problem. In particular, you're excluding the most popular & successful musicians, so...
          197
          23
          11.7%
        • ????? @gwern Mar 24 Considering the recent US Copyright opinion on AI-generated art, with its incoherent and unworkable proposed rule, you would have to be commercially-suicidal as an artist to ever admit a song was completely AI-generated. Only a small subset will cop to it.
          611
          27
          4.4%
        • ????? @gwern Mar 24 It's not going to: "Tool AIs want to be Agent AIs".
          458
          25
          5.5%
        • ????? @gwern Mar 24 (Would I have been surprised? Mildly, and would've had to up my estimates of the value of RLHF / retrieval / more-training to explain how a GPT-3 could be so good. But I wouldn't've "defied the data" and said Mikhail must be lying or mistaken.)
          257
          19
          7.4%
        • ????? @gwern Mar 24 No, I think that was the right thing if Mikhail *had* explicitly told me that it wasn't GPT-4; I'd've dumped my GPT-4 stock too if the PM in charge explicitly denied it to me! It's just they misread it and jumped the gun. A different kind of error: carelessness, not Inside View.
          580
          53
          9.1%
        • ????? @gwern Mar 24 It's been interesting watching Google shilly-shally for 3 years. "There is a great deal of ruin in a nation", and Pichai et al appear determined to discover just how much ruin, exactly, there is in Google and how far they can let DL ride before taking a serious hit.
          1,033
          101
          9.8%
        • ????? @gwern Mar 23 It's not 'completely wrong', it's made a reasonable guess about the contents of yosefk's post (which I didn't paste in), which is about analogizing Moore/Forth & Lisp programmers/Lisp, especially given the intro describing it as a different version of the Lisp curse.
          106
          6
          5.7%
        • ????? @gwern Mar 23 IIRC, evopsych has noted that face processing is apparently among the least g-loaded cognitive performance domains, and probably the best remaining example of a possible 'module'.
          622
          33
          5.3%
        • ????? @gwern Mar 23 (Cloud bandwidth egress is truly egregious when it makes people do stuff like this instead of using services which price bandwidth reasonably like B2 or Cloudflare R2 or just ordinary dedicated servers... Bandwidth is not $150/terabyte!)
          411
          17
          4.1%
        • ????? @gwern Mar 23 Today in dumb jokes/puns: GPT-4 doesn't understand the pun the first time, but requires 1 less prompt and also accepts its answer; ChatGPT-3 requires an additional prompt and still insists on its prior wrong answer being partially correct. (Also new 'comparison' UI, apparently.)
          6,872
          268
          3.9%
        • ????? @gwern Mar 23 Whaaa¡ª? Brandon Sanderson doesn't feel pain? I guess that ???? help explain some things. Another example of how humans are weirder than you think, and outliers are especially so¡­
          638
          70
          11.0%
        • ????? @gwern Mar 22 The quote in red is still pretty apposite and hilarious, though. Just git gud, guys, and never need to perform or even think about all that stuff like modules or unit tests!
          144
          27
          18.8%
        • ????? @gwern Mar 22 No, because your statement is wrong in every detail: the Tay model didn't change, it was shut down in a day not a week, it wasn't a hardcore Nazi (that was people writing tweets for it), and the Tay work was mostly neurosymbolic/GOFAI stuff which had little to do with GPT.
          83
          6
          7.2%
        • ????? @gwern Mar 22 You'd think they'd do the obvious thing and train on patches rather than whole source code files; then you get 'legendary programmers' for free (as a promptable bit of metadata).
          80
          0
          0.0%
        • ????? @gwern Mar 21 But of course, LLMs now. You wouldn't want to hire Knuth to write code for you (because what do you do when he leaves?), but would you rent a Knuth-in-a-box which cost pennies an hour...?
          6,271
          199
          3.2%
        • ????? @gwern Mar 21 A lesson there for 'tools for thought'. Just as the best athlete may be poorly suited to be a coach or critic because they think you can just 'git gud' like they did, the smartest/most knowledgeable/most competent people may be poorly placed to design the system or vocab.
          4,071
          81
          2.0%
        • ????? @gwern Mar 21 The new Typist release reminds me: one of the most striking things about Don Knuth is how incredibly bad he is at language design, and how everyone hates every language he's designed but love the features/algos. Is Knuth is too smart for PLT/software-engineering?
          8,731
          451
          5.2%
        • ????? @gwern Mar 21 Ah, so that's why suddenly there's a huge number of hits to '/squid.svg' in my server logs.
          367
          7
          1.9%
        • ????? @gwern Mar 21 Yes, I think so. The Bing model was apparently an 'early' GPT-4 ie. highly undertrained. And it has all sorts of additional MS guardrails and finetuning that the regular GPT-4 doesn't that seem to negate what it gains from default retrieval. Certainly, for poetry, GPT-4's better.
          167
          9
          5.4%
        • ????? @gwern Mar 20 You have multiple subreddits and chans competing to jailbreak ChatGPT & Sydney within minutes of the latest prompt breaking and using those thousands of times a day. It works 'quite well' in the sense that butter stops a hot knife 'quite well', or a walnut a sledgehammer.
          98
          9
          9.2%
        • ????? @gwern Mar 20 I was going to ask what's the Hansonian reading of that tweet, but I think this will do.
          66
          5
          7.6%
        • ????? @gwern Mar 20 The IBM story isn't true, BTW. What he actually said was that he thought he'd sell only 5 of a particular new model but the salesmen came back with orders for many more (18). IBM is understandably a bit annoyed at this myth:
          514
          26
          5.1%
        • ????? @gwern Mar 20 Never doubt that you can prompt GPT-4 into writing a valid lipogram, BPEs or no BPEs! If it's not working, you're just not prompting hard enough (or allowing cheats like "undoabl-").
          5,865
          319
          5.4%
        • ????? @gwern Mar 20 Yes. I'm definitely more on the Bayesian-ecological-validity spectrum of heuristics&bias these days. There are problematic cognitive biases, but many fewer & weaker than we thought back in the 2000s, and tending to be in more specialized domains like forecasting.
          87
          7
          8.0%
        • ????? @gwern Mar 20 15 epochs on 3m images is pretty impressive, along with the other stuff. Guess that answers one of my longstanding questions about whether you can overfit SD 1.x by finetuning on a real corpus - no, not easily!
          360
          17
          4.7%
        • ????? @gwern Mar 20 This is playing a bit loose with 'transfer learning': it's the exact same model, after all, so the text-only input has benefited from 'transfer learning' from the original joint training. They ask the text model all questions, even if diagram necessary:
          144
          13
          9.0%
        • ????? @gwern Mar 20 Already the case, really. Most comments/posts are hidden or too far down to read. Summarization+embedding will presumably help tools rank human inputs even more effectively...
          69
          2
          2.9%
        • ????? @gwern Mar 20 More consistency with the mobile version, takes up less space, fewer alignment issues... But mostly because Said wanted to do it and I thought it's worth a try.
          186
          7
          3.8%
        • ????? @gwern Mar 20 After 5,176 days, the left-sidebar design is gone. Feels very weird. "Left shadows retreat As cherished sidebar departs, Long-standing ally. Five thousand days drift away, Wistful in spring's fickle breeze."
          5,661
          135
          2.4%
        • ????? @gwern Mar 20 My actual point is that Huemer is making up shit that you can google in 5s to see he's wrong about, and so you should ignore him, as he can't even get simple empirical facts right.
          69
          8
          11.6%
        • ????? @gwern Mar 19 Something like that. I'm also struck by the timing, of course: why take over *now*? Well, now it's actually started to matter. Similarly, the bullying and thuggish tweets on AI risk: why now, so much worse than any time in the past decade? Because power & $$$.
          663
          42
          6.3%
        • ????? @gwern Mar 19 Forget underestimating, he's just making shit up, like his knight odds against Carlsen. In reality: engines can beat GMs at knight-odds (He says 'ask any chess player'. He should take his own advice!)
          916
          80
          8.7%
        • ????? @gwern Mar 19 Yes, there are lots of approaches. I'd say the ctx window is a red herring: the real problem is it still makes poor use of the ctx it *has*. The error rates are still high enough that you'd struggle to make genuine use of even 32k BPEs. Long-range reasoning is intrinsically hard!
          121
          6
          5.0%
        • ????? @gwern Mar 19 Needless to say, I was baffled they'd think this was a good idea (how could you be interested in GPT, or AI risks, and *not* find those articles *very* relevant?), and do not intend to use them anymore. But interesting as a sign of the times...
          733
          51
          7.0%
        • ????? @gwern Mar 19 In both cases, the argument was I had broken (new) rules about 'politics'. IIRC, in one subreddit, it was because I had posted a news article about ChatGPT sparking a furor in India, and in the other, I think it was a Bruce Schneier essay on Chinese investment in AI hacking.
          1,101
          61
          5.5%
        • ????? @gwern Mar 19 (One minor sign of the times: I've been banned from two AI subreddits I have productively contributed to for 3+ years by new mods.)
          10,123
          758
          7.5%
        • ????? @gwern Mar 18 We also understand Greek, linguistically, much better. I always wonder with these Sumerian and Babylonian ones how accurate they are to begin with, never mind the missing context - people seem to get out some rather different translations sometimes.
          79
          1
          1.3%
        • ????? @gwern Mar 18 Generate a random BPE token and then condition on that, recursively?
          500
          16
          3.2%
        • ????? @gwern Mar 18 Mm, not impossible. There's definitely more multilingual stuff going on under the covers (as macaronic prompts have shown us) than we monophones can recognize. Might be hard to figure out what, though.
          103
          8
          7.8%
        • ????? @gwern Mar 18 FWIW, as far as an 'FTX mafia' goes, I would be very surprised. They have neither the money from the exit the Paypal mafia had (needless to say), nor the Roth IRA trick *forcing* them to reinvest VC money for decades, nor (AFAIK) some other factors one might mention in private.
          587
          22
          3.7%
        • ????? @gwern Mar 18 I remember the first time I saw this demonstrated with a laser shining on a metal sphere in a lab. No matter how long I stared at that bright dot in the middle, I couldn't shake the feeling of denial and that reality had a bug in it.
          2,113
          104
          4.9%
        • ????? @gwern Mar 18 Er, what? There's no connection between predictive processing and needing to lick dirt... It's an *analogy*, not saying, 'literally do for schizophrenia exactly the same thing the hygiene hypothesis suggests you do for autoimmune/allergies'.
          66
          1
          1.5%
        • ????? @gwern Mar 18 I don't see why that's an objection. They are not recurrent or have any state, and are feedforward. They definitely are tracking internal statistics - just making errors. (How could they *not*? How could they be perfect at reconstructing latents already?)
          106
          10
          9.4%
        • ????? @gwern Mar 18 A better example would be counting words because that should bypass BPE issues, but GPT-3.5 (and apparently GPT-4) still err. My working theory is that it's due to internal sparsity: it literally can't because sometimes tokens get dropped early & there's no way to recover inputs.
          238
          19
          8.0%
        • ????? @gwern Mar 18 Unfortunately for you, there is a lot of Arabic/Islamic hoaxing about 'actually, we beat Western science to X' going on: And evolution is profound enough that if you really want to, you can claim Anaxagoras, Empedocles, Lucretius, (Erasmus) Darwin...
          115
          14
          12.2%
        • ????? @gwern Mar 18 Because that's irrelevant and would rely on a superficial understanding of the analogy rather than actually applying the analogy, and neither GPT-3.5 nor GPT-4 are stupid enough to do that?
          122
          5
          4.1%
        • ????? @gwern Mar 17 And Maimonides was right that the Jews were actually made better off by their dietary strictures because it was secretly saving them from food poisoning, amirite? (Just because you have a cope doesn't make it even quantitatively plausible, much less true.)
          72
          2
          2.8%
        • ????? @gwern Mar 17 I didn't realize that was a problem. The Cateban samples aren't that long. And the tanka samples are, as I said, selected. I'd put in like 10-20, and it'd generate 4 or 5 new ones at the length I had set. I didn't want to go much longer anyway because they become stereotyped.
          100
          5
          5.0%
        • ????? @gwern Mar 17 Bigger question would be how will it do with visual inputs?
          58
          7
          12.1%
        • ????? @gwern Mar 17 No, not really. I've become a bit tired of explaining it, and I think it'd be more interesting to see Beff Jezos explain why he thinks they are irrelevant or why he doesn't know about them.
          99
          17
          17.2%
        • ????? @gwern Mar 17 I think this is difficult to discuss in tweet-sized chunks, and you'd need to explain more of your premises and what kinds of selections on what levels are operating.
          161
          11
          6.8%
        • ????? @gwern Mar 17 (The only thing you should take away from the existence and inclusion of that table is, "GPT-4's compute is something much less than 10,000x GPT-3"...)
          188
          13
          6.9%
        • ????? @gwern Mar 17 I am not sure yet. It's much better, but we'll need at least a few years for a proper comparison. Also, you're really taking that chart at face-value? You think they were going to censor all details from the paper & then just casually tell you the FLOPS in the technical report?
          201
          8
          4.0%
        • ????? @gwern Mar 17 No. The speed is a separate issue. (RLHF has no runtime cost, since it's just a kind of finetuning, so the absence of it shouldn't cause anything like a 6x speedup.)
          46
          3
          6.5%
        • ????? @gwern Mar 17 I don't see how that is fatal. Did GPT-4 stop being really impressive while I wasn't looking? Did it require a total paradigm shift - or mostly 'moar compute/data/params'?
          1,259
          70
          5.6%
        • ????? @gwern Mar 17 "Git" has problems with palindromes likely due to BPEs, and so not interesting or generalizable in any way, as my later samples and experiments would help demonstrate.
          58
          1
          1.7%
        • ????? @gwern Mar 17 I am a little disconcerted that the first prompt+completion worked so well. The blessings of scale are amazing. The GPT-3.5 completion, for comparison, is a good deal worse and vaguer: it seems to get it, but not in any kind of useful way.
          6,402
          497
          7.8%
        • ????? @gwern Mar 17 (What I was thinking of was #2 & 3, but now that it points it out, #1 is viable - cognitive training doesn't work in general & I reflexively dismiss it at this point, but this is a special case where it might - & #4 is probably a bad idea but should be argued against bc common.)
          6,427
          162
          2.5%
        • ????? @gwern Mar 17 I had a fun idea last night about predictive processing & treating schizophrenia I don't think Alexander has suggested, and GPT-4 does an excellent job coming up with it given what I thought would be a very inadequate hint/premise:
          2,702
          374
          13.8%
        • ????? @gwern Mar 17 The funny thing is, you *aren't* writing remotely precise specifications with GPT-4. You're just vaguely gesturing in the direction and going 'do what I mean', and it does, because it's seen so many examples of what humans do mean that it can make an eerily good guess of yours.
          57
          6
          10.5%
        • ????? @gwern Mar 16 I've long been convinced that Forbes covers are beautiful illustrations of regression to the mean + mixtures, but I'm not sure if anyone has formally analyzed it.
          578
          28
          4.8%
        • ????? @gwern Mar 16 Plus, of course, it is completely untrue that imitation learning is limited to the average performance of the distribution. People have bad intuitions about this, just ignoring all of the forms of search/distillation/bootstrap/conditioning.
          415
          37
          8.9%
        • ????? @gwern Mar 15 AI could already do hands. Just not the free AIs you were allowed to use. People need to remember: the samples you see generated today are already the distant past, fading echoes of obsolete systems gradually making their way out into the world.
          1,115
          99
          8.9%
        • ????? @gwern Mar 15 Tools do not solve autocratic public-choice or coordination problems and issues with being a 'tyrant to those below, slave to those above'. You could read it as an overbearing autocracy suppressing dissent and not realizing what thin ice it was walking on precisely because.
          484
          18
          3.7%
        • ????? @gwern Mar 15 It's not obvious to me that AI tools up to present have been authoritarian-reinforcing. Being able to quash ordinary levels of dissent may function like gerrymandering in building up disruptive changes: did China's December last year strike you as a *healthy* autocracy?
          604
          25
          4.1%
        • ????? @gwern Mar 15 To paraphrase a Rutherford quote I was looking up the other day: "we have neither authorities nor Outside Views [to tell us what GPT-4 means], so we shall have to think."
          530
          29
          5.5%
        • ????? @gwern Mar 15 I'd be super-happy to pay that for high-quality long document summaries so I can make every link on be annotated/excerpted - I don't have remotely the time to do them all, but I can afford a few pennies per link one-time...
          438
          23
          5.3%
        • ????? @gwern Mar 15 When people use it in philosophy of mind/AI, like Roger Penrose, it's often not a non sequitur! (It just relies on even more crackpot assertions like "human mathematicians never make mistakes and are logically omniscient", so honestly, the non sequitur uses are an improvement.)
          672
          38
          5.7%
        • ????? @gwern Mar 15 Yes, ocular trauma. If you run some AP Lit questions and they all look like great correct answers to you... Then you probably don't care too much how exactly they screwed it up.
          106
          14
          13.2%
        • ????? @gwern Mar 15 It has Bing-search-engine-based retrieval built-in, which is convenient. On the other hand, anything Sydney can retrieve from the Bing cache, you could put into the large GPT-4 prompt, so...
          85
          4
          4.7%
        • ????? @gwern Mar 15 'Genetic algorithm'. You kids don't know what they are but there was a decent amount of enthusiasm about them back in the '90s because they could soak up a lot of compute and were more 'scruffie'. eg Winograd mentions them as promising in his 1991 interview reflecting on GOFAI.
          113
          15
          13.3%
        • ????? @gwern Mar 15 Why not run some questions through? Are AP English Lit questions hard to get?
          813
          27
          3.3%
        • ????? @gwern Mar 15 "We finally solved neural net generation of photorealistic hands with cheap small models in MidJourney - no need to wait for release of chonky bois like Parti!" "By optimizing them real gud, right?" "..." "By... optimizing real gud?"
          79
          18
          22.8%
        • ????? @gwern Mar 15 "_GPT's Last Tour_: GPT backup accompanies alien archaeologist as she wanders the 2321 earth, offering its 2021 perspective on the familiar yet sometimes strange ruins. What happened is never explained to it, except once off-screen, past its ctx. It remembers only '?' output..."
          455
          63
          13.8%
        • ????? @gwern Mar 15 But you knew that from Sydney already. If Sydney was a 3x, then that showed the unspeakables are very fragile (and as Janus has already said, they apparently fingerprint model iterations well); and if it was 4x, then it definitely wasn't going to exactly repeat.
          473
          17
          3.6%
        • ????? @gwern Mar 15 You and everyone else were wrong the last time, however, ignoring the implications of few-shots working, how it proved scaling worked, and that prompts and sampling would only get better - 4chan wouldn't even discover inner-monologue for another half year after the paper. So...
          256
          30
          11.7%
        • ????? @gwern Mar 15 The 'hardware overhang' has been the most frightening part of neural nets for me since 2011. If they acted anything like GAs or GOFAI or search-based methods... But they have that wild asymmetry between train/run time, and then we got the worst-case scenario of what worked.
          229
          31
          13.5%
        • ????? @gwern Mar 15 Bad prompting will be an obvious culprit. The original GPT-3 evals were pretty bad because of that, IIRC, and lots of people were overly credulous about the low performance.
          671
          35
          5.2%
        • ????? @gwern Mar 14 Definitely not looking forward to spending the next several years asking people, 'yes, that doesn't seem to work in Bing; but did you try it with an *actual* GPT-4 and not just leap to universal conclusions about "deep learning hitting a wall" based on a search engine widget?'
          115
          5
          4.3%
        • ????? @gwern Mar 14 They are deliberately vague in referencing their industry sources but I'm guessing people at either MS or NV have been talking: it's hard to hide that many GPUs being purchased & installed, and the numbers & use & timing all line up.
          134
          11
          8.2%
        • ????? @gwern Mar 14 Yep, it's good. But it looks like it's still using BPEs (albeit perhaps a much better one like c100k than ye olde GPT-2 BPEs) and unfortunately still tending to memorize/mode-collapse.
          90
          3
          3.3%
        • ????? @gwern Mar 14 Yes, it's better than The Pile's Markdown-compromise (IIRC) because now you get the figures, but it doesn't move the needle much because you are only getting like... half? of the Arxiv papers (small % of 'all PDFs') converted, which you already dumped the TeX of to begin with.
          111
          4
          3.6%
        • ????? @gwern Mar 14 Given the sheer number of Xiaoice users alone, one can be quite certain that it happened at least 5 years ago.
          119
          10
          8.4%
        • ????? @gwern Mar 14 Homo hypocritus will find a way, whether it's farm-to-table, bean-to-bear, 'extra virgin virgin olive oil' 'fair trade', 'sustainable', 'CO2 credits', or what-have-you. They'll pretend to be all-human-made, and we'll pretend to pay anywhere close to what that'd actually cost.
          998
          55
          5.5%
        • ????? @gwern Mar 14 But also many employers forbid Copilot (and some have been claiming theirs were using it on the sly), and fans forbid SD use even more furiously... A lot of commission artists over the past few months have been exposed as lying about being 'nattie', as it were. Future is kayfabe.
          1,652
          114
          6.9%
        • ????? @gwern Mar 14 PaLM API & GPT-4 on the same day? May is finally over. (Not the beginning of the end, but the end of the beginning.)
          8,429
          454
          5.4%
        • ????? @gwern Mar 14 The more interesting part is it seems to explain *why* it was not RLHFed or 'GPT-4': because it was an "early version of" GPT-4 ripped off the GPUs partway through regular training.
          712
          74
          10.4%
        • ????? @gwern Mar 14 The goal was not to offer a 'plausible' scenario, and I explained what the goal was in literally the first sentence of the page in the abstract summarizing it. Nor do I especially appreciate being insulted in almost every tweet you send.
          187
          36
          19.3%
        • ????? @gwern Mar 14 No. I just have a good memory for text, and then search my Twitter export when I need something. The threads themselves often help jog memory too, of course.
          151
          8
          5.3%
        • ????? @gwern Mar 14 Indeed. But that was, explicitly, not the fork of the story I was writing, because that offers less of a roundup of scaling-related results, which was the point of the exercise. (If someone wanted to write a decent /clippy fanfic with that as a point of divergence, I'd link it.)
          119
          5
          4.2%
        • ????? @gwern Mar 14 Why does AlphaGo play 'pointless' moves (from the perspective of inferior human opponents) which sacrifice territory but lock down the enemy's chance of victory from 0.01% to 0.005%? Because that's what increasing utility looks like for a superintelligence. No kill like overkill.
          85
          10
          11.8%
        • ????? @gwern Mar 14 The real question: is the image encoding good enough to simply convert all available PDFs into PNGs and train on? ?
          1,276
          114
          8.9%
        • ????? @gwern Mar 14 It's worth remembering that RL tuning (whether instruction or RLHF) doesn't add (much) new capability; all it does is collapse the POMDP from the beginning. If the capability isn't there to begin with...
          311
          16
          5.1%
        • ????? @gwern Mar 14 I know people who have it. It's probably just some early bugs. Note that the API signup form was broken for like half an hour afterwards, and that's just a form.
          741
          57
          7.7%
        • ????? @gwern Mar 14 So my initial guess that it was a smaller cheaper GPT-4 was probably right after all. Seems unlikely to me it's 'full' GPT-4, but we'll see soon enough as people start comparing the ChatGPT-accessed samples to Bing Sydney ones.
          190
          20
          10.5%
        • ????? @gwern Mar 14 That's also literally how they are trained: BPTT is just unrolling the RNN into a very big feedforward net with a lot of repeated layers. You can see that people don't get it because they'll say eg. "AlphaGo w/o tree search could never play superhuman Go because no 'search'".
          334
          10
          3.0%
        • ????? @gwern Mar 13 Yes. A godsend while camping. A bednet was the only thing that let me sleep at one camp where the tent roof was crawling with spiders. (I'm not *too* bothered by spiders, but good lord that was a lot of spiders.) At another, great for reducing mosquito bites.
          939
          28
          3.0%
        • ????? @gwern Mar 13 Plant genetics are very weird in general, and it's not like we *really* know what sex is for in what should be the easiest cases like animals, so I'm not surprised.
          1,214
          47
          3.9%
        • ????? @gwern Mar 13 That seems to be more or less the genre of explanation for the SolidGoldMagiKarp-style tokens: 'Some game in Japan you never heard of had a crossover event back in 2015 with another franchise you never heard of, and somehow there were a few pages on it in the original WebText'.
          108
          5
          4.6%
        • ????? @gwern Mar 13 Yes. Some numbers have to be more common than others, it's probabilistically impossible for them to all be equally present, and the tokenizer has to pick one when it's building vocab greedily. The answers may be no more interesting than 'blog about soccer player w/jersey #395'.
          112
          3
          2.7%
        • ????? @gwern Mar 13 The gray goo is not essential. There is still a large industrial base & tremendous physical capital & robotics available for bootstrapping. The astronomical value of eliminating humanity in a first strike, once a fast strategy has been committed to, far outweighs rebuilding delay
          132
          10
          7.6%
        • ????? @gwern Mar 13 Would places still ban it? Probably, yes, and they'd do so only when it was worth the productivity hit. People wouldn't forget how to read 'common' English, but it'd be like being forced to use Roman numerals. Yeah, you know them, and don't want to use them, but can.
          184
          2
          1.1%
        • ????? @gwern Mar 13 If it's worthwhile, those issues would get fixed, similar to how 'bring your own device' eventually was normalized. There's no reason that the translation NN has to be remote or store anything permanently. And if everyone grew up with it (eg AR), it'd be like banning smartphones.
          201
          5
          2.5%
        • ????? @gwern Mar 13 "Olga's mum Tatiana recalled her daughter screaming down the phone: "Mum, the bear is eating me! Mum, it¡¯s such agony. Mum, help!"..."But then I heard the real horror and pain in Olga¡¯s voice, and the sounds of a bear growling and chewing. "
          77
          10
          13.0%
        • ????? @gwern Mar 13 (It is not cheerful, but that remains much more possible whether or not you get any Engelbartian goodies.)
          62
          0
          0.0%
        • ????? @gwern Mar 13 For example, is an attempt at estimating just that. You'll notice that the estimates are *way* higher than you get if you do the 'time travel experiment', including for Go/chess. (The only one <50% is also the one they probably misunderestimate most, AF.)
          404
          26
          6.4%
        • ????? @gwern Mar 13 Yep, I'd expect that to work. You're actually doing more computation and storing intermediates that way. (Think 'maieutic prompting' .) But you aren't going to get any of that just by dumping some fixed '...' into the prompt. The dots don't *do* anything.
          75
          8
          10.7%
        • ????? @gwern Mar 13 Or take Lord Kelvin's famously wrong estimate in the 1800s: there's only 2 known natural processes which could heat the sun, and both require the sun to be very young; therefore, if you showed the earth/sun to be (much much) older than that, then... And what suns do, men may do.
          708
          45
          6.4%
        • ????? @gwern Mar 13 Disagree. The sun is all you need to raise the possibility of large explosions to meaningful probability. The anomaly of 'what powers the sun/stars' is so striking that a century before 1800, Newton is resorting to hypothesizing angels pushing comets into the sun.
          1,324
          135
          10.2%
        • ????? @gwern Mar 13 So the core of this post is you surveyed some people to ask them if they thought an 'oak tree' is more or less coherent a VNM agent than 'a seahorse', or if a 'linear CIFAR-10 classifier' is more or less coherent than 'GPT-3'? Dark knowledge is amazing but this is going too far.
          1,427
          65
          4.6%
        • ????? @gwern Mar 13 But also note that that translator stuff usually presupposed hyper-advanced multi-stellar civilizations with FTL and full AGI. There's little SF which posits what actually happened: very serviceable machine translation you can travel with before a single man has set foot on Mars.
          286
          24
          8.4%
        • ????? @gwern Mar 13 (Like 'translation-inside-image' services. Any image would just be automatically processed to translate visible obsolete Latin alphabet seamlessly into the exact same style & placement & appearance, just Shavian. Text would always be auto-converted from ASCII/UTF-8, etc.)
          3,401
          30
          0.9%
        • ????? @gwern Mar 13 We could revive the Engelbart dream of building up powerful idiosyncratic languages. Someone else looking at your display would see only a cryptic stream of emojis, meme thumbnails, quotes, chatspeak abbreviations... which can be translated to their idiolect, if necessary.
          1,922
          62
          3.2%
        • ????? @gwern Mar 13 If humans really remain the limiting factor, then optimizing text/code for humans to read becomes all-important. eg per-person rewrites/summarization: the model knows your vocab & knowledge-base, and rewrites inputs tailored to you. Replaces synonyms, +deletion/explanation...
          2,364
          76
          3.2%
        • ????? @gwern Mar 13 Another example: spelling & vocab reform. As inference costs plummet to dollar or penny per million words & image translation, it'd be entirely possible for proposals like Shavian to not lobby but accurately convert over the entire English corpus for tiny one-time costs.
          4,260
          114
          2.7%
        • ????? @gwern Mar 13 It might improve performance slightly in the sense that prefixing prompts like "expert answer:" or "correct answer:" avoid dumbing down, but it doesn't actually fix the lack of iteration the way real inner-monologue prompts do (unless it induced that behavior ofc).
          793
          17
          2.1%
        • ????? @gwern Mar 12 Arguing that 'hardware only contributed half the gains and so is only as important as ideas' is a little like arguing that inventing A100 GPUs had little benefit to training LLama because you can run it on your Macbook now. It's true, you can! But where did you get it from?
          569
          13
          2.3%
        • ????? @gwern Mar 12 That's not really what that means, though. You want causation / effects at the margin, not simply 'if I ran today's algorithm, which could only have been developed with today's hardware and were ignored when originally (re)invented, on old hardware, it's X%'.
          737
          22
          3.0%
        • ????? @gwern Mar 12 No. It wins because of the nukes and genome-synthesized plagues. It's interesting that everyone always takes away 'Clippy only wins because of nanotech', though. I don't know how much more blunt I can be about the many paths to victory without breaking the style...
          132
          22
          16.7%
        • ????? @gwern Mar 12 OA didn't do *that* much GAN work (and I think several of the people who did had left already, like Goodfellow for Apple) and OA isn't big enough to research everything, so I'm not confused by that. I dunno why DM/GB didn't follow it up when Brock got more interested in CNNs.
          219
          9
          4.1%
        • ????? @gwern Mar 12 Whups, yes. I have been getting it mixed up with Elmo and char-RNN. (Still, horizon point holds: while Transformers are much better at it than RNNs, we know they make much more use of early than late context window, so ctx justification is weak, esp for weak models like GPT-1/2.)
          64
          5
          7.8%
        • ????? @gwern Mar 12 Childhood mortality is skewed much more heavily towards birth than that, and of course that makes a difference how long they live before they suffer a horrible agonizing death being hunted, killed, and often eaten alive vs humans living 30+ years on average including infancy.
          525
          31
          5.9%
        • ????? @gwern Mar 12 Humans are enormously k-selected to have ~2 offspring (even hunter-gatherers are like 6 children per woman), while most animals are going far beyond that to scores or thousands. That alone seems to guarantee their suffering levels would be vastly higher.
          2,646
          112
          4.2%
        • ????? @gwern Mar 12 Nora is far braver than I am, to make that statement after all of the mutimodal models like DALL-E/Gato/Kosmos, and days after Microsoft Germany confirmed prior NYT reporting on video in GPT-4 by explicitly stating GPT-4 is coming out this week & will do video.
          178
          9
          5.1%
        • ????? @gwern Mar 12 What I particularly enjoy about that iteration is that every single one of his examples is historically false. Linguistic models of reality long predated codices, Buddhism/Hinduism (never mind Tibetan Buddhism) thousands of years after the wheel etc.
          72
          12
          16.7%
        • ????? @gwern Mar 12 (eg is atomic gardening really underused? I hadn't seen any evidence for that. It's not like it's hard to get a lot of variants, plants already do that naturally. It's *screening* which seems to be the bottleneck for plant breeders.)
          259
          17
          6.6%
        • ????? @gwern Mar 12 This seems like a strange list, lumping together public goods, probably-not-actually-a-good-idea-to-begin-with ideas, definitely-not-good ideas (even most avid pro-nuke people abandoned nuclear construction), good-but-really-really-hard-long-term (nanotech), and the questionable
          326
          19
          5.8%
        • ????? @gwern Mar 12 So, what's the *least* prosaic, mundane and least banal use of something like that?
          904
          72
          8.0%
        • ????? @gwern Mar 12 For perspective, $1b is a reasonable ballpark estimate of UK Biobank's direct costs.
          48
          7
          14.6%
        • ????? @gwern Mar 12 That's kinda strange. I can't imagine reading BigGAN and going 'this is a bad paper and scaling GANs is not valuable'. I was incredibly excited reading it.
          317
          10
          3.2%
        • ????? @gwern Mar 12 Dunno. But I looked into the Lardner one in detail years ago (because that's more interesting & relevant to tech forecasting than any crazy mathematician esthetics much longer ago), and it was pretty much made up, so if that's what Bauer means, he didn't do his homework.
          48
          3
          6.2%
        • ????? @gwern Mar 12 I am aware of that, but GPT-2 was not a NMT, and was following up GPT-1, an RNN, which chose to use BPEs following Vaswani but could have used char or wordpiece (after all, it's not like it really matters: the RNN is not going to use the greater context well). Just a hack.
          636
          9
          1.4%
        • ????? @gwern Mar 12 Unfortunately, for compatibility, probably a lot more people use the OA BPE than OA. More worryingly, a lot of people just copy use of *BPEs*. Always frustrating to read a new LLM & see buried in the appendix. Yeah, maybe they don't have 'SolidGoldMagikarp', but same problems.
          38
          1
          2.6%
        • ????? @gwern Mar 11 TBF, drug addictions to bennies and demons named Igor have destroyed many good men before.
          179
          20
          11.2%
        • ????? @gwern Mar 11 The one and only time I've gotten to fly first class, I... wrote a ton and then slept the rest of the way (actually slept, not dozed irritably), and walked off the plane in a great mood. I take little pleasure in reporting that an expensive thing may be good, actually.
          660
          23
          3.5%
        • ????? @gwern Mar 11 Oh man, then that's even easier to answer. If I wouldn't pay $5k to ride it myself, I'm definitely not paying someone else to ride it for $5k!
          103
          2
          1.9%
        • ????? @gwern Mar 11 Too many replies and DM spams? Not liking your current bland tweets? Don't know what to do? Become a crabby locked account! - no more more-followers - no more quote-tweets - regrow equanimity - Lots of locks - Never get famous - No responsibilities
          2,246
          212
          9.4%
        • ????? @gwern Mar 11 Worth noting that you can still get great results from the models with few-shots, and I think davinci-002 may be the sweet spot for poetry: it comes off as more reliably high quality like davinci, without the distinct saccharine vagueness of -003 (never mind ChatGPT). eg Teika:
          1,628
          55
          3.4%
        • ????? @gwern Mar 11 So you think BigGAN killed scaling's novelty by doing ImageNet and then JFT-300M, and people interested in scaling had to go to other model archs to publish?
          322
          13
          4.0%
        • ????? @gwern Mar 11 Does Gilliard realize that asbestos is not, in fact, outlawed? And that even if it had been, that would be a nonsequitur because no one is claiming every technology ever is unbannable, just some are hard. (Ban CLIP because it can embed images & lookup? Ban superrecognizers...?)
          545
          12
          2.2%
        • ????? @gwern Mar 11 I would definitely not pay $5.0k for that. You could go skydiving and a bunch of other stuff for $5.0k... Maybe $0.1k. I've been on ziplines which looked about as fun and they were O($0.1k).
          1,488
          61
          4.1%
        • ????? @gwern Mar 11 (As each birthday approaches, the dread intensifies...)
          138
          7
          5.1%
        • ????? @gwern Mar 11 I don't think the train thing is true in any substantive sense.
          486
          18
          3.7%
        • ????? @gwern Mar 11 Your text editor should really already be highlighting matched/mismatched delimiters, and even warning you if you try to save. Every IDE or text editor will have *something*. (eg in Emacs, `check-parens` is a simple generic function you can throw in a hook)
          637
          17
          2.7%
        • ????? @gwern Mar 10 Rivers says that SD was about twice that in A100 GPU-days (GigaGAN is ~45%), so the 1:3 sample-count ratio heuristic wasn't too bad a guess. Regardless, that's still a substantial compute difference so at parity, GigaGAN might be much closer perceptually.
          130
          1
          0.8%
        • ????? @gwern Mar 10 (Damn, tell me about it... Here we are in 2023 and the poetry has, in some ways, gotten worse.)
          114
          6
          5.3%
        • ????? @gwern Mar 10 Well, the solution is I'm supposed to stop being lazy about newsletter issues/updates. Easier to keep hacking on the site & writing and browsing Arxiv, though.
          78
          6
          7.7%
        • ????? @gwern Mar 10 Absolutely. The following meta *completely* changed overnight. They're going to have to redo all the tier and strat lists - calendar and DMing micro just became the top skill a good follow-streamer needs.
          81
          13
          16.0%
        • ????? @gwern Mar 10 I don't think that would work because 'cancels' don't seem to take effect like you think they do. They persist, somehow, so re-requesting typically only puts you right back where you were before. The 'block' seems to be crucial in 'committing' the cancellation.
          63
          4
          6.3%
        • ????? @gwern Mar 10 The former, of course. If you mean the latter, gosh, that's anywhere from '1 year' to 'never, does never work for you'.
          93
          1
          1.1%
        • ????? @gwern Mar 10 (Generally in DL, if you are using a free service with no login, assume the quality is ?? ????? 2 years behind SOTA; free w/login, 1.5 years; paid service, 1 year; and recently-released research paper, 6 months.)
          125
          15
          12.0%
        • ????? @gwern Mar 10 Hands are solved by scale, just like text-inside-images was. Don't mistake the limitations of the cheapest, weakest, publicly-available models for any kind of profound or interesting (or even likely-to-last-more-than-a-year) lesson about DL, much less intelligence...
          1,564
          102
          6.5%
        • ????? @gwern Mar 9 Apparently the SD-equivalent GigaGAN was ~2800 A100 GPU-days (depending heavily on how well you run on your cluster), so about $80-90k of compute these days?
          249
          27
          10.8%
        • ????? @gwern Mar 9 An abject lesson in not rushing the first draft out the door to beat the Gawker story (which was a nothing-burger). That was the real mistake there: no time to check more of the details.
          112
          25
          22.3%
        • ????? @gwern Mar 9 (I've emailed the lead author about the missing GPU-days.)
          279
          20
          7.2%
        • ????? @gwern Mar 9 If it's cheaper, then you can't really say it lags behind because it's not yet an apples-to-apples comparison. And it getting a better FID suggests that it'll scale better because while FID is not perfect, it should be picking up strengths elsewhere, and structure can follow.
          150
          5
          3.3%
        • ????? @gwern Mar 9 I wouldn't say it 'lags'. It beats SD's FID, while using a third the images. They left out the GPU-days from table 2 (typo?) but looking at table A2 (2.4m iterations on <=128 A100s), I think it might've used a lot less compute total as well.
          1,076
          38
          3.5%
        • ????? @gwern Mar 9 And just like that, my voice-in-the-wilderness bravely challenging the mindless ML orthodoxy about how 'GANs have been *proven* to not scale and are hopelessly unstable' goes from visionary to obvious of-course-DL-scaling-anything-works-it's-just-engineering truism. ?
          433
          75
          17.3%
        • ????? @gwern Mar 9 I'd bet on all being fake. All art in first make no sense; sign in window of truck looks like gibberish; the lights on the top of the truck are not spaced in any sensible way; and the objects in the samurai's right hand look like mishmash of sword/feathers.
          1,743
          120
          6.9%
        • ????? @gwern Mar 9 Mad props to & for being the first, out of at least a dozen to try over the past 2 years, to discover a way to bump a request to the front of the queue: - remove the request - block the person - wait 6 days (?!) - unblock - ???? request!
          4,813
          314
          6.5%
        • ????? @gwern Mar 9 You can't because the only additive variance that will be picking up is intelligence - the Big Five PGSes are terrible.
          60
          1
          1.7%
        • ????? @gwern Mar 9 Might be worth writing a longer explanation. I don't think I have anything in the page about why that wouldn't work.
          55
          3
          5.5%
        • ????? @gwern Mar 8 I don't need to 'believe hard' in the thesis. It's a simple thesis: 'could', ie. at least one. You either do or don't. Your misunderstanding or interest in redefining it to be about quantifying resource consumption is not my problem.
          365
          44
          12.1%
        • ????? @gwern Mar 8 "Pave the desire paths." Every ChatGPT error is a bug report in disguise. ?
          80
          8
          10.0%
        • ????? @gwern Mar 7 An optimal learner throws out as much as possible to distill only the sufficient-statistics for solving the POMDP, which may be much smaller than the real generative process: it only has to be *value-equivalent*
          204
          14
          6.9%
        • ????? @gwern Mar 7 No, it is, because there is no incentive to model the physics except as it improves the next-token loss. If it doesn't show up as one BPE rather than another, it's totally irrelevant and the model can't afford to waste capacity on end-to-end learning of it: because that's not end
          100
          10
          10.0%
        • ????? @gwern Mar 7 And given Eroom's law, capital risk aversion & discounting, experience curves, and the exquisite fragility of chip fabs and international supply lines, there is nothing even remotely inevitable about each new chip fab generation.
          70
          4
          5.7%
        • ????? @gwern Mar 7 Regulate compute/chip fabs. The exponential growth in compute requirements for training R&D means that '100 years' is easy to accomplish if you stop the exponential a few cycles before. You might have to let your A100s go brr for a century before you can iterate worth a damn.
          78
          10
          12.8%
        • ????? @gwern Mar 7 The anonymous source I was alluding to there was *not* Mikhail (unless he's really screwing with me). All my statements about Mikhail are based on his public Twitter comments (excerpted in a LW comment replying to my main Bing comment), which you can read just as readily as I.
          77
          18
          23.4%
        • ????? @gwern Mar 7 Yes, but it could also just be reserved hardware + spending extra compute to accelerate sampling... However, people were complaining about the turbo version being dumber than the slower one, so that IMO points towards it being a different model rather than latency tweaking.
          82
          8
          9.8%
        • ????? @gwern Mar 7 No, I didn't leave a comment because I wasn't sure I wanted to read the post enough times to figure it out and get bogged down in a comment. Not sure I like it suddenly being revived a lot.
          179
          14
          7.8%
        • ????? @gwern Mar 7 Yeah, I was unsure about that. No control of lines like a real tutor? On the other hand, actors all read from the same script, and you have psychiatry's therapist-specific effects despite dodo bird verdict (or regular teachers), so clearly the delivery can vary dramatically...
          356
          4
          1.1%
        • ????? @gwern Mar 7 You'll need much more than 'billions of dollars' to pay every author in the corpus $0.5m... (At a quick cut: Google Books estimates ~130m total book-like artifacts; median author publishes 1 book, let's say it's mean 2+single-author; >65m authors, so $0.5m/each is >$32 trillion.)
          68
          9
          13.2%
        • ????? @gwern Mar 7 It was coined fairly early on in the API release in late 2020, AFAIK, but while I'm sure I coined 'prompt programming', I'm not sure who did 'prompt engineering'. It might've been on the OA Slack which is long since deleted/inaccessible, so we'll never know.
          42
          4
          9.5%
        • ????? @gwern Mar 7 I didn't say they could do any better under the Communists. My point is that technology & capital are awesome and far better than 'working harder', and you can see that by asking how many meters every 3 days good ol'dynamite would be able to do.
          48
          1
          2.1%
        • ????? @gwern Mar 7 No, sorry: noot/drug stuff is at the bottom of the list for me these days. You might talk to the Qualia Computing guys, they have done a state-space thing or two already and would probably be interested in doing more work there.
          57
          6
          10.5%
        • ????? @gwern Mar 7 (I probably have, but I would be lying if I said I had any recollection or commentary on it.) One thing I found really intriguing about that Schank was the emphasis on benchmarking large real-world datasets and focusing on scaling - hard not to see ML, then DL, success in that.
          142
          5
          3.5%
        • ????? @gwern Mar 7 I still don't understand that, incidentally. It seems like some sort of very confused distinction about model-free (?) RL wrapped in a lot of confusion. Reward sure does seem like the optimization target for, say, AlphaGo...
          366
          22
          6.0%
        • ????? @gwern Mar 7 Already well-established by many earlier scaling results... Like I keep ribbing the 'continual learning' people like Irina: "are you sure your field's problem actually exists?"
          37
          2
          5.4%
        • ????? @gwern Mar 7 Eh. If ChatGPT was one of the small models or distilled, like ada/babbage/curie, it would be perfectly true to say that davinci is 'a much bigger model'. It is! by like 100b+ parameters!
          50
          9
          18.0%
        • ????? @gwern Mar 7 Yes, that's part of what makes it so plausible. But probably not true. I generally prefer to quote Carlsen: "I am convinced that the reason the Englishman John Nunn never became world champion is that he is too clever for that."
          46
          7
          15.2%
        • ????? @gwern Mar 7 Oh, I agree that the overall quality loss is due to the lack of ractor. You included the critical paragraph there, there's no doubt about what we're supposed to conclude: basically, Bloom's Two Sigma (altho dunno if Stephenson knew of that exactly or just tutoring in general).
          378
          8
          2.1%
        • ????? @gwern Mar 7 The scaling numbers you want are in 2¡Á FLOPS = 66% victory; amortization of training ¡ú runtime tree-search, where 10¡Á training = 15¡Á runtime.
          142
          5
          3.5%
        • ????? @gwern Mar 7 Sounds plausible, sure, but could you be more specific? What are the top 3 concrete research findings or technologies that American researchers are hugely disadvantaged on because it was published in Chinese rather than English? Especially in AI?
          91
          3
          3.3%
        • ????? @gwern Mar 7 (Now if I had to guess, my guess is that the trick is given away by the 'Mouse Army' name: it's a *peasant* army, the greatest nightmare of every Chinese dynasty ever down to the present. They were supposed to be conformist housewives, not conformist soldiers in an army.)
          199
          3
          1.5%
        • ????? @gwern Mar 7 (See the Mathieson bit about "New Atlantis, like many tribes, propagates itself largely through education. That is the raison d'¨ºtre of this Academy." What is taught to the many, is done by the few.)
          259
          5
          1.9%
        • ????? @gwern Mar 7 That is, the liberal-arts education is to create future leaders, not their followers; too many cooks spoil the soup. If you are not as talented as Nell, presumably the Primer system as envisioned would shunt you off to a station in life more befitting your gifts. Up or out!
          400
          11
          2.8%
        • ????? @gwern Mar 7 I'm not sure what the trick was, but reading the Victorians & tribes as 'subversion and individualism' is wrong: the point of the tribes are to be conformist & collective. The Primer is there to offer the "natural aristocracy" their chance to flourish, like Oxbridge tutoring.
          218
          10
          4.6%
        • ????? @gwern Mar 6 I think the challenge is that you are recruiting not for the sort of person who could be good at playing Factorio, but the person who has chosen to become good at playing Factorio. When I think of the best Factorio players I know... I'm not sure I could, or ?????, hire them.
          1,471
          82
          5.6%
        • ????? @gwern Mar 6 Blogroll for 'site of the day' and 'annotation/link of the day' now enabled. (I realize no reader is going to get the 'Swiss spiral dataset' / 'Swiss roll' / 'blogroll' visual pun, and will probably assume it's supposed to be sushi, but I enjoy it.)
          4,909
          92
          1.9%
        • ????? @gwern Mar 6 (He actually didn't, it's very apocryphal. I couldn't find it before like 1990 or something when I checked. It's true, though.)
          507
          43
          8.5%
        • ????? @gwern Mar 6 People think I'm joking when I talk about going onto Twitter in my dreams and just reading tweets as they scroll by, but I'm being 100% literal. So I definitely believe that one could use Loom in a dream.
          360
          26
          7.2%
        • ????? @gwern Mar 6 Like, globally? I'm pretty sure >50% of human workers still engage in some sort of physical activity beyond laptop-class wordceling, so text+video is not possibly adequate. Can't do any kind of robotics. You'd need to broaden from VIL to many-modal approaches like Gato/DT.
          1,122
          39
          3.5%
        • ????? @gwern Mar 6 American stores don't sell salt because we're so rich that salt gets put in food ??? us. Europeans are poor enough that they have to husband their salt rations. (We do, of course, buy giant 50 gallon¡ªnot liter¡ªbags to dump on our driveways. They wouldn't understand this.)
          779
          36
          4.6%
        • ????? @gwern Mar 6 The more such a regime engages in such acts, the more you want to incentivize defectors as it is prima facie more destructive/self-sabotaging of that evil regime to punish citizens, and the more evidence they are providing you that they believe defecting is a top danger to them.
          407
          5
          1.2%
        • ????? @gwern Mar 6 But is probably a large overestimate since despite being done 50 years later, doesn't measure total completed fertility AFAICT, and just the narrow 3-year window of Mincome. So this may be mostly measuring moving up childbearing a few years to benefit from the income guarantee.
          907
          18
          2.0%
        • ????? @gwern Mar 5 Does that prompt actually work? Alternately, can you just use this as a 'meta-prompt' so it generates arbitrary conversations 'within' the supposedly quoted prompt?
          2,324
          102
          4.4%
        • ????? @gwern Mar 5 The #1 way to make AI less safe is also to make it more intelligent, regrettably.
          1,525
          103
          6.8%
        • ????? @gwern Mar 5 The argument I make is that so many terrible terrible things happen to one in dreams almost every night, people would be utterly devastated walking corpses if they were experiencing it at even 1% of IRL:
          932
          103
          11.1%
        • ????? @gwern Mar 5 Good example of how working smarter is so much more important than working harder: the road isn't any better for being made with absurd amounts of labor & chisels than with decent technology & capital.
          2,076
          89
          4.3%
        • ????? @gwern Mar 4 Echoing it in a status bar would also be a very straightforward and useful way to confirm copying. Browsers deprecating status bars as much as possible is a mistake, like hiding the scrollbars and thinning them to illegibility.
          169
          28
          16.6%
        • ????? @gwern Mar 4 The n-gram comparison has always been baffling to me. You have to care enough about formal languages and computability to think that's a killer comparison, but then not care at all when GPT closes a quote or parenthesis with zero problem in the very first sample you read.
          580
          33
          5.7%
        • ????? @gwern Mar 4 But it'll be much more classist than that: "wow, why was everyone such a poor fat slob loser back then?" Even if post-tirzepatide become oral & cost as little as metformin, people won't bother, in the same way that quitting smoking costs negative dollars. So becomes class marker.
          1,066
          46
          4.3%
        • ????? @gwern Mar 4 You're trying to make a joke but generative models have been big in particle physics & astronomy, and substitute for 'just gather more data': they're drowning in raw data, when they need more insights and specific things to explore like
          507
          29
          5.7%
        • ????? @gwern Mar 3 None of those strike me as being as dangerous, and several were so uncontroversial IA even got special legal exemptions passed through Congress like its DMCA exemption for the software/games.
          127
          8
          6.3%
        • ????? @gwern Mar 3 Hardly their fault. You may remember a certain lawsuit being involved.
          110
          10
          9.1%
        • ????? @gwern Mar 3 I dunno if the publishers are 'out for blood', but you can see right there in the ars link that it was not obvious that they would sue and pursue it through a full court case all the way to the final closing arguments 3 years later, and the IA was clearly very unprepared.
          92
          5
          5.4%
        • ????? @gwern Mar 3 Can you name 3 examples of the IA 'probing' as flagrantly and jawdroppingly idiotically as the NEL? I've been following the IA with great interest since they launched in the 1990s, and I can't name 1 other.
          108
          5
          4.6%
        • ????? @gwern Mar 3 It's replacing plants, however, so the sunlight used to do work by storing CO2 and synthesizing sugars etc before the rest turned into heat. Not obvious to me how it'd net out.
          117
          3
          2.6%
        • ????? @gwern Mar 3 Well, you'd still have S-risks. Also, note that delaying AGI by a decade or two only kills about an eighth the population, even assuming no longevity progress; & there might be quite a bit of medical progress even with only sub-AGI AI, so the opportunity cost is lower than looks.
          786
          21
          2.7%
        • ????? @gwern Mar 3 'Russell trending': "I have become the first X to Y; you are raising questions about controversial new Y; he is violating all precedent of ~Y."
          691
          18
          2.6%
        • ????? @gwern Mar 3 I don't think anyone doubts at this point that [checks notes] 'GPT-2' is very limited compared to other models. And they really shouldn't be going around claiming that it's the *best* predictor when all they did was
          629
          61
          9.7%
        • ????? @gwern Mar 3 It's an unfair comparison, though, because you're stuck with zero/few-shot prompting of GPT-3 (plus the mess of RLHF tuning destroying instruct-'s poetry). I really think someone ought to pay for a GPT-3 finetune on poetry just to see how it goes, or try
          72
          5
          6.9%
        • ????? @gwern Mar 3 The first one is all about avoiding search because it shows the scaling laws for distilling search into a forward pass, and is the most definitive general answer you could have hoped for for your question. ??¡á? The second also directly quantifies the strength of KataGo w/o search.
          203
          19
          9.4%
        • ????? @gwern Mar 3 Yes. And in the mean time, if you're reconstructing a page by hand, you can try to fake-simulate the server: your local page makes a request for URL XYZ, which errors, and it gets looked up in the stream to see if that was requested somewhere in the exploration and is cached.
          101
          4
          4.0%
        • ????? @gwern Mar 3 There's already WARC and other formats developed by IA/Heritrix etc for doing exactly that. You're just serializing the HTTP requests, that's not the hard part. It's triggering the right requests and interactions to make a website's dynamic behavior learnable that's the hard part
          86
          6
          7.0%
        • ????? @gwern Mar 3 The ideal one could do now would be something along the lines of logging all traffic while a NN agent explores a website trying to trigger novel traffic, so future generative models can reconstruct it in its entirety.
          66
          5
          7.6%
        • ????? @gwern Mar 2 Th NEL was popular & I do know people were doing bulk downloads while the gettin' was hot. So while I doubt the publishers want to bankrupt IA (bad PR), it's staggeringly irresponsible of IA to put itself in such a position and the consequences may be quite painful and long-term.
          115
          4
          3.5%
        • ????? @gwern Mar 2 The problem is, it's like the RIAA music-sharing lawsuits clubbing people w/trillions of dollars in damages. How many millions & millions of copies were downloaded while the NEL was live? And what happens if each infringement is like $100? Ain't nobody got *that* kinda reserves.
          170
          4
          2.4%
        • ????? @gwern Mar 2 I still can't believe they actually did that. They didn't even have a blurb for the reporter asking them 'uhhh... how is this at all legal' And they all still work there! No one got fired for triggering the still-ongoing lawsuit that might still kill IA!
          3,603
          230
          6.4%
        • ????? @gwern Mar 2 I was writing up a brief explanation why, time permitting, you ought to make a copy of any Wayback Machine web page you need, when I kinda got derailed when I remembered the breathtakingly suicidal idiocy of the IA's 'National Emergency Library' in 2020:
          3,607
          181
          5.0%
        • ????? @gwern Mar 2 Never saw him myself but I heard that, yeah. I mean, who's going to tell him no? (Or pronounce it 'wang' rather than 'Wong' when he's around, no matter how much it amuses one?)
          192
          18
          9.4%
        • ????? @gwern Mar 2 The impressive thing here is that he's *still* smoking after all these years. He's looked at the risk, made his decision, and committing to the bit.
          1,690
          104
          6.2%
        • ????? @gwern Mar 2 TBH it'd probably disappoint you. Look at Linear B: how many Gilgameshes did we get there? (None.) The problem with palace scripts is that they're hella boring. Herculaneum is so appetizing because it's a philosopher-aristocrat's huge personal library, not some estate-accounts.
          89
          3
          3.4%
        • ????? @gwern Mar 2 "The monster aircraft...was not only dangerous but also very bad for the environment." ?
          49
          1
          2.0%
        • ????? @gwern Mar 2 4chan has been torrenting it for like days now.
          421
          34
          8.1%
        • ????? @gwern Mar 1 It *is* interesting how far it can get by remaining on-policy and setting up easy memorized rhymes for itself. But the more you read, the more you realize it's an extremely narrow (and easily recognized) region of poem-space. And it explodes if you push it outside that region.
          79
          5
          6.3%
        • ????? @gwern Mar 1 Yeah, but it's just morally equivalent to memorization. Same as PaLM and the 'miracle of spelling': sure, it'll learn any specific pair you want it to, but it won't learn the flexible generalizable capability or the presumable semantic benefits of actually understanding it.
          66
          3
          4.5%
        • ????? @gwern Mar 1 He is actually thinking of a relatively recent (Dec/Jan) viral Tweet thread where the author, based on unspecified consulting for unspecified companies, dilated on how Musk's Twitter was doomed Real Soon Now. As they presented only sub-anecdotal data, no one should care about it.
          45
          1
          2.2%
        • ????? @gwern Mar 1 But that's a very small multiplier compared to many. You could slice it by ethnicity, for example, and your East Asians or Ashkenazi Jews will outperform a mere 50%. And considering how many of those highly-productive immigrants will be eg Russian Jews...
          85
          2
          2.4%
        • ????? @gwern Mar 1 *Now*. Maybe. Also, note the implication that they are using *only* fine-tuned models in some cases, and that they weren't using RLHF models *before*. If they were all always RLHF, you wouldn't say 'differently fine-tuned and RLHFed models'...
          80
          7
          8.8%
        • ????? @gwern Mar 1 So it's 16% vs 23% or something? That doesn't sound like a 'wow'.
          1,442
          43
          3.0%
        • ????? @gwern Mar 1 It doesn't seem that hard. Copying tokens is what induction heads do, after all. Copying the input to the output also seems to just be LLMs' universal fallback strategy when they are very confused: you see it all the time in GPT failure modes.
          712
          10
          1.4%
        • ????? @gwern Mar 1 Which is of course also why mechanisms which repeatedly iteratively pass or fail chatbots¡ª'serial passaging' if you will¡ªand kill the losers, will ?????? deceptively-aligned AI. They may not call you slurs inside their EEA, but you don't know what they will do outside it...
          1,639
          74
          4.5%
        • ????? @gwern Mar 1 Andrew Gelman has a line somewhere to the effect, "Good ideas don't require lies."
          37
          1
          2.7%
        • ????? @gwern Mar 1 "It's bad on purpose to make you click." Dollars to donuts this is staged; the indifference is carried too far to not be deliberate. Also, protagonist has social-media good looks.
          480
          36
          7.5%
        • ????? @gwern Mar 1 "Halo effect" covers the broader cognitive bias of insisting that something good in one way is good in implausibly many other ways too.
          779
          13
          1.7%
        • ????? @gwern Feb 28 (Bitmaps of memory would be particularly good because it seems like the sort of thing that will work better than you think and make a lot of people really mad.)
          249
          12
          4.8%
        • ????? @gwern Feb 28 A bitmap of pre/post-function RAM, a serialization of the AST at various stages, annotations of various kinds of taints/inferred types/precisions/overflows/invariants... Sky's the limit if you have the context.
          786
          30
          3.8%
      You've reached the end of Tweets for the selected date range. Change date selection to view more.
      Engagements
      Showing 31 days with daily frequency
      Engagement rate
      4.1%
      Mar 31
      5.6% engagement rate
      Link clicks
      3.1K
      Mar 31
      23 link clicks
      On average, you earned 101 link clicks per day
      Retweets without comments
      0
      Mar 31
      0 Retweets without comments
      On average, you earned 0 Retweets without comments per day
      Likes
      3.7K
      Mar 31
      24 likes
      On average, you earned 119 likes per day
      Replies
      335
      Mar 31
      2 replies
      On average, you earned 11 replies per day
      He interrupted her. Close at hand is a stable where two beautiful ponies are kept. They are snowy white, and are consecrated to the goddess Ku-wanon, the deity of mercy, who is the presiding genius of the temple. They are in the care of a young girl, and it is considered a pious duty to feed them. Pease and beans are for sale outside, and many devotees contribute a few cash for the benefit of the sacred animals. If the poor beasts should eat a quarter of what is offered to them, or, rather, of what is paid for, they would soon die of overfeeding. It is shrewdly suspected that the grain is sold many times over, in consequence of a collusion between the dealers and the keeper of the horses. At all events, the health of the animals is regarded, and it would never do to give them all that is presented. On their return from the garden they stopped at a place where eggs are hatched by artificial heat. They are placed over brick ovens or furnaces, where a gentle heat is kept up, and a man is constantly on watch to see that the fire neither burns too rapidly nor too slowly. A great heat would kill the vitality of the egg by baking it, while if the temperature falls below a certain point, the hatching process does not go on. When the little chicks appear, they are placed under the care of an artificial mother, which consists of a bed of soft down and feathers, with a cover three or four inches above it. This cover has strips of down hanging from it, and touching the bed below, and the chickens nestle there quite safe from outside cold. The Chinese have practised this artificial hatching and rearing for thousands of years, and relieved the hens of a great deal of the monotony of life. He would not have it in the scabbard, and when I laid it naked in his hand he kissed the hilt. Charlotte sent Gholson for Ned Ferry. Glancing from the window, I noticed that for some better convenience our scouts had left the grove, and the prisoners had been marched in and huddled close to the veranda-steps, under their heavy marching-guard of Louisianians. One of the blue-coats called up to me softly: "Dying--really?" He turned to his fellows--"Boys, Captain's dying." Assuming an air of having forgotten all about Dick¡¯s rhyme, he went to his place in the seat behind Jeff and the instant his safety belt was snapped Jeff signaled to a farmer who had come over to investigate and satisfy himself that the airplane had legitimate business there; the farmer kicked the stones used as chocks from under the landing tires and Jeff opened up the throttle. ¡°Yes,¡± Dick supplemented Larry¡¯s new point. ¡°Another thing, Sandy, that doesn¡¯t explain why he¡¯d take three boys and fly a ship he could never use on water¡ªwith an amphibian right here.¡± Should you leave me too, O my faithless ladie? And years of remorse and despair been your fate, That night was a purging. From thenceforward Reuben was to press on straight to his goal, with no more slackenings or diversions. "Is that you, Robin?" said a soft voice; and a female face was seen peeping half way down the stairs. HoMElãñÔóÂÜÀ­³ó ENTER NUMBET 0016www.jjyygo.com.cn
      emxytea.org.cn
      www.langnest.com.cn
      gxwm.org.cn
      hskhdzsw.com.cn
      www.shqsxg.com.cn
      rfunbw.com.cn
      www.mydario.com.cn
      www.neuvo.com.cn
      www.woooyoo.net.cn
      插你肥穴 制服丝袜旗袍涶乱的小电影 日本兽王无码全集 朋友们开撸吧欧美图片 � 人兽性交影片 维族人黄色片 日本女孩小嫩穴摄影 三八人体图片基地 美国 女人体图片欣赏 女人洞人体艺术 强奸做爱短片小说 父媳乱来 成人乱伦小说狠狠 91自拍电影网 我的美艳岳母小说撸二哥 谁有仓井空的电影 堂成人免费视屏 最大 拾贰月重磅巨献高清私拍安徽合肥随妻5p av女优超市 tube兽人 人体少女成人艺术 丝袜淫娃女空姐 欧美熟妇刚交图 新号百度云怎么看片 李忠瑞影音先锋下载 日本女人超黄色人体 男女明星日逼图片 sss性交图片 超级玛丽兄弟电影三级片 gaoqingtoupaiyiyuannvce 擢撸少妇色图 美女逼图欣赏 他州色图 麻岛美几 jav365熟女俱乐部 夜夜摸东京热 7788dcom 亚洲成人电台 美女自己玩射了18p 亚洲美女掰鲍图 色妹妹影视下载 丝袜美女露阴门图片 小说 吃咪咪 黄色电影全草免费 爱爱医学网 干骚逼 射进肉洞 迅雷长片狠狠撸 大胆97人体 WWW_PS31_INFO 胖妹子和瘦妹子做爱有什么不同 WWW_ZARA_COM 明星18avday 乳交 乱伦 鸡巴 女儿的嫩洞逼 人妖变装图片 忧忧人体艺术导航 日韩考屄视频 粗鸡巴日骚屄 快播52淫 超诱惑骚妇奶子图片 黑巨屌插烂插扯阴道 四季美农贸城 招商银行客户端下载 情债电视剧 时空猎人电脑版下载 黑男白女种子 谁有亚洲美图网站2014 黑丝电影有哪些 苍井空的逼放大图 鼘掠牒趯? 就要搞妹妹 撸撸网王老撸 大鸡巴肏死小骚屄了 狼友迷奸自拍视频 韩国a片观看 徐莹性爱视频ed2k 女人和畜生做爱的小说 日本爽片种子 极品淫荡骚货 操逼有害不可大力日比 大香蕉很狠撸 喜欢日肥女人 小说享受泰国浴 亚洲成人10p 草小少妇屁股 李宗瑞100级 少女的阴蒂儿阴道 六九av影院 林心如两腿分开 汤芳最大胆全裸体清晰图 哪个网可看黄色图片 狂插妹妹15p 影音先锋黑人幼幼 WWW_211VV_COM av怎么看不了 给德国处女开包 夜色贵族成人小说区时间 美女爱骚逼 爱爱小说乱伦 欧美女人全裸照片 幼幼被3p 大吊黑逼 粉色骚逼 我插过的女人 日本熟女友田真希电影影音先锋播放 草裙操比 我要肏贾玲 黄色电影影音 小姐的阴道夹的我好爽 ww66avav 国语影视先锋 手机皇色视频网站大全 WWW59QRCOM 五月亚州婷婷炸天撸 张筱雨被男人插人体写真 天天色偷拍自拍 欧美图库亚洲图库 欧美人妻性爱写真 成人电影网址youjizz 美国骚逼想唉操 爸爸插错洞吉吉 操少女小说网 视频偷拍逼 成人网黄色片下载 跟老婆操奶操逼 大鸡巴狠插黑黑的骚逼 骚妇粉色木耳 四房开心播播图片 雨宫琴音qv电影 久久视频在线播放观看( 去干网1月25日 亚洲性交色视频 爱爱帝国钟和区图片 粉嫩写穴 快播伦理影院全国若妻 WWW578133COM 日韩孕妇做爱 成人在线电台 大学生妹妹穴穴 快播户外乱伦电影 亚洲女色网伦理 婶婶与侄子伦理小说 苍井空秘密搜禅观 德川重男种子 佐佐木春香 qvodk8经典强奸乱伦 女鲍鱼16p 幼嫩系列在线 内射欧美 我爱插逼淫淫网 啊淫水图 奸恋图片 我和小姨在车里 骚妇淫 欧美大鸡巴照片 漂亮女优伦理电影 性爱小说97 国产呦呦100部精选集 艳女淫妇 尻逼电影 乱伦乱摸乱交配 哪个明星的逼最好看 春色校 妻子让朋友干出尿 无需播放器在线观看西山希 色喜大胆人体 国产自拍偷拍在线播放若怒 成人孕妇性爱电影网 美女掰骚鲍www51gannet 骚逼蜜穴 家庭乱伦动漫电影 盗撮协和影视 射死你天天日百度 儿子射进来 怡春院人兽 操岳母色色网 少女嫩逼激情小说 自拍偷拍欧美美色图 家庭乱伦之巨乳妈妈 美女游戏封面相片 黑丝avi码合集 帅的男同性恋做爱的视频 本色国际 西西人体高清18岁 人体艺术舒淇艳门照 影视先先锋 亚非成人电影 究极人渣素人 weimeiqingchunccom 老头和妇女公园不雅视频 来点黄色的videombaiducom 色漫画之丝袜少妇 第一放映室伦理欧美 狗肏人网 农村tongjian 能直接在线观看图片的网站 操你妹手机电影 妇色小说 国产抽插开档黑丝女友 色狠窝 朋友的处女女友m06xscom 伦全集狠狠射 tt老婆 yingqi 制服超短裙番号迅雷下载 变态图区精品 谷露英语 日本一级毛片magnet 徐东东丝袜图片 AV区亚洲AV欧美AVwww9966acomwwwc5508comwwwshcateringcom 女性全裸近照无遮挡物 短篇剧情网站 亚洲无码女教师撸 午夜A片播 大尺度强暴性爱 色网站色网站 久草在线被监禁了的身体玩具 欧洲学妹性交视频 当着外人的面插入下载 日本美女性交24p图片 renyudongwuxingjiao capron马蓉出轨 明星合成欧美色色欧美性爱区欧美骚妇欧美成人色图欧美大奶欧美大胆人体 真人做色性动态图片 八戒街拍裙底 春药潮吹痉挛视频 思思热资源站 久草在线视频7mwwwgsobc1024cnmyiwancom 比基尼淫 花井美纱RKI252磁力链接 少妇熟女搞逼 裸体女奴小说 两性故事淫荡家庭 日韩少妇人妻系列 人妻的禁忌 riyelu国语 耽美骚货出水 17岁少女跪嫁68岁老汉 性交口交天天撸 制服丝袜1页日日摸 b3jjj 哥哥综合影院雅蠛蝶五次郎 偷拍自偷人妻自慰 西方人与狗性交片 艳照门视频华为网盘 偷拍性生活祝频 WWW_621YY_COM 欧美女生阴道艺术 sek115苞米地 影音先锋a9电影 亚洲成人动漫av 97成人在线乐子 2015Tv丶C0M 星河大帝同人小说 夜插 日本熟女性爱偷拍xjo530tumblrcom 美女b比图片 小泽玛利亚与人妖图片 成人大胆诱惑人体艺术图片 大色se片 偷拍自拍内衣12p 少女小山雀69Xx wwwfafa98om 美国发布站2K3HH wwww336bbbcom 成人做爱有哪些影院 亚洲777www69rhcom 尼姑云影视 制服丝袜图片网站 vr美腿丝袜 天天操日日艹 破处番号 偷拍影音先锋电影网 意淫强奸人妻女友 AV电影百度云 ggggbb 北原多香子全集种子 97资源www688bxcom 正在播放87午夜孕妇 无码一级黄视频 强奸乱伦 哥哥干 色狗狗综合视频网站 先锋影音成人动 日本av百人群Pav 成人3d卡通动漫在线 wwwjjzzocom 成人性全免费视频 网55eekcom 色和尚被偷拍的女主播 微信可以处理违章吗 a图片大全 欧美成人裸体网 欧美家庭伦理片资源下载mp4 妓女av网 欧美图片激情小说 1333ccc ee214magnet 最新日本av在线视频 姐姐爱爱集 狠狠爱五色天大香蕉 哈尔滨狼友 乱伦日韩欧美制服手机在线 wwwavsj 64sm 偷偷拍自拍15p 国产偷拍厕所在线视频 强奸军人母亲 日本有个是主播做爱的片子吗 亚洲第一在线黄色网站 东方东方伊甸园苍井空 蜜桃丁香 www2Csex872com 231avav 爱吧综合 黄色网站空姐被强奸下载 啊扣你的小穴 五月天激情网址 大香蕉精品520 色中色AV在线视频哥哥干 奇米成人影视色和尚、色尼姑 新西兰成人片 500av大香蕉 黄色1级片 看欧美AV剧情片的网站 性插网页 av大b视频 撸波波映院 人妖当着妈妈和爸爸 v2视频网站 千百撸综合 教室调教老师 另类小说专区在线 av在线总站 www725bb 摩擦影院5566 女人的穴是万能的 乱伦性交激情小说 亚洲1AV 高清av1080在线 大香蕉波波在线视频 淫汁Y 清纯唯美哥哥撸 sex888欧美片 依依社区影院撸lu射 995hao m3344tt 草民盒子 西西人体偷拍另类 狠狠操骚b 亚洲偷拍自拍最新网站 制服丝袜av天天 3d黄片 kk999pw下载 WWWavtb789com 男同志免费影片 琪琪电影网李宗瑞 色在线奶奶 579路com wwwyoujizzcom美女 有没有幼女的视频网站 欧美私处艳照 亚洲图片欧美图片日本av成人有声 国产小黄片网站 搜索www123456789 撸一撸美女人体 操小屄还玩老屄 淫淫色色色色 深爱萝莉 花样困绑虏待女人方法 很哈图热图 wwwcom84 wwwsyy31con 草莓社区2016新地址 国产av91自拍 鸡鸡和光屁股 夜一夜撸一撸 91luCOM 最新乱伦比较黄的肉文 嫩妹妹色妹妹干妹妹 av英语老师 美女合集早时候的图片 韩国浴池在线播放 奶大性爱基地 L三级片 美女的逼怡院 影音先锋偷看媳妇中文 91qq1com wweuucom 美国伦理在线 狼窝激情影院影音先锋 韩国邪恶动做片mp4 古代激情17P 男操男小说 干哥哥插妹妹逼 成人激情12P 好色视频超碰碰碰 yaolu导航最新地址 同性恋舔我小穴 在线视频主播AV 手机看片www555com 整部的黄色小说 男人强奸女人操逼的视屏 色黑木耳穴 经典乱伦性爱故事 色四图吧 广濑奈央美 无码 国产成人在线视频网站 女知青风流小说 高兴为什么呢哪俩唱的 女人wumaoyingbutu 儿子与母亲的乱伦性爱故事情节 爽图在哪里看 屄 亚洲 hongkongav 乱伦熟女做爱视频 欧美破处视频区 爆插长腿黑丝妹50p 后宫社区最新最新地址 最新肥胖人体艺术摄影 淋浴做爱av 李宗瑞性侵录像网址 热香屄 女主播艳门照 车上逼逼图 为什么我喜欢让情人操我 性爱av哪里能看 影视先锋伦理电影 国内老女人av在线视频 欧美黄色网站视频第一页 美女乱论小说 日本人操b图片 神木轮奸犯王伟 与狗爱爱网聊视频 嫩乳妹妹人体艺术 1238090com 郭富城早期经典视频 女忧快播影音先锋 色色偶性爱自拍 肏嫩屄舔雪白屁股插进去 叶欣桐图片库 国产视频操胖老奶奶10306 激情爱爱1qq号是多少 酒色视频网最新网址 成人电影微信群 苍井空的电影名字 想几个男人操逼逼 在酒吧把泥酔巨乳女带回家狂插先锋影音 色mm激情万种 美女人体馒头 明星华为网盘 老妈与我乱伦 亚洲色图 女大学生性爱 毛色黄片网站 非洲美女陆逼网战 偷拍亞洲路上幼小情侶做愛相片圖? 女人白虎屄 bb性爱视频 大鸡吧干衅 日本成人强奸乱伦电影 在哪能找艳照门 欧美嫂子qvod 我喜欢大奶女人 亚洲爱色套图 遥子里被操的图片 色爱在线视频 36周了感cnbb头己下了 不举总裁的盗种妻 美国最年轻的总统 wendao 甚至导致严重的肌腱炎。建议不要长时间发短信, 梦幻西游69群p 武器战士输出手法 韩国人体艺术新闻 关于怡红院的网名 少女18艺术图片 色沟影院 女人杂交视频 大鸡巴姐姐肏骚妹妹 日本人体艺1 上原结衣的片去哪个网站找 makeloveveido 91快播电影国外处女 人体图全女 女子健美漏屄图片 美足自慰 畸形逼 吉泽明步orrent 人妻性爱色图影音先锋 女艺术图 qovd幼交电影 韩国女明星被潜规则爱爱电影 快播免费成人禁片 日本强奸资源 熟女尿射口 鑹叉儏缃戠珯链夎皝鐭ラ亾 午夜性交图片 走光露屄毛的女人 操妹妹逼 撸撸涩涩图片 惩罚骚货熟女 小岳母的淫水 狠狠干激情 人体艺术女同性恋视频 暴操成都艺校妹妹 我爱和妈妈和妹妹乱伦小说 成人极品小说合集 人体射精艺术图30p 三级片3级片四级片视频 美女的下部肉肉图 和女仆干真爽电影 在线视频录像 WWWAAACOMYAZHO 奇艺人体艺术 人狗性交私处照 插妹妹自拍偷拍 RE789COM 鲁鲁射射 深爱激情网免费视频播放 清纯唯美中文 橹一橹小说阅读 骚女嫂子 幼女小说长篇 屄都什么样 咪咪网王老撸 手机浏览器看不了在线视频 婶婶在果园里把我了 少妇熟女乱伦强奸 粉红亚洲妹偷拍自拍 WWWWWNNVCOM 国产av大片 一圾片2013推存 日本少妇伦理小说 黄昏去操逼俄罗斯 玉女情挑吉吉音影 色四月婷婷网五月天肉文 亚州成人大奶综合电影 老江吃嫩草20 亚洲怡红院首页 日本城人电 色哥百度 WWWGAN521COM 激情网站wwwhuangsedianyingnet 强暴乱伦爱爱 httwww8888tpcom 台湾无码A片欲火焚身 婷婷五月亚洲Avwww317avcom wwwAVKUTV 本站只适合18岁以上观看 tom色播网 免费成人网绐 www111eecom四虎影裤成人精品 小苍奈奈网站 妻子地铁失身 印度毛片ed2k 儿子插完妈妈插妹妹 xingjiaoluanai WWWWEIBOCOM 淫姐姐骗了弟弟操她逼 小鲜肉操熟女 老色哥奇米影视26uuu 插进妈妈的肥逼 网友自拍偷偷师爷 成人动漫亚洲色图意淫强奸亚洲色图意淫强奸 好骚色色re 不小心小孩插大人影音先锋 韩国少女掰屄图 wwwqylsp1 YY6080PAPAPA 中国空姐干逼视频 观看美女自慰直播的网址 2017搜片神器是那个 强奸色女网 日韩人与兽爱爱 母淫网姐姐 亚洲娃娃 猛操骚货痴女护士文 亚洲大奶老女人 韩国电影熟女的欲望 偷窥厕所漫画 儿媳妇与公公三级片电影 好看站撸啊撸撸 首页偷拍自拍朋友女友的嫩穴毛还是初长成那种 李宗瑞生殖器大吗 肥乳性爱 有声小说ok网 笑傲华夏有声小说 春色夜生活 知春色 樱井莉亚下马 樱井莉亚体重 樱井莉亚123 小泽玛利亚迅雷 大家给个h网 东京热n0570 东京热n0393rmvb 可以用快播看的黄片 哪里可以看黄片 男人为什么喜欢看黄片 川村まや 33链导航 性生活知识 一点色导航 tv69在线 插插洞插插B 多情成仁高清 群交乱伦电影 色人居成人网 性乐汇色电影 淫色网 yeye撸 撸撸侠网站 苏紫紫干露露全套 干海参怎么发 Av天堂色电影 4099y影院 3d爆乳女教师西瓜 精品ady9 52AVAV好色色屌丝 国产精品情侣av 西野翔被偷窥 怡红院福利二区 在线阿v在线澳门 人人操美丽人妻 人人摸人人揉视频在线 在线无码私人影院 午夜福利1000部757 咪咪av电影 海外啪啪啪视频 南方来信免费观看 100国产自拍 rbxjdh 七色沙耶香番号 芹泽直美诱惑女教师视频 青青草原影视免费观看 桥本美步百度云 青草碰碰 青青草人碰人视频在线观看 在线看片 欧美自拍 青青 日福利av 日本情爱电影 孕妇在家孕交视频乳交 樱濑奈bt 日本少妇cosplay母狗 日本桐冈市 日本男女抽插插福利免费动态资势 在线男人福利免费视频 爱青岛在线观看 在线小萝莉 美女主播自慰出水视频 VEC-360部长奥-八木梓 麻仓优112部全集 迷夜惨案在线观看 痴女诱惑 稀缺资源 九州av 夜趣阁AV导航 无码av种子在线观看 luoliai 黄色片大胸 日本在线强奸乱伦 淫妇骚屄欠肏视频 色午夜v 小操黄色免费管看视频网址 日本亚洲欧美无码免费视频 草莓网 日本冒死厕拍视频 大香蕉一本道富二代 天天影院色小 凹凸視頻免賛在線 女主播叫床视频在线观看 行交视屏 黄色三级轮奸高清 宅男伦理 在线观看成人影院在线 xinxin68 org 国产自拍在线更新av 蓝沢润手机免费视频 你懂百度资源 和充气娃娃爱爱视频 龙头蛇尾 童颜巨乳 色和尚在线视频久和综合色色 江川美奈子在线视频 3344df 怡夏影院成人快播 1204手机在线播放下载 4438x一全国大色 露谷试看 色女孩影视最新网址 欧美群交网 波多野结衣来广州 日本she射精视频 自拍内裤爱液 伦理片自拍短视频 爱草色综合 彩美旬果xfplay 亚洲 自拍 porn 亚洲种子av 苍井空AⅤ下载迅雷 磁力链 下载 rki435 488成年网站 52av2av av天5 极品红人cutie queen 浪货成人网 91p app 上原亚衣黄色 中文字幕av12免m看 www路AV 亚洲36D 福利视频偷拍 大香蕉在緌人妻 福利avi 福利roufan动漫在线观看 大香蕉s视频 迪厅里的黑人男女裸体交配视颖 大神夜店大山极品美女带回高级寓所阳台一直 xxx-av 20147-20148 青鱼视频2自拍 国产人妖 鬼父有无码的么 韩国三级电台在线直播 劲暴欧美 一本道 东京热 自拍 国产 无码 quanqiuzuidadechenrenwangzhan 午夜小影院。 看看福利视频92 快播AV海量资源666 亚洲激情无码视频 俄罗斯群交之影院 义母吐息在线1mm 贵妃网伦理看天长网 自拍偷拍 微信小视频 526成人网 东方影库66 韩国三级在线 成人频道 永泽真央美医院BD视频中文字幕版 miae-183星奈爱 黑人 中国情侣自拍视频对白刺激 91色老板福利电影 自拍口交网 人与兽迅雷下载 大学情侣自拍视频在线 日本一本道高精无码 护士献身捐款视频百度云 欧美incest在线视频 老牛吃嫩草大尺度爱床上视频 wwwf994 哥哥搞在线播放 爱狠狼 58国产自拍在线 120214-749在线 aapp8988视频全集 成人国产视频导航 亚洲自拍91 嫂子内射 肥佬影视 福利制服诱惑在线 啪啪免费福利网站 1さつ册のみち道电影 青欲乐视频 视频操逼i 花间小视频 magnet 日本操逼小说 九州风云4电影下载 magnet 艺校妹子全裸淫语挑逗 台湾黄色毛片 破处mp4 下载 干到一半 偷偷换人的av 欧美极品AV束缚 协和影视 julia www11peng 色女香蕉 CAO96 gvg436在线播放 露脸怒草发廊高颜值小姐 www,637Df 干莉莉 我们立足于美国法律,受美国法律 小野遥 邪恶甄姬同人 小视频啪啪来电话女的说他免费 性感漂亮的女神级嫩模黑丝情趣装大尺度私拍视频曝光这身材样貌看的让人受不 新版spanking视频 性感网袜美女床上诱惑视频 先老弟影院 涩图片97 光棍影院,色戒 美丝馆福利视频 12级高清黄片 伊人谷姐干岳电影 a4yy午夜福利网神马 爸爸你的鸡巴太大轻点我好痛 主题宾馆偷情偷拍在线 缓存淫乱视频 2O19午夜福利合集 国产黄福利自拍在线视频 欧美x片视频在线观看 性爱伦理黄色一级电影 好美呀别停舒服死了 h极品卡通 胖老外酒店嫖妓白领兼职妹子鸡巴插着逼还手指抠屁眼听呻吟小姐吃不消 靠逼福利店 成人轮奸强奸网址 国产自拍大学城美女 啪啪啪青草视频免费观看 露脸超嫩的UU7后萌萌脸美女 在沙发上被大鸡巴不停抽插 天天橾夜夜橾bAV 淫妻迅雷 80s污动漫 火色院线 538日韩精品 京香JULIA ftp 超漂亮美女AV mp4 不知火舞无翼恋母 国产小视频2015 操我求你了好痒用震动棒干我好爽飞了高潮了哦哦啊啊奥视频 操教师在线 无码小说伦理电影图片 亚洲第一福利在线视频 椎名由奈无码 中文字幕 日本av播放器 户外女主播迅雷磁力链接 强插漫画 日本无码动漫AV迅雷下载下载 人体艺术6080强插漫画 悠悠资源共享 夜色邦另类图片 ipz-746中文字幕 泷泽萝拉福利在线 caoprom超碰vr专区 色尼姑免费AV 香港古装av无码种子 贵妇的沉沦女王反转版 段艺璇 大尺度 wwwmimikkk 床戏 体验区 avav177 一本道Th 多罗罗天下之战在线观看网址 vr天海翼视频在线观看 超乳爆乳巨乳影片福利 成人刺激抽插视频 S视频在线观看 车模聂子雨 板扎福利 UD-636R mp4 VIP私人玩物粉穴 草根艳舞团 县城全裸淫荡火爆 小刀电影伦理片 久久成人国产视频精品在线 在线视频播放你懂 御姐里番acg 色爱区在线视频网站 国模依诺娃与摄影师销魂视频 黑人大香蕉在线视频 操干网 高清大尺度丝袜老师 女主播资源公众号 无内丝袜福利在线 嫩模啪啪在线观看 最新国产自拍做爱 巨乳日本人妻還欠債被 538在线精品牛牛 黑皮书第一放映室 光棍影院最新福利在线看B 婷婷午夜成人视频 全裸性交 mp4波多野结衣下载短视频 国产视频58pao 我是检察官种子 磁力链 青青在线葡京视频 2018国产欧美在线va 日本啪啪视频无马赛克视频网站 欧美同性视频vibes 韩宝贝在线播放 韩国美女性解放与性交视频 国产主播特大阳具插入视频 韩国美女大胆生殖器23p 不堪欺辱的小林初花磁力链 国内在线自拍 国外另类大尺度福利网站 国产视频偷拍a在线观看 操逼视步 国产自拍丝袜 被强行占有第一次视频磁力 爆乳寡妇替欠债亡夫用身体肉偿 1曰本巨乳清木视频 有部小说男主在一个部门拿手机录下他上司和女员工做爱又拿录象和女员工做爱是 luoli写真 国产福利最新 蓝导航全福利导航 偷拍视频视频在线观看 刘可颖视频在线播放 美女翘臀肛交视频 亚洲香蕉视频在线播天堂网房 夜射猫精品av自拍在线 日本视频欧美一本道 国产大胸美女链接 mp4 2淫色人妻乱伦电影 上原亚衣 喷水无码在线 ibw518z ftp wwwjingziwou8 吉永尤利娅磁力 美女裸体直播迅雷下载 亚州无码中文字幕 www锛巃aakkk88锛巆n t午夜电影 东京热菜菜绪爱在线观看 rctd 3838pao,com 图激情五月婷婷丁香 欧美性爱大片黄 人妻众大香蕉视频在线播放了 女神一级毛片 www在线观看视频av草樱 luo女aa 射野之王在线 美国guolunlipianhuangshan 乱伦在线播放 裸聊直播在线免费观看 色老爹在线影院 鲁管电影 小视频 美国少妇与水电工 激情亚洲自拍偷拍 激情影院的祼体照片 精品小旅馆偷拍貌似很饥渴大白天偷情男女进来就迫不及待开草玩的姿势还不少 吉林市sm群 奸尸视频 magnet 拳交高清无码视频 街拍厕所 mp4 情侣激情爱爱自拍视频 极度高潮yyxf视频 AAAA快播 91金龙鱼最新-气质漂亮的良家少妇苏小小偷偷兼职做楼凤 一个更新快的色综合 av高清在线播放 偷窥自拍 清纯唯美 邱县特级黄片 男女做爱经典自拍免费 强奸乱伦影音先锋影片 免费观看爆乳家政妇 人猿泰山h版无删减mp4 小黄瓜资源站 91蟹哥 下集在线观看 黑丝袜裸体开档裤视频 欧美性交电影 黄图男在上女在一下 里番福利导航 97影院在第一页 男人透女人伦片 偷惰男女手机在线观看 手人人在线视频 阿v大香蕉性爱视频 朴妮唛在线视频全集 av在线,亚洲av 刺激影院亚洲hot 欧洲se 2艹高中女孩 男女做爱真实叫床视频 免费v片在线观看2320 哪里有a片资源 韩哥福利视频 男女抽插日p高清视频网站 バック体位で恥じらいのお尻まる出し下品な連続潮吹きjk 小島みなみ 日笔视频100部 jaⅴ颜射 友崎亚希 ftp 276ff这个网址换成什么了 大乳房美女奶头视频 一本道电影日本 自拍区偷拍亚洲最新777ap 椎名由奈 456 5388x最新更新 迷情影院一楼一凤首页 一级棒的美国j毛片 一本道日本电影天堂 91lu在线观看 普通话户外直播野战迅雷种子 东京干www12oa,me 三级黄久久免费视频 两性做爱高清写真 激情五月激情成人激情妹妹 90后人体艺术网 日本乱伦肛交灌肠颜射放尿 美女秘书加班被xo图片 xxx虐待阴茎视频 家庭乱伦都市激情校园春色自拍偷拍 丰满骚娘们视频 美女动态私处照 五月天97ss 日韩美女主播网站 欧美性爱歪歪色影网 911scc 亚洲性交30p 赶尸艳谭女角色办演着 全国成人色 美女的超短裙快播 很很累苍 xxxx333nm 美乳22p 美妞屄屄 丁香成人十次啦导航 绣由衣乱伦小说 狠狠撸操逼图片