lãñÔóÂÜÀ­³ó

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

      Tweet activity

      October 2022

      Your Tweets earned 798.3K impressions over this 31 day period

      50.0K20Oct 2Oct 9Oct 16Oct 23Oct 30
      Your Tweets
      During this 31 day period, you earned 25.8K impressions per day.
      • Impressions
        Engagements
        Engagement rate
        • ????? @gwern Oct 31 (Ah, thank you. That was actually the one I was trying to think of because of the completely absurd logos/branding combined with being water - not even carbonated, originally, just plain bottled water!)
          144
          4
          2.8%
        • ????? @gwern Oct 31 But it ???? make a great Halloween theme, so in honor of the occasion... ? Also hints at how a rubricated light-mode design could work: very thin red lines aren't overwhelming & provide structure. (And for dark-mode, switch to a green-blue?)
          913
          19
          2.1%
        • ????? @gwern Oct 31 Agassi was tennis, not chess. If you mean the Polgars: because his kids liked it and competed against each other, and after he hit the broad side of a barn, he drew a circle around 'chess'.
          123
          7
          5.7%
        • ????? @gwern Oct 31 I'd say beats the hell out of Agassi or more tiger-parenting like piano or violin lessons. (NPR last night: "...XYZ started violin lessons at age 3..." ME: "Uh huh, and how miserable were the other 99 kids who never touched a violin again after pissing away childhood on it?")
          1,556
          37
          2.4%
        • ????? @gwern Oct 30 (If you think about it, the real conspicuous consumption here is implying you either have learned the extravagantly useless Elvish to visually type, or to touchtype entirely.)
          3,537
          60
          1.7%
        • ????? @gwern Oct 30 Yeah, it's highly speculative but I *want* the ancient-Egypt theory to be true and not just some boring 'very slow universal soft selection towards domesticity'. Cats deserve cool backstories. He would disagree - I don't use the ripple rug or feed him *nearly* enough Greenies!
          128
          7
          5.5%
        • ????? @gwern Oct 30 (If you think it feels bad now, wait until a few years from now when you run a SOTA model over all your emails and texts to get a few KB/MB embedding finetuning it to 'you', and none of your friends can tell it's not you when you prank them with it.)
          114
          10
          8.8%
        • ????? @gwern Oct 30 Yeah, best-of is not a very good search strategy. Models should probably just be self-distilled/finetuned on a bunch of best-ofs for a one-off performance boost like inner-monologue-finetuning, and with the spurious sampling noise removed, real search strategies used.
          49
          2
          4.1%
        • ????? @gwern Oct 30 I have a water fountain; a relative feeds him once a day; and since he's an indoor-outdoor cat and this is a safe area (the road is quite far away) and the weather is nice, I leave a window open so he's not terminally bored & shutup.
          63
          3
          4.8%
        • ????? @gwern Oct 30 (It just smells a little funny and talks about how it can do kinda cool ImageNet samples too, it's still cool, right...? We need BigGAN Gaiden.)
          116
          6
          5.2%
        • ????? @gwern Oct 29 Hotlinking seems like a design decision to regret, given the ever increasing clampdown on referrers or cross-site anything, not to mention brutally high linkrot rates in general. The only time you can count on something working is the instant the user grabs it. Caching is safer.
          145
          2
          1.4%
        • ????? @gwern Oct 29 DreamBooth embeddings. (OK, I don't actually know how many KB it would be and SD checkpoints can be anywhere from like 4GB to 9GB depending on combos of quantization/optimizer/state/etc so I don't know a finetuned one would be 7GB, but the line is too good that way.)
          563
          18
          3.2%
        • ????? @gwern Oct 29 He's blocked me, but in any case, I disagree that it winds up being a huge advantage or the most important thing about diffusion modeling. Even in RL/planning contexts where iteration is really important, the restriction to a single computation is a painful one.
          145
          8
          5.5%
        • ????? @gwern Oct 29 I like the implication that if he had bought some *real estate* and been rentseeking on it the whole time and that's why he's rich, then that would be a just wealth and not usurious at all. But running or founding Tesla or SpaceX? Heaven forfend. Issue a bull against that ASAP!
          4,584
          147
          3.2%
        • ????? @gwern Oct 29 Marketing or not, I'd rather download 7 kilobytes than 7 gigabytes.
          489
          29
          5.9%
        • ????? @gwern Oct 28 They won't be loling so hard when you innocently ask them for the prompt and then replace them with API calls costing 1000x less.
          3,574
          74
          2.1%
        • ????? @gwern Oct 28 Not sure how that's relevant? Of course non-diffusion models like ThisAnimeDoesNotExist exist and can be CLIP-guided etc.
          76
          0
          0.0%
        • ????? @gwern Oct 27 The loser, who has to be fertilized, I am given to understand. (I wouldn't know first-hand, as I have never lost at penis fencing.)
          135
          20
          14.8%
        • ????? @gwern Oct 27 Fellas, is it gay to go into the pvp bathroom and hide in the stalls just to get a backstab bonus / attack of opportunity on people from behind before the penis fencing proper starts?
          3,603
          98
          2.7%
        • ????? @gwern Oct 27 As a naive cousin from 'cross the Pond, I'm quite intrigued to hear you mean something different!
          289
          7
          2.4%
        • ????? @gwern Oct 27 /shrug Not sure they are going to have much of a choice in the matter, and the longer they wait, the less any of that software will matter.
          97
          8
          8.2%
        • ????? @gwern Oct 27 Rarely. I am still traumatized by the two-faucet sink system, or questioning if my whole life was a lie given unrefrigerated eggs, or how many hours I wasted because I just didn't know you had to dial a number with a leading 0 or whatever... But probably early next year.
          205
          22
          10.7%
        • ????? @gwern Oct 27 When compile hits an unknown link w/no annotation and tries known-sources: Arxiv/Biorxiv/Medrxiv, Crossref-known abstracts by DOI (low hitrate), some sites known by an R library like Pubmed/PLOS, a page which can be scraped for an 'abstract' class div...
          74
          3
          4.1%
        • ????? @gwern Oct 27 Yes, all the summaries are pre-generated. Either by writing, generated at compile-time, or fetched by the client from an API like the Wikipedia API (defaulting to the introduction/first-section). At some point I'd like auto-summaries but haven't quite found an easy+reliable one.
          35
          3
          8.6%
        • ????? @gwern Oct 27 We know synthetic non-natural data like fractal pretraining works surprisingly well (it's a simplicity/Turing-machine prior); is agency also pretrainable from 'synthetic non-agent-generated data' like generative modeling of raw physics? Possibly:
          4,653
          145
          3.1%
        • ????? @gwern Oct 27 I obviously am not going to go into that, and anyway, "change must come from within". I like what I've heard about Tenstorrent but am wary of over-updating on lack of evidence + few legible external signals like Keller joining.
          96
          3
          3.1%
        • ????? @gwern Oct 27 (Since we're playing the VC oracular underrated/overrated gossip game here, it's about decrementing timelines eg 2030¡ú2029.)
          85
          2
          2.4%
        • ????? @gwern Oct 26 Apparently. I haven't tried the non-guest trick yet myself, but none of them have demanded a hotel card or ID. And I did tip at the Venetian but mostly to not feel like an asshole, I could've not paid.
          170
          6
          3.5%
        • ????? @gwern Oct 26 I doubt it's "all" (what is? even WiFi or running water isn't "all"), but the last one was not very nice at all and still did it without a qualm.
          198
          8
          4.0%
        • ????? @gwern Oct 26 (Or to put it another way, this argument also disproves the existence of the Russian silver fox breeding program, which was based on an earlier captive-bred fur population.)
          61
          5
          8.2%
        • ????? @gwern Oct 26 That paper can't show that, because all it shows is a population bottleneck. You need to show that the selection on domesticity-boosting traits happened then. The bottleneck, of course, must happen *before* any implicit selective breeding...
          83
          3
          3.6%
        • ????? @gwern Oct 26 "Luggage storage". Hotels, even ones you aren't staying at, will check your baggage, and you can pick it up on your way out of town. One employee found it funny that I had no idea about this; I told her, "it's not like anyone *tells* you this when you don't know you don't know!"
          233
          25
          10.7%
        • ????? @gwern Oct 26 Everyone always said that Half-Life 3 would only be released when they had a game engine breakthrough to justify it...
          116
          3
          2.6%
        • ????? @gwern Oct 26 19. The AI safety/capability relationship has always been ironic, and will continue to be so. No solution there but luck and speed. 20. Overfitting, reproducibility, hand-tweaking aren't the biggest intellectual fraud in AI¡ªbut the denial of trial & error and serendipity.
          3,997
          86
          2.2%
        • ????? @gwern Oct 26 16. The US continues to snatch defeat from the jaws of victory with visas; UK interesting BATNA... 17. Do you do research & need compute? Give TRC a shot, seriously, it's <5min signup form & TPus have gotten a lot easier to use w/PyTorch. 18. There's plenty of room at the top.
          5,616
          147
          2.6%
        • ????? @gwern Oct 26 11. AI investment numbers are at least a bit misleading because so much of it is zero-sum compensation inflation 12. People still can't ????? like big models. 13. Stability GPUs go brrr 14. Genetics is great & all, but still too slow 15. Every year SF is shabbier
          3,553
          94
          2.6%
        • ????? @gwern Oct 26 7. gain-of-lambda-function research continues apace 8. if Musk isn't at least a little bipolar, I'll eat my hat 9. "What does everyone know that I don't?" Thus far for music/GANs/chip sanctions/emergence/¨€¨€¨€¨€-¨€¨€¨€¨€¨€¨€¨€: nothing good. 10. Groq is doomed.
          4,263
          210
          4.9%
        • ????? @gwern Oct 26 4. self-driving cars+, Waymo LLC? 5. My travel life got abruptly better when I learned "baggage check" included non-airplane things. 6. the future is even more unevenly distributed than usual¡ªeven in their subfield, most people are way behind on their reading
          4,353
          139
          3.2%
        • ????? @gwern Oct 26 Learnings: 1. it may not be possible to genericize nouns by pluralizing without sounding like a douches 2. AI timelines ?1y if you are not named 'Gabe' (also: prediction market & timeline discussions??) 3. the avalanche is caused by both the snowfall ??? the last snowflake
          2,945
          147
          5.0%
        • ????? @gwern Oct 26 You're either cool enough to have a job that will send you or cool enough to not have a job, sorry, I don't make the rules.
          12,301
          1,043
          8.5%
        • ????? @gwern Oct 26 Spend 24 hours flying to the land of dropbears and dropspiders? ? How about 'never'? I can pencil that in now.
          222
          6
          2.7%
        • ????? @gwern Oct 25 Change of plans: Biergarten apparently is closed so I will meet everyone there and we can go to the little park next to it since the weather is nice.
          5,714
          133
          2.3%
        • ????? @gwern Oct 25 Do *you* want to hear execs like Cook or Ballmer go on stage and shout about 'blobs! Huge tracts of blobs! We have the biggest blobs in the world, to blob your business into the blob new age!' Hm. Maybe that wouldn't be so bad, actually...
          156
          5
          3.2%
        • ????? @gwern Oct 25 Hm, I heard it several times on my trip so far. Seems to be making a fight of it.
          216
          6
          2.8%
        • ????? @gwern Oct 25 You say that, but you're not the one who has to try to explain at parties that you write about chonkies and trying to understand how chonking too hard could lead to absolute units madladding humanity into can't-evening and everyone should realize it's srs bsns.
          627
          91
          14.5%
        • ????? @gwern Oct 25 (Look, I'm not saying terminology merits all the bikeshedding it inevitably receives; just that your life is now quantifiably better because I went with "the scaling hypothesis" instead of the alternatives like "the big blob of compute" or "one chonky is all you need".)
          8,647
          368
          4.3%
        • ????? @gwern Oct 25 (But is DT bad at generalization in the way that all Transformers were bad at generalization until they weren't, or in a different way?)
          100
          8
          8.0%
        • ????? @gwern Oct 23 Why does that prove too much? Isn't that exactly how these models became of interest with all their capabilities, and became worth either RL or data finetuning further?
          207
          6
          2.9%
        • ????? @gwern Oct 23 (I don't need to hand you my driver's license to go to a cafe and drink tea and chat, you know.)
          397
          28
          7.1%
        • ????? @gwern Oct 20 It can only copy the human raters with high accuracy by learning what the human raters want, creating a capability which can be prompted simply by the 'human rating'. No separate model, additional head, classification loss, PPO pain, KL constraints, degeneracy, multiple phases...
          187
          7
          3.7%
        • ????? @gwern Oct 20 My point there is that evaluation *is* generation. Whatever evaluation raters generate, that's just more text to generate. I don't find it obvious that 'handing that generated text to a surrogate reward model to then blackbox optimize the generative model' >> 'more generative'.
          163
          7
          4.3%
        • ????? @gwern Oct 20 DT can condition on desired performance, including perf never observed, so it is not simply imitating or stitching together expert trajectories. Also, it's not like InstructGPT is 'superhuman' so that's not even a requirement here.
          124
          4
          3.2%
        • ????? @gwern Oct 20 I don't buy that here because there is no principled difference in the generative setting, eg Decision Transformer. Use of rankings or comparison is just more data to predict; anyway, lots of InstructGPT examples are clearly '1 right answer' and classic supervised learning.
          179
          13
          7.3%
        • ????? @gwern Oct 20 It could, but why on earth would they have dropout or similar mechanisms activated? They have not historically screwed up the API like that, and checking that you return the same answers for the same input+model is one of the first tests anyone would set up for the API.
          77
          1
          1.3%
        • ????? @gwern Oct 20 Not sure about. The former has a lot of practical advantages and is scalable, smoothly turning into Decision Transformer. While no one wants to mess with PPO, and RLHF comes with its own deep problems like completions collapsing to dull single possibilities or adversarial optim.
          2,352
          86
          3.7%
        • ????? @gwern Oct 20 Deciding to solely make jokes of the 'haha only serious' form definitely does not bar trolling.
          485
          15
          3.1%
        • ????? @gwern Oct 20 What is this, a suggestion for an ant researcher? I'd need, like, at least 8 times more free time to replicate that!
          79
          4
          5.1%
        • ????? @gwern Oct 20 A point that would've made sense in 2012 when people were amazed by the obvious success of the marijuana campaign... but forget to update that it's now a decade later, and it's absurd. No one could imagine success¡ªafter *11* states had medical marijuana?
          8,947
          62
          0.7%
        • ????? @gwern Oct 20 Ah, but is Ilya tweeting about AGI because OA work is going so well it's urgent to prepare the ground, or because it's going poorly and he's bored and rather enjoys trolling everyone on Twitter, or neither and he just enjoys trolling? ? Admit it, you might do the same thing too.
          3,853
          115
          3.0%
        • ????? @gwern Oct 20 (As much as I would love something as wild as 'ants have a quasi-language', that's almost as far into "extraordinary claims require extraordinary evidence" territory as 'ants passing the mirror-test'. (Maybe further? Hm...))
          89
          5
          5.6%
        • ????? @gwern Oct 20 Yes, that's one of the things I am thinking of. All of the travel & anthropology accounts of non-First-Worlders having much better smell skills, and then an experimental demonstration that getting better isn't even *that* hard, we First Worlders just don't do even that much.
          206
          14
          6.8%
        • ????? @gwern Oct 20 I remembered that a bunch of states had already legalized medical marijuana in 2002 and so Agrawal's claim was absurd; I did have to look up the exact number of ~11, yes, but that's what keyboard shortcuts & WP are for.
          272
          18
          6.6%
        • ????? @gwern Oct 20 (oh no so the availability was better than that of video game consoles in 2020, however could anyone have survived such tragic first world problems or have thought that marijuana legalization was winning, it is a mystery)
          129
          9
          7.0%
        • ????? @gwern Oct 20 Our olfaction is apparently way worse. Everything is scented/perfumed, sanitized, food is safe/fresh/preserved, bodies washed, etc.
          3,807
          112
          2.9%
        • ????? @gwern Oct 20 That sounds kinda useless. Wouldn't people rather just use the tags directly instead of groping around with prompts?
          37
          1
          2.7%
        • ????? @gwern Oct 20 I don't know if they explicitly do it, but I wouldn't be surprised if that's implicit. It's often cited in leaky or metered paywalls, and definitely more than 1 econ modeling paper has treated adblocking that way. (Tho I don't read them too carefully because generally no data.)
          42
          2
          4.8%
        • ????? @gwern Oct 20 And Canute ordered the tide back, yes. Resorting to sending goons in to medical clinics is what 'losing' looks like. A winner would never be forced to such a resort, having won long before.
          166
          10
          6.0%
        • ????? @gwern Oct 19 Reading about the sheer totalitarian brutality of the American WWI treatment of German-Americans and any kind of dissent, the level of censorship and jingoism, is extraordinary. The US wasn't even in any danger! Nothing during eg the Iraq War or COVID came close.
          2,151
          76
          3.5%
        • ????? @gwern Oct 19 People were speculating the price cuts were due to quantization or something on the backend. If they heavily quantized it, then GPU nondeterminism might be greatly amplified?
          163
          14
          8.6%
        • ????? @gwern Oct 19 But yeah, for most practical purposes, what a critic might deem 'memorization' is all that is necessary. People who want to generate anime images for art or games etc probably don't care much about prompts like 'a horse riding an astronaut' - they want astronauts riding horses.
          71
          4
          5.6%
        • ????? @gwern Oct 19 I expect larger models (esp larger text models) to resolve it. If you don't want to pay the scaling tax, you can try to hand-engineer it in. Synthetic datasets to encode relationships into images can probably push much smaller models into eliciting the desired reasoning skills.
          72
          7
          9.7%
        • ????? @gwern Oct 19 The papers explain how causal attribution can go wrong and simply cannibalize or steal credit or yield spurious results. These problems are admitted by ad sellers who say they've fixed them... tacitly admitting the old numbers were bogus and fooling everyone - so why believe now?
          63
          4
          6.3%
        • ????? @gwern Oct 19 Yes, but you both *do* use adblocking and watch YT... That's price discrimination/self-selection in action. And Google has done some of the research on ad avoidance, so presumably YT has not forgotten it and found that those ads are at an optimal misery point.
          17
          1
          5.9%
        • ????? @gwern Oct 19 P&G is one of the sources for the ad critiques, and in any case, considering the extreme variance, I wouldn't expect to see much of a broad correlation. The effect size cannot be that large.
          41
          1
          2.4%
        • ????? @gwern Oct 19 (That feels a little creepy and insulting, tbh. I liked Ross, even if he tried to do some terrible things and screwed over a lot of people with his mistakes.)
          312
          12
          3.8%
        • ????? @gwern Oct 19 Uninformative because massively confounded, and the split experiments they briefly review are little better.
          35
          2
          5.7%
        • ????? @gwern Oct 19 The scenes aren't very weird, so he would say that's just the memorization at work. (Tags don't encode arbitrary relationships well, they are bag-of-words or pre-specified relationships, so it'd be hard to use the hard-image benchmarks or his challenge examples.)
          53
          8
          15.1%
        • ????? @gwern Oct 19 One thing I have learned the past two years is that bees are smarter than I realize. Not just play, but stuff like counting. (I am still waiting for some hostile replications of the ant mirror-test before I start taking *that* too seriously, however.)
          1,647
          58
          3.5%
        • ????? @gwern Oct 19 In 2002, 11 US states had legalized medical marijuana.
          5,451
          198
          3.6%
        • ????? @gwern Oct 19 (Softmax is just the poor-man/Bayesian's Thompson sampling. ?)
          2,441
          78
          3.2%
        • ????? @gwern Oct 19 If you want to make the diffusion more sensitive to higher-order like curvature, could you build that in w/o 2nd order by a pixelwise loss for the destination pixel, similar to ? So at every denoising instance,the model is pulled towards next denoised+true.
          37
          1
          2.7%
        • ????? @gwern Oct 19 Note that with new methods like DreamBooth or retrieval methods on Stable Diffusion-class and better models, transfer learning may require <10 samples, and may be doable simply by prompting various keywords or similar author names to find the target artist/style 'in between' them
          62
          3
          4.8%
        • ????? @gwern Oct 19 Depends. Have you heard of the 'dodo bird verdict'? If not, I think a lot of people would like to make an appropriately-phrased bet with you...
          179
          12
          6.7%
        • ????? @gwern Oct 19 Or possibly something morally equivalent to a lr drop - maybe a warmup hyperparam or a moving average somewhere. (I like 'emergence' as much as anyone, but that the curves look less wiggly is pretty suspicious. Don't think things like the induction bump usually look like that.)
          162
          3
          1.9%
        • ????? @gwern Oct 18 (The screaming boys had also bought all the records.)
          158
          4
          2.5%
        • ????? @gwern Oct 18 ? Not sure how lucrative ads to teen girls would've been - after all, they already were buying all the Beatles' stuff. And I know network advertising in that era was typically pre-sold in large blocks in advance, so probably the advertiser neither knew nor cared.
          136
          5
          3.7%
        • ????? @gwern Oct 18 You can argue if they hate ads as much as they claim (I don't think they do, I was shocked the effects are even as large as 10 or 20%), but they do clearly have revealed preferences against ads per se.
          1,843
          38
          2.1%
        • ????? @gwern Oct 18 'Targeting'? On the Ed Sullivan show? One of the most famously universally watched pieces of Americana, for the whole family? Not quite sure what you have in mind there. Those seem about as targeted as possible to be for such an audience.
          118
          3
          2.5%
        • ????? @gwern Oct 18 Bad statistics, universally shared folk psychology intuitions, minimal evolutionary pressure, and especially internal principal-agent problems of the sort you've written about many times eg the excess of mergers.
          2,852
          169
          5.9%
        • ????? @gwern Oct 18 Sounds very wholesome. Everyone needs some flour to bake with, and clean clothes. Would you have rather had instead ads on how 9 of 10 doctors prefer Lucky Strikes smokes?
          333
          14
          4.2%
        • ????? @gwern Oct 18 Yes! John Backus took me there a while ago. It was very impressive. (Albeit in a "look at all the things you can't use or play with because you're just visiting" sorta way...)
          952
          32
          3.4%
        • ????? @gwern Oct 18 In randomized experiments of ads, people *do* stop watching entirely, and the buying effects are also extremely small. I've linked you both literatures on this.
          9,119
          421
          4.6%
        • ????? @gwern Oct 18 I'll be in SF next Sunday afternoon (2022-10-23) & probably Monday, maybe Tuesday, if anyone wants to hang out. Suggestions for new/unusual things to do also welcome.
          29,558
          1,129
          3.8%
        • ????? @gwern Oct 18 (FWIW, please don't use that name for your future work. Aydao and Tensorfork worked very hard to make and it was pretty impressive work at the time, nor was it obscure.)
          255
          38
          14.9%
        • ????? @gwern Oct 18 That would be a slamdunk analogy if not for risk homeostasis.
          2,265
          40
          1.8%
        • ????? @gwern Oct 17 'Placebo algorithms'? I suppose a lot of things could look like that. 'Money', for example.
          561
          17
          3.0%
        • ????? @gwern Oct 17 (I feel a little schadenfreude that even professionally-edited Google PR months in the making still cannot correctly spell "LaMDA", thereby suffering my pain. Don't call it a grave typo¡ªit's the future you chose.)
          2,010
          35
          1.7%
        • ????? @gwern Oct 17 (Neither a squirrel Bar Mitzvah nor a squirrel Bat Mitzvah, but a squirrel third thing...)
          588
          21
          3.6%
        • ????? @gwern Oct 17 You should tweet some fake prompts at some point, just to see who has, like me, given up distinguishing impossible from possible prompts.
          1,846
          54
          2.9%
        • ????? @gwern Oct 17 Indeed, but how? The dirty secret of dating apps seems to have always been that they have little idea who will click (), and function mostly to get people to go on dates at all. So they are vital, like oxygen, but not having any clear effects...
          912
          56
          6.1%
        • ????? @gwern Oct 17 This is a good question to pose people who think self-attention/Transformers are magic pixie dust and alternatives like MLPs couldn't beat them: if Transformers like 175b are ??????? almost entirely MLPs by FLOPS, why are you certain the self-attention is irreplaceable?
          193
          10
          5.2%
        • ????? @gwern Oct 17 (I would also note that even if this is 100% true at face-value with no caveats, it's still not an answer to the question of 'is Jewish elite dominance decreasing', as there is more to life than academia, and elite patterns have changed a lot eg tech.)
          227
          13
          5.7%
        • ????? @gwern Oct 17 I don't see how that gets you a five-fold collapse in 30 years (ie. <1 academic-generation). Other things sound more likely: changes in response patterns (like the supposed dectupling or whatever of LGBT over the same period), hiring patterns, sampling frame...
          598
          48
          8.0%
        • ????? @gwern Oct 17 After 51 CPU-core-days on my Threadripper, ? = 133,787,000 turns out to be $6,038.22 vs the ¡Ì? approximation of $6038.78162 for a difference of $0.56 or a percentage error of 0.009%. Yeah, I'm going to say that it is a square root and no need to try to calculate further.
          199
          5
          2.5%
        • ????? @gwern Oct 16 If you have more faith in 'peer review' than most of the people who do it, perhaps.
          133
          8
          6.0%
        • ????? @gwern Oct 16 (Also worth remembering that "attacks only get better"/"sampling can show the presence of knowledge not the absence": capabilities never go away or get worse, and whatever the *current* quality of samples, people will be learning how to do better ones for a long time to come...)
          3,293
          52
          1.6%
        • ????? @gwern Oct 16 Just pull'em out of IA and host them yourself. They're not going to care.
          711
          19
          2.7%
        • ????? @gwern Oct 16 It's been 13 days since launch and Pixiv is reporting >36k 'works', which can have dozens of images in each, so easily >100k images? Capital is a complement to labor until it becomes a substitute. It's quite something to see all that GPU capital go brrrr once past the threshold.
          7,516
          328
          4.4%
        • ????? @gwern Oct 16 Probably because "no drug works for Alzheimer's" is not news, and quite depressing.
          3,951
          162
          4.1%
        • ????? @gwern Oct 16 It wasn't, but it does feel tempting to leave it there or comment on it...
          106
          6
          5.7%
        • ????? @gwern Oct 16 The current diffusion hierarchical approaches would lend themselves well to iterative AI edits: a tree of frames, and you just edit a frame, clamp it, re-diffuse parent frames, and diffuse back through the rest of the movie, propagating edits, until the whole thing stabilizes.
          74
          6
          8.1%
        • ????? @gwern Oct 16 A probability like 2% across an entire phenome means larger deviations are guaranteed. (How many thousands of variables does UKBB measure? And the # of human diseases runs as many thousands as you care to define) Plus, you need to calculate the tail risk of event, not percentile.
          66
          18
          27.3%
        • ????? @gwern Oct 16 Thoughts on how 'good design' pays off: One example: Eliezer Yudkowsky gets a lot of criticism, good & bad, unsurprisingly. But surprisingly, many people misspell it with 'i' (despite big visual difference)... and yep, those are invariably bad criticisms.
          3,138
          211
          6.7%
        • ????? @gwern Oct 15 Ah, that makes more sense. I'm not *too* surprised by that because IIRC even the original residual paper observes that residual layers are often additive and can be dropped randomly and don't form a clean low-level -> high-level hierarchy like one likes to imagine.
          102
          3
          2.9%
        • ????? @gwern Oct 15 Mildly confused - what does it mean to permute the layers of a U-net?
          91
          1
          1.1%
        • ????? @gwern Oct 15 There is no standard dollar. I'm not sure even IMF drawing rights can be considered constant. ?
          572
          11
          1.9%
        • ????? @gwern Oct 15 Think of the consequences of a critical node, already teetering over bankruptcy, when a bunch of employees leave literally overnight, orders are canceled, exports/imports abruptly impossible. They shut down. Now everyone who depended on them suddenly can't produce anything...
          54
          3
          5.6%
        • ????? @gwern Oct 15 From a systems perspective, the timing isn't going to get better later, and the economic crisis is a force-multiplier: chip is an ecosystem, and if you can attack it hard enough at a low point, you may bring the entire thing down, force complete liquidation and permanent exodus.
          78
          3
          3.8%
        • ????? @gwern Oct 15 Is that more or less mindblowing than that StyleGAN was so linear that it could swap layers?
          5,496
          153
          2.8%
        • ????? @gwern Oct 15 I've really always thought of you as more of a scifluencer and thinkmancer.
          128
          7
          5.5%
        • ????? @gwern Oct 15 (It's a little funny thinking back to then and spending months on GANs and making Danbooru20xx, and telling people, "just you wait! in 5-10 years, we'll be making human-level anime images & videos with this stuff!" "So, how well does it work now?" "er well it'll get better ok.")
          113
          12
          10.6%
        • ????? @gwern Oct 15 Ah, this takes me back. Below left, December 2015, Soumith Chintala's DCGAN-torch, Asuka-anime-faces; below right, February 2017, original WGAN-GP. (Exponentials are a helluva drug.)
          2,111
          137
          6.5%
        • ????? @gwern Oct 15 Sounds susceptible to shortcut solutions. And anyway, we didn't evolve against machine cheating detection.
          383
          10
          2.6%
        • ????? @gwern Oct 15 Hm, a lot of potential here for using blackbox methods instead: define a bunch of SVG vector art gears and then optimize their locations and parameters for the face and subsequent animation, instead of smooshy pixel-level gears.
          1,687
          30
          1.8%
        • ????? @gwern Oct 15 Again, you simply wait for a fluctuation which provides a vacuum around it, which may be much more unlikely, but again, happens indefinitely often (and you won't observe the crashed-in brains at all if they don't exist long enough for observer-moments).
          48
          3
          6.2%
        • ????? @gwern Oct 15 (And then you can run anime neural net models on it!)
          38
          1
          2.6%
        • ????? @gwern Oct 15 My understanding is that brains can survive a second or two conscious without any oxygen/blood-flow, so in a fluctuation large enough to give rise to a brain surrounded by a vacuum (which will happen indefinitely many times), a brain would be conscious for ~1.5s.
          73
          6
          8.2%
        • ????? @gwern Oct 15 I think the PR by emad & SD about thinking of the models as 'search engines', while true in an abstract sense like Borges's library, is catastrophically bad PR and probably responsible for why a lot of laymen think it's just literally searching & copying images - he says so!
          37
          5
          13.5%
        • ????? @gwern Oct 15 No, because 'echo' is dumb. It's not even a hack, it's working exactly as both designed & intended. There are a bazillion more interesting hacks, like 4chan manipulating votes on a poll so the initials spell out a slur, or something.
          123
          11
          8.9%
        • ????? @gwern Oct 15 No, they don't love cat earwax nearly as much as they love human earwax. Best theory I have so far is some overlap with valerian:
          25
          2
          8.0%
        • ????? @gwern Oct 14 (That is, the AI safety challenge is not to imagine in 2022 how inner-monologue could be dangerous, but to imagine it in 2019.)
          65
          4
          6.2%
        • ????? @gwern Oct 14 I wouldn't go quite that far. It was a tradeoff, definitely. But for future scale-ups, Stability should probably be more careful, and less restrictive, about their filtering.
          74
          6
          8.1%
        • ????? @gwern Oct 14 (I think the quality jumps of Waifu Diffusion & NovelAI, and the blind spots in SD like Pokemon, vindicate this view: SD is not *so* good that it only needs to locate characters. It has a lot of learning left to do.)
          65
          3
          4.6%
        • ????? @gwern Oct 14 Yeah, it's been frustrating watching Tay just enter the It Is Known stage of AI ethics, but also it is now so mythologized (on the basis of some throwaway media reporting) that based on my tanks work, most people will never believe you even if you get the original programmers.
          728
          26
          3.6%
        • ????? @gwern Oct 14 (Supposedly JPEG does support CIE LAB, but I wasn't able to get ImageMagick to spit out a JPEG which it reported to use LAB rather than sRGB, and figuring out what's going wrong there is beyond me.)
          157
          5
          3.2%
        • ????? @gwern Oct 14 Impressive how there's NovelAI samples everywhere now. I've been checking in on AIBooru (NSFW) every few days and it's up to 3k images now and increasing in quality/diversity as people explore prompts. And it's still objectively a pretty bad model!
          47
          1
          2.1%
        • ????? @gwern Oct 14 (Yeah, I deleted them all and set up some rewrites. Abstracts look substantially more readable now that there's less techno-babble hype.)
          34
          1
          2.9%
        • ????? @gwern Oct 14 You wouldn't mistake them for real photos. Makes me wonder if conditioning is a crutch. Perhaps training should anneal from conditional to unconditional? Better latent space, more generalization, handle weirder prompts...
          61
          6
          9.8%
        • ????? @gwern Oct 14 In the future, everyone will be a Boltzmann brain for 1.5s.
          7,834
          80
          1.0%
        • ????? @gwern Oct 14 Probably impossible, given that even at GMU, Hanson (PhD, publications) only got tenure because, as he describes it, the other faculty were pissed about the media sliming the 'terrorism futures' project (which we know from later IARPA work would've been useful).
          36
          0
          0.0%
        • ????? @gwern Oct 13 Well, that's what you would ablate by deleting each one and looking at a set of samples, and seeing how minimal (and thus, educational & reusable) a prompt you can get. Instead of cargo-culting 'greg rutkowski masterpiece trending' everywhere.
          280
          23
          8.2%
        • ????? @gwern Oct 13 Yep. 'I worked through it line by line and there was some non-zero probability it would answer literally any question outside the one line by line dialogue, you couldn't actually move blocks lol' is... not how it was described in any of my textbooks or Hofstadter's GEB, for sure.
          109
          18
          16.5%
        • ????? @gwern Oct 13 Seems optimistic to assume that SHRDLU matched the SHRDLU system: (I too can write SHRDLU if I only have to make it give the answer to a single precanned script for a demo never run again; just make a series of 'echo' calls...)
          2,654
          50
          1.9%
        • ????? @gwern Oct 13 (For any prompt this long, I think it should be ablated because probably at least several of those phrases aren't doing anything.)
          5,206
          170
          3.3%
        • ????? @gwern Oct 13 (Very cute, grasshopper, but you realize Takahata was criticizing the antithesis of Buddhism just as much as the thesis of the aristocracy, right?)
          121
          13
          10.7%
        • ????? @gwern Oct 13 Very common. Unfortunately, with rare books, much less alpha because individuals can't just chuck them into a debind-then-sheet-fed-scanner-to-Libgen pipeline for relatively low cost. Have to leave that stuff to organized efforts like IA/GB with proper camera-scan setups.
          62
          9
          14.5%
        • ????? @gwern Oct 12 Nat... are ??? a promptie? I'm not sure I could still be friends with a promptie.
          1,765
          103
          5.8%
        • ????? @gwern Oct 12 (All I ask is to be remembered as a humanitarian and cat-lover.)
          168
          14
          8.3%
        • ????? @gwern Oct 12 First Pokemon, now geese? We must declare war on the esthetic filter because clearly peace was never an option.
          157
          16
          10.2%
        • ????? @gwern Oct 12 T-I/DB already failed on Canadian geese? OK, that's surprising. I wonder what happened there? The esthetic filter really hates geese?
          227
          11
          4.8%
        • ????? @gwern Oct 12 What's wrong with the other two? You think SD doesn't know what geese are?
          162
          7
          4.3%
        • ????? @gwern Oct 12 I suggest "promptie", by analogy to "skiddie". People who can only ask the AI for stuff because they have no independent skills to program or edit. "That Jack is such a promptie, he can't even fix that hand or finetune SD without the Colab."
          7,191
          255
          3.5%
        • ????? @gwern Oct 12 I am way behind on inner-monologue literature, but do we have self-critique/refutation yet, like prompting it for explanations of why an erroneous answer is wrong & the right one? Testing with InstructGPT, I got 2?3 working, so PaLM might be in the convincingly-working regime...
          3,459
          115
          3.3%
        • ????? @gwern Oct 12 Remarkably, Khoth & FeepingCreature kept on with the C versions & got it down to 0.5s (!) to calculate the exact value of ?=100,000 cards. (My ?=133,787,000 run is still going.) This lets us confirm the random-walk conjecture that its value is pretty much exactly = 0.5¡Á¡Ì?.
          379
          8
          2.1%
        • ????? @gwern Oct 11 I have no idea. Algo just popped up her tweet and I was fascinated by her and her many fav/rts' belief that there was literally no career option for a Western women other than being a housewife in 1972, and how that seemed like this sort of logarithmic compression of history.
          129
          5
          3.9%
        • ????? @gwern Oct 11 (OK but seriously, what makes Musk Musk? I found him easier to understand when I started to think of him as¡ªlike many entrepreneurs¡ªan untreated-because-so-high-functioning manic-depressive, complete with world-peace plans during manic phases. So that, or moral equivalent.)
          611
          38
          6.2%
        • ????? @gwern Oct 11 The 3 hacker genders: 1. "Can I run ???? on it?" 2. "Can I run P2P on it?" 3. "Can I run NetBSD^WLinux on it?"
          6,700
          72
          1.1%
        • ????? @gwern Oct 11 ?50ya from the '90s is a good comparison! ?50ya from 2022, however, to 1972 (and incrementing) is not; eg RBG was already on her (first) tenured law professorship at that point. ?30ya likewise: from '90s, to '60s, lots of useful comparisons. 2022 to... 1992? Not so much.
          5,453
          77
          1.4%
        • ????? @gwern Oct 11 That is an almost maximally non-random selection (and that's why it's so useful, of course).
          43
          1
          2.3%
        • ????? @gwern Oct 11 Thanks. My first reaction is: these are a lot worse than conditional. You set text guidance to 0, you didn't just use a "" prompt? Any idea why they are so bad? I know that guidance improves the quality but I didn't think the difference would be *this* large.
          217
          6
          2.8%
        • ????? @gwern Oct 11 Yes, one would expect a commercial chess site to be very reluctant to accuse professionals playing in real-money tournaments of cheating & expel them. Their bias is towards sticking their heads in the sand, not to accusing at the drop of a hat!
          74
          4
          5.4%
        • ????? @gwern Oct 11 Meh. That's not novel: people were doing that with AI Dungeon 2 Dragon back in like July 2020 (discovering inner-monologue that way, incidentally, long before the researchers reinvented it), so well over 2 years ago.
          100
          5
          5.0%
        • ????? @gwern Oct 11 Randomness is pretty hard; I know in MCMC stuff, it can take a lot of random steps to 'wash out' any initial bias or starting point. You also have no idea what the latent space connectivity or dynamics are once you stop sampling from the training/generative process.
          80
          4
          5.0%
        • ????? @gwern Oct 10 Really part of the same hack. Also not sure how much people care about that? I haven't followed NovelAI text stuff in detail, but didn't seem astounding and seemed more or less a 'AID2 if they hadn't screwed it up'?
          137
          0
          0.0%
        • ????? @gwern Oct 10 Hm. So there's a interesting larger model leak you just can't run? What's that?
          163
          1
          0.6%
        • ????? @gwern Oct 10 Come to think of it, is this the first hacked/leaked neural net model that anyone really cares about using? SD 'leaked' a week or two early but that hardly counts; the AID2 incidents didn't leak the GPT-3 models; nothing else comes to mind.
          6,490
          157
          2.4%
        • ????? @gwern Oct 10 The latent diffusion people are rederiving _z_ and the progressive distillation are reinventing progressive growing, so I assume next up is data-augmentation diffusion and auxiliary critics. ? But since time is a flat circle, hopefully diffusions will be much better this time.
          3,055
          23
          0.8%
        • ????? @gwern Oct 10 Fair use is a defense against copyright infringement; if it's transformative, then there's no infringement you need any defense for. (Simply affecting market prospects is not definitive - copyright is not a UBI or sinecure. Artists aren't owed jobs, if there is a competitor.)
          95
          2
          2.1%
        • ????? @gwern Oct 10 So... what are the learned prompts? They should be text and readable, and presumably very interesting to look at and modify.
          2,538
          37
          1.5%
        • ????? @gwern Oct 10 I saw nothing about using an offline WP dump or anything like that. By WP API, you mean PaLM autonomously accesses the live MediaWiki APIs over the Internet?
          89
          20
          22.5%
        • ????? @gwern Oct 10 It's a good thing to know, but unfortunately, it's obscure so it'll cause you various problems. For example, until recently, the W3C HTML checker would warn you if you linked to #?top because it thought that anchor was undefined.
          2,332
          12
          0.5%
        • ????? @gwern Oct 10 (I also doubt that it would have the benefits you expect on AI capabilities, for the reasons we discussed before: much of the relevant data would not be copyrighted or would be already licensed, assuming it was not just ignored entirely.)
          1,437
          28
          1.9%
        • ????? @gwern Oct 10 I mean, if you can get a law through US copyright narrowing transformative use and redefining AI training to be copying, then yeah sure, I will agree that they are then legally equivalent and thus morally equivalent in that respect. I don't think that's going to happen, though.
          2,802
          85
          3.0%
        • ????? @gwern Oct 10 A very loose lowerbound of several percent for something that should be, and was believed to be, ~0%, and the rather casual way in which this information came out long afterwards, is quite alarming. Would you adopt the same blase attitude towards, say, DNA/wrongful executions?
          77
          3
          3.9%
        • ????? @gwern Oct 10 A legal difference is always a moral difference, because higher-order effects are important, and that difference is always relevant when someone is claiming they are identical (ie. have no differences at all).
          3,239
          90
          2.8%
        • ????? @gwern Oct 10 Tragically, no; and it doesn't involve self-attention *or* Decepticons either! Language is such a gyp.
          242
          10
          4.1%
        • ????? @gwern Oct 10 Yes... One potential problem I see is finetuned models: I don't know if *any* of the finetuners have made a point of dropping out the embedding unless they copied code which did that for them. So this might be true of SD but then finetuned models lose their unconditionality... ?
          69
          6
          8.7%
        • ????? @gwern Oct 10 (lman, you think you're being very clever and insightful, but you're failing at the first step when you specify that it is an exact replacement and it's just a derivative work from then on out.)
          416
          17
          4.1%
        • ????? @gwern Oct 10 No, because that's where the lamppost is. would not expect to detect cheating offline. If they cheat that much where they can be detected...
          214
          8
          3.7%
        • ????? @gwern Oct 10 (He's busy and probably wouldn't know a detail like that.) I asked and she says SD *was* trained with dropout, where 'zeroed out' is the CLIP embedding of "", so sampling "" *should* be equivalent to random unconditional samples from the full SD distribution.
          64
          5
          7.8%
        • ????? @gwern Oct 10 I think this might depend on what SD did for the embedding and/or classifier guidance. Did they drop it out during training as some do? If so, then I *think* zeroing it out and then diffusing would in fact be supposed to be a pure random sample from the generative distribution.
          53
          1
          1.9%
        • ????? @gwern Oct 10 Seems to me what counts as 'transformative' in copyright law is defined by copyright law.
          517
          19
          3.7%
        • ????? @gwern Oct 10 I dunno. How random *is* that? I've seen plenty of so-called 'random samples' go wrong and not be all that random. (Like our anime BigGAN was being sampled completely wrong and we didn't figure it out for a long time.)
          57
          1
          1.8%
        • ????? @gwern Oct 10 I mean... what's there to discuss? is fairly clear. SD uses an image #1234; it does not output #1234 unless you make it do so and does not replace original, and the training does indeed "transform a particular article into a different state or thing".
          432
          29
          6.7%
        • ????? @gwern Oct 10 Hm, well, I'd expect two vertical regions for RNN small-compute Transformer large-compute at small n for RNN (inductive bias eventually crippled by history penalty), passing to CNN-below-Transformer for the rest, giving a 2x2 look. What's your intuition how it'd look?
          47
          1
          2.1%
        • ????? @gwern Oct 10 That would start every sample at the same point (right?) and latent space is big, so I would be skeptical such a biased sample would after some kicking give me the equivalent of sampling a multivariate gaussian for a GAN's _z_.
          128
          2
          1.6%
        • ????? @gwern Oct 10 Q. Does anyone have a batch of random samples from Stable Diffusion I can browse? I mean completely random, just random samples from its distribution¡ªnot variations on a prompt or a random walk from a prompt or multiple decodes from a prompt, which is all everyone seems to have?
          7,594
          97
          1.3%
        • ????? @gwern Oct 10 (An increasingly awkward comparison after Magnus was beaten by a cheater and just told everyone last week that detecting cheating is routine among the top 100, and they are all let back in.)
          6,004
          176
          2.9%
        • ????? @gwern Oct 10 Source code copying is not a transformative use.
          12,532
          358
          2.9%
        • ????? @gwern Oct 10 I bet it's to make handling key-chords easier. If it comes with a canonical sort, then the API user can simply write a sorted literal to match it instead of messing with set equality or sorting. `If x == [Control, Command, V]`, instead of every permutation of CCV emitted.
          44
          0
          0.0%
        • ????? @gwern Oct 10 I disagree. Such legislation and changes do not drop out of the sky, and the changes precede the legislation: groups gain power first. When they do drop out of the sky like the endless lists of rights enumerated in the USSR constitution & other socialist states, they are useless.
          47
          8
          17.0%
        • ????? @gwern Oct 10 Yes, that's what I meant. You're gonna have to bite this bullet at some point or those texture artifacts are never going away. If the VAE encoding is suboptimal, the model can't fix it because it can only shuffle around and rearrange the same suboptimal 'pixels'.
          55
          4
          7.3%
        • ????? @gwern Oct 10 The discussion by the author, and all the women quoted, indicates it was routine, common, unremarkable, and not special. You would not be surprised or offended if you were only 1 of 10k privileged women out of >100,000k women and ran athwart default assumptions; so, they weren't.
          40
          4
          10.0%
        • ????? @gwern Oct 10 Yes, but that still doesn't explain the paradox itself because you might just assume that in very misogynist societies, the girls wouldn't be allowed to go to college at all, as indeed was true for most of history.
          92
          4
          4.3%
        • ????? @gwern Oct 10 On the contrary, it makes clear repeatedly that many women had their own CC & there was no ban. I don't know how you can read its many vignettes about eg women being annoyed by being expected to give up their single CCs for joint CCs on marriage and conclude 'no women had CCs'.
          44
          7
          15.9%
        • ????? @gwern Oct 10 If it was American culture, the gender egalitarian paradox wouldn't be a general phenomenon applicable even more strongly to eg the Scandis.
          123
          5
          4.1%
        • ????? @gwern Oct 10 Yeah, the inverse correlation has drawn a lot of commentary. The usual explanation seems to be that STEM degrees serve as a kind of "mrs. degree" in those countries because a 'real' STEM degree is socially acceptable for women to go get; thus, inequality drives STEM equality.
          294
          8
          2.7%
        • ????? @gwern Oct 10 And obviously Greenberg was referring to it because it was 2 years later, the only thing that comes up when people discuss it, and most people operate under a "if it's not mandatory, it's forbidden" framework where that's an ordinary mistake; nor does he deny it.
          133
          6
          4.5%
        • ????? @gwern Oct 10 Look it up. First, common sense: why would ban *discrimination* if reality was *no* women had credit cards? You don't 'ban discrimination' against marijuana, you 'legalize' marijuana. Second, lots of coverage at the drop of a google eg "Credit-Ability for Women", Garrisopn 1976.
          152
          6
          3.9%
        • ????? @gwern Oct 10 Yes, something like that... "Immortal time bias", perhaps?
          30
          1
          3.3%
        • ????? @gwern Oct 9 Like I keep telling people, everything on the Internet linkrots and you have to make your own. The IA is great, but as you can see, they can't keep up with the degradation of webshit. Here's my mirror:
          362
          55
          15.2%
        • ????? @gwern Oct 9 Substack is only a 'great reading experience', for low values of 'great', because everything else is so terrible and they haven't experienced any pressure. But every time I have to delete the 'subscribe' screen blocker popping up at random, I think, 'Medium used to be OK too'.
          96
          5
          5.2%
        • ????? @gwern Oct 9 Obvious presentation would be a 'This estimate uses: A B C D E' union, then on each page hyperlink to the source for used ones, and dim out unused ones.
          78
          0
          0.0%
        • ????? @gwern Oct 9 I assume that the OA legal structure isn't the *same*, if only because of the whole capped-returns stuff they came up with for OA specifically because of the possibility of extreme returns.
          560
          7
          1.2%
        • ????? @gwern Oct 9 To think my hair rates any comparison to Connor's, i am blush ?
          1,102
          30
          2.7%
        • ????? @gwern Oct 9 It's something of a beginner vs poweruser divide: a page is packing in way more than a standard LW2 is, but also has, despite my best efforts, a steeper learning curve and more visual elements. ?
          81
          9
          11.1%
        • ????? @gwern Oct 9 I would phrase it as 'weaker', at least in terms of presented information. Not 'simpler' - LW2 has a ton of features like editors or chat or crossposint which doesn't need - just weaker, in that everything like popups does less & is easier/simpler.
          82
          4
          4.9%
        • ????? @gwern Oct 9 Yes, it's now clearly real, too many people have tried it and want to keep using it (leading to the AUTOMATIC1111 drama). It could be trojaned, but by this point also enough people have tried it that someone would've noticed that too.
          56
          7
          12.5%
        • ????? @gwern Oct 9 The non-profit still exists. It pretty much has to. You can't just flip a non-profit into a for-profit and gift all the share-ownership of its assets to insiders. It's not that unfamiliar a structure. Every time you bite into a Hershey's Chocolate you're acquainted with it.
          967
          29
          3.0%
        • ????? @gwern Oct 9 (It's especially fineprint because one would stop reading after the bold headline version listing the markets it's sourcing from. I read the page and found the information I was looking for, centered, highlighted, bold, linked. Why would I have kept reading into the fineprint?)
          82
          2
          2.4%
        • ????? @gwern Oct 9 Yes, they did in the USA. There was no law banning credit cards for women; many routinely had cards, even excluding household financial units. Lenders could choose to give women CCs or not. What you are mistakenly referring to is a law banning lenders from choosing not to.
          552
          27
          4.9%
        • ????? @gwern Oct 8 (Not that the small # of markets, possibly just 1 (PredictIt), matters here, since they all have long-shot bias, and whether sourced from 1 or 5, the Harris probability will still be very low.)
          1,145
          15
          1.3%
        • ????? @gwern Oct 8 And, as I already said, that is useless because that is clearly not what they are doing since Polymarket doesn't even *have* a market, as far as I can tell, for winning the US election that they *could* be pulling odds from. That was the first one I checked, and I stopped there.
          2,484
          41
          1.7%
        • ????? @gwern Oct 8 Seriously, it's terrible. Here's like 144 uses in Arxiv abstracts, and not a single one means anything but puffery which can be deleted with zero loss of meaning. In fact, aside from maybe 5 instances, it remains just as grammatical as before. Utter misleading waste of space...
          135
          20
          14.8%
        • ????? @gwern Oct 8 No, it's not, but it's unsurprising that you think it is.
          113
          15
          13.3%
        • ????? @gwern Oct 8 PMs have a well-known long-shot bias (2.5% Michelle Obama, really?), and I'm not sure what those 5 markets are since it doesn't say - doesn't include Polymarket, eg there is no presidential market I can find
          1,462
          27
          1.8%
        • ????? @gwern Oct 8 If I was going to defend the position that Jon Stewart was a good comedian, my argument would not be that an episode of his comedy had 'the progressive position', but that it was funny.
          106
          17
          16.0%
        • ????? @gwern Oct 8 In the universe where Harris somehow gets the nomination despite being a nonentity with near-zero chance right now, something amazing must have happened to her prospects, like Biden keeling over dead. This is semi-interesting, but doesn't mean she's actually going to get elected.
          71
          6
          8.5%
        • ????? @gwern Oct 8 No kidding. And think, it's only May 2020 so we've still got 9.5 years to go!
          271
          23
          8.5%
        • ????? @gwern Oct 8 (Why waste my time in search-and-replace trying to disambiguate between the 3 real meanings when even the authors would often have to admit that they aren't claiming any of them & it's just the style of Arxiv abstracts to use that word endlessly?)
          75
          3
          4.0%
        • ????? @gwern Oct 8 Could be worse, though, at least there are 3 meanings there. In Arxiv abstracts, I've noticed that 'significan*' get used most times as mere copy-pasted puffery where it doesn't mean anything. I've begun seriously considering just deleting it automatically from Arxiv abstracts.
          60
          4
          6.7%
        • ????? @gwern Oct 8 (Concrete example: the Waymo simulations of fatal accident scenarios to benchmark how many their full stack would've avoided. I'd expect that sort of expert knowledge engineering to take the same people *about* the same amount of time if they left... once they built the AI, ofc!)
          81
          4
          4.9%
        • ????? @gwern Oct 8 Self-driving cars isn't an area where I'd expect some cheap ideas shared between companies to help much. It seems much more infrastructure/code/data-heavy: vast databanks of very rare accident scenarios to benchmark your fleet on in silico endlessly, eg. Can't take that with you.
          120
          12
          10.0%
        • ????? @gwern Oct 8 The conditionality here is a bit misleading - like Harris my be P(40%|nom) but she is dead in the water, her shares aren't even trading note the complete lack of any blip when she was reported to be in a car accident a few days ago!
          1,571
          36
          2.3%
        • ????? @gwern Oct 8 Yeah, it's become a countersignaling thing for elite leftists to talk about how, now everyone's seen it, _Hamilton_ is overrated and overpriced and something something racism (just invert all the previous talking points about its brilliance).
          183
          8
          4.4%
        • ????? @gwern Oct 8 (Thankfully, like _Hamilton_ not being that good, you just have to wait a few years. The owl of Minerva etc.)
          481
          22
          4.6%
        • ????? @gwern Oct 8 That's such a 2000s thing to rant about. Might as well rant about the threat of Intelligent Design.
          2,245
          16
          0.7%
        • ????? @gwern Oct 8 It was both, more or less. He and Colbert could be decent during the GWB years, but as soon as Obama was elected, all the comedians died (despite Obama continuing so many GWB policies!). They could go after the out-group, but not the in-group.
          7,184
          240
          3.3%
        • ????? @gwern Oct 8 True, true, there are a lot of suns out there to keep burning, and everyone must do their part.
          149
          8
          5.4%
        • ????? @gwern Oct 8 It's tricky, though, because the scaling laws lit also shows that models were often being trained much too long (eg that RL paper last week), because of a default assumption you should try to reach convergence. The key part is the *more data*, seems - which is usually skipped.
          664
          23
          3.5%
        • ????? @gwern Oct 8 (Looks like the model hasn't been trained on enough snuff/gore.)
          142
          6
          4.2%
        • ????? @gwern Oct 8 (No, it was, it's just that the $1t+ of cleanup costs is off-books for the Project eg we're still cleaning up Hanover because there's dozens of tanks of serious shit no one knows how to deal with and so they just sit there in the ground.)
          66
          4
          6.1%
        • ????? @gwern Oct 8 (I am impressed by $100b of R&D. I'm not impressed by $10b of R&D done 5 times over and then another 5 R&Ds based on organizational sunk costs that no one would do when starting from a blank sheet today.)
          1,063
          52
          4.9%
        • ????? @gwern Oct 8 I tend to suspect a lot of that car research was DOA and completely wasted. Anything based on a lineage out of 200s-era DARPA competitors seems like a dead end today. Plus, *huge* amounts of reduplication due to total secrecy about everything.
          2,119
          82
          3.9%
        • ????? @gwern Oct 7 A superintelligent AI plucked from the space of random programs is as morally valuable (or esthetically interesting, or...) as the output of a random page of x86 assembler opcodes. A non-random one, like one built by the Aztecs, could be arbitrarily bad.
          204
          15
          7.4%
        • ????? @gwern Oct 7 (well which is it, was the password 'password' or did they use a 0day on a JS lib etc)
          65
          10
          15.4%
        • ????? @gwern Oct 7 Some would say that Western civilization is already a good 4 or 5 levels of this deep.
          4,113
          177
          4.3%
        • ????? @gwern Oct 7 But I expect most DM compensation is in the form of dollar-denominated Alphabet stock, and most of the technical expenditure is for electricity, which they would be purchasing in the USA (and if not, the EU energy cost increases will be relocating their cloud workloads ASAP).
          111
          5
          4.5%
        • ????? @gwern Oct 7 Hard to say. After all, it's easy to spend tons of R&D with nothing to show for it (especially in DL...). Doing the right thing is much more important than doing expensive things; nothing so costly in R&D as bad taste combined with insistence on being penny-wise pound-foolish.
          1,063
          36
          3.4%
        • ????? @gwern Oct 7 That requires a framing that the $4.5t was necessary and not the result of deliberate choices in terms of funding research/prevention, such as refusal to do challenge trials so vaccine rollouts were delayed. (Plus, the relief was super incompetent even strictly on econ grounds.)
          446
          14
          3.1%
        • ????? @gwern Oct 7 Well, go spend an hour applying SD to a photo or video of your house (putting the effort consumption up front so no one can miss it), and see if you can get millions of YouTube video views and Smithsonian and BBC etc articles on it.
          1,128
          46
          4.1%
        • ????? @gwern Oct 7 A good example of art as conspicuous consumption/leisure. The lack of interest any of the coverage has in looking at the doodles up close or for more than an instant shows that everyone is only interested in the sheer profligate waste of human time/labor. If it was done by AI...
          10,096
          434
          4.3%
        • ????? @gwern Oct 7 I'm not sure I'd even give it that much. There are loads of pidgins or smaller or simpler alphabets or scripts. Just pure network effects.
          72
          3
          4.2%
        • ????? @gwern Oct 6 A bigger model. Xin et al 2022 is a familiar Bitter-Lesson methodology result (eg recommender papers): the fancy algorithms reporting increasingly better perf over the years are getting better perf... it's just all from more n/compute/param, not the researchers' hard work.
          173
          6
          3.5%
        • ????? @gwern Oct 6 I didn't see any discussion or allusions indicating how the hack was done. (Might be more clues in the torrent, but I didn't download it.)
          394
          20
          5.1%
        • ????? @gwern Oct 6 I suppose. Presumably you could compare the ckpts and only parameter values should differ and no new Python op codes involved. Or load it with Python libraries intended for this, I assume someone's written pkl sanitizers.
          197
          7
          3.6%
        • ????? @gwern Oct 6 [kneejerk response to Nat: "Oh, OA already is."] Now the interesting question is, how much do those fluent errors matter for training not ASR but pure *text* models? Machine translation and LM work in general seems like it'd suggest the harm is minimal, & the bigger data better.
          223
          8
          3.6%
        • ????? @gwern Oct 6 Nah, they're mediocre at best. As symbols, they have flaws like several being equivalent under symmetries, wildly varying stroke styles/counts, inefficient to write. As phonetic encoding, lacking sounds, illogical, parts obsolete. Lots of better shorthands or scripts like Hangul.
          2,444
          75
          3.1%
        • ????? @gwern Oct 6 "CSS and usersheet styles. You've invented usersheet styles."
          5,769
          139
          2.4%
        • ????? @gwern Oct 6 3 year update: still frozen in amber. The latest post: tldr: Brin discusses, for the millionth time, how he is the best predictor in the world (fans will be shocked by lack of mention of _Earth_), and the Republican Party is the worst thing in the world.
          79
          8
          10.1%
        • ????? @gwern Oct 6 I am pretty sure they said publicly it was SD, so not being SD-finetuned would be the surprise.
          718
          27
          3.8%
        • ????? @gwern Oct 6 (It's especially funny because the hacker doesn't sound like he's angry about 'theft' or 'copying'. He's just cheap.)
          11,064
          161
          1.5%
        • ????? @gwern Oct 6 Today's anime AI drama: it appears 4chan may've hacked & leaked in a torrent 's model-checkpoints & website code? No clear confirmation in the threads yet. (I'd be wary of running since as a loudmouth anon points out, ckpts=code.)
          10,303
          396
          3.8%
        • ????? @gwern Oct 6 No. Name a correlation metric where it's correlated. Pearson? Spearman? Mutual information? If you can, you've broken it.
          122
          14
          11.5%
        • ????? @gwern Oct 6 As in, "The output of an encryption algorithm is uncorrelated with its inputs, and uncausalated with stock market returns."
          1,794
          27
          1.5%
        • ????? @gwern Oct 6 Either way, should've spent more. Vaccines were an open field.
          388
          6
          1.5%
        • ????? @gwern Oct 6 (Eh. I think they're substantially worse - or if they are, they are only reasonably comparable within a very, very narrow domain. (ProGAN could do faces great, but doesn't get you to Imagen.) TADNE is much better, and compare TADNE samples to Waifu Diffusion or NovelAI.)
          1,215
          12
          1.0%
        • ????? @gwern Oct 5 Well, no one's going to write '1986 36ya'. Visually nonsensical, nor do even scientists write like that. But the long-forms they use like '1986 (36ya)' aren't going to catch on either. I also don't intend to write it out explicitly given it changes; it'd be a plugin, of course.
          56
          1
          1.8%
        • ????? @gwern Oct 5 This raises an interesting point: apples are supported by 'clubs' of patented apples. Why not new bananas? Patentable, and the giant quasi-monopolies would be able to reap almost all of the gains. Are they just too fat and lazy to try to improve their monocultures?
          4,779
          132
          2.8%
        • ????? @gwern Oct 5 (The two major exceptions for native support that come to mind are Twitter, which doesn't support any formatting, and, bizarrely enough, which apparently is why Substack (!) doesn't support <sup>/</sub> or footnotes.)
          58
          2
          3.4%
        • ????? @gwern Oct 5 Nah, they're easy to use. They are some of the most basic formatting there is, supported by every markup language in common use, and as long as Unicode is supported, you can still hack it as I demonstrate there. The markup is usually short too: '~36ya~' in Markdown, eg.
          59
          3
          5.1%
        • ????? @gwern Oct 5 (One can speculate on what's going on there. I suspect it's a mix of 'context collapse' in all eras of media available intermixed, power laws, and extremely lengthy careers - after all, how long ago could The Beatles have been, really, if Paul McCartney is *still* concerting?)
          4,184
          66
          1.6%
        • ????? @gwern Oct 5 Like, does Tom Cruise in _Top Gun_ ???? 37 years ago? If I mention '_Scooby Do_ started in 1969', does that ???? 53 years ago? Do Top Gun/Scooby-Doo ???? twice as close in time as Top Gun/you? They still don't to me, and I just did the arithmetic!
          3,392
          23
          0.7%
        • ????? @gwern Oct 5 Another subtle cognitive distortion I've been noticing: date intervals or durations. It seems like people are increasingly losing their grip on intuitive temporal durations, thinking in a Perpetual Now, and are increasingly surprised when I tuck in an explicit 'X years ago'.
          8,064
          116
          1.4%
        • ????? @gwern Oct 5 Yes, the analysis is an amazing read. And the author is a college student who became a hardware guy as an outgrowth! I also thought the Dream Minecraft one was interesting as an intro into Bayesian reasoning for lots of youngsters, and his fans' non-reaction to his confession.
          57
          0
          0.0%
        • ????? @gwern Oct 5 Yeah, that's pretty crazy. My first guess is that it's because people namecheck or mention Engelbart all the time, but usually won't formally cite 'Engelbart 1963' (much less some sort of 'Engelbart 1968' for the Mother of All Demos). Bet something like Google Books hits sensible
          1,592
          47
          3.0%
        • ????? @gwern Oct 5 (And recently added to Libgen, along with the rest of her missing books. RIP.)
          40
          2
          5.0%
        • ????? @gwern Oct 5 IIRC, the rationale was actually to get an updated version. I assume they also keep it continuously updated by polling IDs/RSS or something.
          135
          8
          5.9%
        • ????? @gwern Oct 5 That's incorrect. NovelAI told me they did their own scrape.
          3,816
          77
          2.0%
        • ????? @gwern Oct 5 What they want is him to be forced to buy it, then sell it at a loss, so they can enjoy the schadenfreud twice and then have a reliable fallback for the next few years "Elon Musk paid $0 in taxes in 202x!" when the outrage wells run dry.
          4,093
          150
          3.7%
        • ????? @gwern Oct 5 I kinda want the Will Smith meme now 'SMITH: "Can a North American river otter derive backpropagation? Does a river otter know how semiconductors work?" OTTER: "Do ????" SMITH [in tears]: ""'
          2,390
          87
          3.6%
        • ????? @gwern Oct 4 "The report says dozens of grandmasters have been caught cheating on the website, including four of the top-100 players in the world who confessed." whoa ?
          101
          12
          11.9%
        • ????? @gwern Oct 4 But now that we can see how much PaLM can do, and have directions like Gato/Multi-Game-Transformer and startups like Adept, scaling DRL is becoming more relevant.
          122
          7
          5.7%
        • ????? @gwern Oct 4 DRL was already over-invested in if you look at the compute-history graphs: projects like AG/A0 were trend-breakingly expensive, while having no practical use. Meanwhile, vision-language models do lots of emergent and useful stuff, and would empower DRL agents as foundations.
          136
          9
          6.6%
        • ????? @gwern Oct 4 No. It's fun but I'm not convinced, or if I was half-convinced then, years of watching intellectuals be politically craven on both sides of the aisle has disabused me of the notion there's anything inexplicable about a lickspittle Machiavelli requiring such an explanation.
          475
          35
          7.4%
        • ????? @gwern Oct 4 I see. Confusing, but it is what it is. I guess the tree search self-play does make it very unclear, compared to eg ALE episodes, what the equivalent of 'tokens' would be in trying to distinguish between those scalings.
          53
          2
          3.8%
        • ????? @gwern Oct 4 So they were too small under both Kaplan & Chinchilla scaling laws? (I'm left a little confused what is proven in what settings given the factors like only 2-layers and the problems being small enough to reach optimality etc.)
          40
          11
          27.5%
        • ????? @gwern Oct 4 Of course you can tell them, and I'm sure they did just that. People make comparisons all the time. "We want to make a movie like X, but better, with blackjack and hookers." "Z is like X meets Y! You in?"
          75
          2
          2.7%
        • ????? @gwern Oct 4 You're the one proposing one of the largest expansions of government-enforced monopolies ever, I think you ought to rotate on it a lot more than you have. No one said the artist's name has to be involved. Consider embeddings. 'disenfranchised' is meaningless rhetoric.
          1,251
          94
          7.5%
        • ????? @gwern Oct 4 Yes. All RL agents are extremely unsafe by default, because their entire point is to *do* things, not 'merely' generate data; therefore, anything implying large easy capability gains in RL is much worse than the same news anywhere else.
          113
          6
          5.3%
        • ????? @gwern Oct 4 OK, so that's nothing like your Ghibli example, then, as I thought. And it also has no real relevance: oh, you can't use their name in the title? Fine, people just won't share their prompts lest it imply endorsement or authorship. Wow, that was easy.
          54
          2
          3.7%
        • ????? @gwern Oct 4 Really? Enforceable under what grounds? What could Studio Ghibli sue its ex-staff at Studio Ponoc for? (Or Gainax Trigger?)
          84
          4
          4.8%
        • ????? @gwern Oct 4 Fashion is one of the largest and most influential creative industries in the world, not 'an edge case'. Mickey Mouse is a recognizable character and *not* a style. 'Steamboat Willy style' is instantly recognizable... and owned by no one. Unprecedented = giant monopolies bad.
          811
          47
          5.8%
        • ????? @gwern Oct 4 Because it is the most dangerous and also relatively useless place for scaling to hold. It's like discovering that any weed grow farm or Bitcoin mine can synthesize weapons-grade plutonium from ingredients under a kitchen sink.
          142
          17
          12.0%
        • ????? @gwern Oct 3 Leaving aside the dubious sense in which buying stock is actually equivalent to either investment or involvement or effort, you can't buy stock in either, either. You can only buy stock which is like <10% 'Cruise' or 'Waymo' (ie GM or Alphabet).
          78
          8
          10.3%
        • ????? @gwern Oct 3 It's also hard to contribute in a meaningful way from the outside, short of getting hired by Cruise or Waymo. If you or I or Tanaka wanted to contribute 10 hours or $1,000 to Self-Driving Cars?, we would do this... how?
          483
          23
          4.8%
        • ????? @gwern Oct 3 Why do you think that is necessary? Fashion doesn't have copyright on designs, much less 'style' What a nightmare it would be if someone owned 'styles' like 'anime', which is exactly where rentseeking would take such proposals...
          7,018
          297
          4.2%
        • ????? @gwern Oct 3 The sentencing will also always be tainted by the massive (and probably still incompletely-purged) LE corruption. Hard not to feel that played a role in the judge going berserk on Ross...
          105
          11
          10.5%
        • ????? @gwern Oct 3 (As I have been saying all along, there is - very unfortunately - no good a priori reason to think that scaling laws will not apply to RL just like they apply everywhere else, and the reward vs sample/compute curves available have always indicated as much.)
          3,143
          90
          2.9%
        • ????? @gwern Oct 3 (Damn, that's not even close to a boil-down of it. That's "I didn't read past the thumbnail" boiling-down.)
          39
          1
          2.6%
        • ????? @gwern Oct 3 ATTENBOROUGH: "As the parasite matures, its ordinate sacs swell with green fluid. Soon, one day in autumn, a frost will trigger a hormone surge, and it will be time. The parallel sacs burst, seeding the wind with spores that, should they chance upon a host, become¡­ architects."
          261
          28
          10.7%
        • ????? @gwern Oct 3 Somehow, they didn't feel the need to snub Jennifer Doudna for merely 'implementing' the CRISPR techniques discovered by others (or the cryo-EM triple Nobel or...).
          875
          32
          3.7%
        • ????? @gwern Oct 2 But it knows the important thing about guns: trigger discipline.
          2,078
          38
          1.8%
        • ????? @gwern Oct 2 I spent plenty of time using darcs & Mercurial to collaborate with many others, and never had a fraction of the trouble I have collaborating with a single person on code in Git. No, it's just the worst DVCS.
          63
          9
          14.3%
        • ????? @gwern Oct 1 Look, if they ????¡¯? approve his grant application for field trials of his innovative new "CPR for bats" first aid technique, that would be letting the conspiracy theorists win?
          657
          47
          7.2%
        • ????? @gwern Oct 1 But they are typically not simultaneously masters at anything important like rocket science.
          326
          18
          5.5%
        • ????? @gwern Oct 1 (I guarantee you that if a lawsuit dumped the complete texts of Terry Tao for hostile journalists to dunk on, we'd find at least one joke on the level of '5318008'.)
          1,331
          65
          4.9%
        • ????? @gwern Oct 1 If there's one thing we should've learned by now, it's that everyone sounds dumb in their texts, which is why Warzel conspicuously fails to mention any comparisons, merely comparing against some imaginary 'actually-impressive' moving-goalpost set of texts.
          6,172
          228
          3.7%
      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
      3.4%
      Oct 31
      2.9% engagement rate
      Link clicks
      4.7K
      Oct 31
      277 link clicks
      On average, you earned 150 link clicks per day
      Retweets without comments
      1
      Oct 31
      0 Retweets without comments
      On average, you earned 0 Retweets without comments per day
      Likes
      7.4K
      Oct 31
      133 likes
      On average, you earned 237 likes per day
      Replies
      493
      Oct 31
      13 replies
      On average, you earned 16 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 0016ecjdsk.com.cn
      www.mcchain.com.cn
      www.ep500.com.cn
      ghvf.com.cn
      rhiwip.com.cn
      smzc1688.com.cn
      svenya.com.cn
      qhdp.com.cn
      nianz.com.cn
      mkiusi.com.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