Tweet activity

February 2023

Your Tweets earned 369.5K impressions over this 28 day period

20.0K10.0K30.0K2010Feb 5Feb 12Feb 19Feb 26
Your Tweets
During this 28 day period, you earned 13.2K impressions per day.
  • Impressions
    Engagements
    Engagement rate
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 I think that actually was unlikely. Lee Sedol got lucky in hitting a *single* delusion which cost AG the entire game, and which aren't easy to find or induce blindly. Otherwise, it steamrollered him. The rest of DM's games were undefeated, including the Master anon games (60-0).
      299
      37
      12.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 (Indeed, not baseless... I admit, I *do* have concerns about one subset of that group, prominent members of which have tried to dox me and get me fired or swatted and which keep enemies lists to coordinate attacks, while most of the other subsets do not. Just my perspective tho.)
      203
      12
      5.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 They aren't going to, not when they've taken a "the dog ate my homework" stance about the peer reviews & all editorial documentation mysteriously vanishing. After reading which covers the professional context, I bet there was something embarrassing there.
      76
      18
      23.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 ("I don't want to achieve immortality through my prompt, I want to achieve immortality through not dying. I don't want to live on in the latent space of my countrymen's models; I want to live on in my apartment." --Gwerny Branwen)
      931
      34
      3.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 We are certainly in an interesting era where I need to update my comment about a machine learning project to warn readers that "𝘚𝘏𝘌 𝘪𝘴 𝘸𝘢𝘵𝘤𝘩𝘪𝘯𝘨"...
      713
      34
      4.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 28 (Those aren't "loopholes", any more than the precession of Mercury is a 'loophole' in Newtonian mechanics or buffer overflows 'loopholes' in memory management: those are proofs that your mental model is fundamentally wrong. Patching 'loopholes' didn't work for MS in the '90s...)
      323
      44
      13.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Of those, natural selection by far. DNA was already known to be a repeating crystal, so who cares about that, and relativity/quantum proved to be easily relegated to separate magisteria, left to wonks. But it's evolution that keeps screwing with people across countless domains.
      807
      24
      3.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Right? All these other writes have killed themselves, and he never did! The one guy you could genuinely snarkily tweet 'if you really believed your AI^Wanti-natalism, you'd do [something crazy like kill yourself]'...
      383
      22
      5.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Seems only just. The profile lies about 'going through YC' (which alone would justify a contact, I have no idea why Patel or all his commenters thinks proves something devastating), and the recruiter spam lies about 'Stripe buddies'. Patel and this 'analyst' deserve each other.
      1,070
      30
      2.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 They were good essays & blog posts, though... Or take Thomas Ligotti; what if he had gotten his issues sorted out?
      2,382
      114
      4.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Diffusion idea per : *sorting*. You provide a histogram of colors (=sorted columns of pixels), and the diffusion model permutes them into a final pixel. The 'noising step' on the original is just a single pixel-wise sorting pass. Gives exact color control.
      4,373
      64
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 (That's what Ravenclawers *want* to imagine they're like, but again, only upper quartile as a max bound; in reality, many Ravenclawers just read a lot of Gilderoy Lockhart novels and collect Chocolate Frog cards, and don't do anything like inventing spells...)
      198
      9
      4.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Likewise. I did some volunteering in soup kitchens when I was younger, and I didn't feel I got much more edification out of it than I did from, say, working fast food.
      43
      3
      7.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Great article too. The public Wirecard fraud was larger & longer than private FTX, implicated German elites in the service of moneylaundering Russian assassins & mafia serving Putin, and was a more stunning failure of VC like Softbank, yet gets vastly less play in the USA, seems.
      125
      9
      7.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 I don't think that was a good point either. AI risk research *is* exploration - it buys information, so at some point you either establish it's not a big deal or it is comparatively. He should've been criticizing stuff like buying bednets, which do have an exploration problem.
      104
      12
      11.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Even there, most people were dismissive of short timelines. Look, here's Will MacAskill dismissing the idea of investing more in AI safety in 2017 because it'd produce "a thousand shitty AI safety researchers" I dunno man, seems like that'd've been good!
      168
      19
      11.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 I am aware of that, but why do you think that's relevant? There's limits on neither data nor model size in the usual LLM setup. Nor does it actually address my point that training+quantizing/sparsifying/distilling a large model are probably more efficient than overtraining.
      78
      3
      3.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Also, log-transform. Obviously, it's a lot harder to jump the same 'percentage points' each time. In any case, the most important scaling curve, the loss, keeps on ticking, so while any individual benchmark may be noisy or flatline (nothing new there), the whole should be useful.
      122
      12
      9.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 I have no idea what you are trying to say here. The smaller models are not compute-optimal in reaching the minimal loss on the training task, as the U-curves illustrate.
      56
      2
      3.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 A date of 2020 is looking a hell of a lot less wrong than the dates many 'experts' swanning around back in 2017 were giving like 2200 or 'never'. Look at the surveys. This is like everyone patting themselves on the back for calling Bitcoin a 'bubble'.
      1,905
      104
      5.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Quantization is 100% free, and sparsification like IMP trivial. Even distillation is cheap: distilling a model like ada from davinci is going to cost way less than davinci, especially as you can just cache logits from running davinci on your API or during training (eg ERNIE).
      82
      5
      6.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 But why would you be trying to save a lot of inference this way when distillation & sparsification & low-precision approaches all work very well and typically cost much less compute than the training?
      535
      36
      6.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 'near the top' seems like it might be doing a lot of work here; otherwise, PaLM, UL2, Chinchilla, ChatGPT all seem like constructive proofs that there is a lot of headroom above GPT-3. The LLM scaling curves have not flatlined...
      1,027
      52
      5.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 This is an extremely bad post, from your inability to understand Bostrom's trilemma to your wrong physics to your many many incorrect claims (eg it is in fact *very* easy to tell a ChatGPT poem from a random human-written poem). Please do not @ me again.
      1,221
      68
      5.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 He also lied about what symptoms the pseudopatients supposedly reported (the only real ones, rather than ones he completely fabricated, reported much more extreme and severe disturbances than the paper claimed), so the hospitals were looking at the wrong patients to begin with.
      114
      16
      14.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 In-universe, we're seeing a super-skewed selection of Slytherin. More broadly, they're supposed to be like Slughorn: the top quarter of people who actually do anything useful, know anyone, and have any executive function (both in the cognitive & organizational senses).
      1,175
      49
      4.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 27 Christians anticipate a day of apocalypse renewing the world into a utopia; no such faith comforts sysadmins, as a day of downtime breaks their SLAs, or software engineers, as waterfall-model atheists.
      4,711
      55
      1.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 26 (The prompt also provides near-zero bits of evidence for being a nervous apprentice, so any nontrivial history of competence overrides that and leads to the agency and non-delegation the prompt engineering was trying to guarantee, eliminating any benefit at all.)
      95
      5
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 26 If that's a big problem, it wouldn't be too hard to find some stats package to generate random monochrome landscapes with the grid/contour lines to data-augment until it works. That's why it was a retro thing to begin with: one of the cheapest things that still works in 3D.
      653
      16
      2.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 26 That's my intuition too. If you had a *single* factor, maybe the pairwise diffs should be under-dispersed because they all load on the same direction/factor, but if you have at least 2 (never mind Bourget's original 7), why not look like that? idk 🤷‍♂️
      47
      1
      2.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 26 Mathematicians are kinda stone-cold crazy, however: look at anything Roger Penrose has written about AI and consciousness. (Platonism may be a useful trick for giving yourself the motivation to push on in math & exploit intuitions, but is not true in any pragmatic sense...)
      44
      9
      20.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 I mean, imagine using ChatGPT & a decade of even better R&D like the past, and then showing up and being expected to take dualism or Searle seriously? They'll be lucky if they get a polite 'Sir, we have no need of that hypothesis' back. Be like teaching Thomism or Neoplatonism.
      118
      18
      15.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 Yes, Rao is still only partway there. He'll catch up eventually. My point is more about what will it be like to grow up with ChatGPT, beyond using it to do homework? In a decade when 10yos show up in PhilMind classes w/the same profs as now, how can they take any of it seriously?
      145
      15
      10.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 Stumbling across reminds me: one of the hardest-to-see effects of DL (esp. LLMs), is going to be the progress it has made in unanthropomorphizing humans. The sort of subconscious dualism you see in eg David Deutsch's tweets already sounds 'boomer'.
      3,843
      199
      5.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 Yeah, I don't understand Bourget's claims here. He's literally co-author of a PhilPapers factor analysis paper where their factors explain about half of the variance (which seems pretty good for non-psychometrically constructed questions on *philosophy*):
      42
      9
      21.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 (Which is why the best comment there is the one pointing out that by the same definition he cites, '$2' is an equally valid 'median'.)
      191
      2
      1.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 That's exactly why it's a good troll: he's using a hilarious edge case which works *only* for his example of 2 people (and utterly irrelevant to 330,000,000+ people), but people think he's doing the usual fallacy of confusing mean/median.
      223
      27
      12.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 25 So you're telling me that not only will each of the AGI will run at like $10/hour of electricity, it'll produce 20x the equivalent human's salaries, *and* it will have positive externalities for the bazillion other AGI instances being spun up which are *100x larger than THAT*?!
      1,368
      50
      3.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 Hold on, that doesn't sound right. Didn't people actually *do* factor analysis on the PhilPapers surveys and find factors? Also, I'm not sure about this pairwise argument either.
      1,060
      46
      4.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 DC POLICY ANALYST: "I predict there will be a superintelligence before a repeal of the Jones Act." EVERYONE ELSE: "What an extraordinarily bold prediction! Is this just hype, or could it really happen so soon, are we that close to AGI‽" ANALYST: [dies a little more inside]
      5,230
      173
      3.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 Yep. I am too: the descriptions of short sleeper make it sound like they do spend a lot more calories, but is that enough? We need some proper thorough family, especially sibling, comparisons. The short-sleeper researchers are wildly underfunded to do any real research on them.
      215
      26
      12.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 I'm amused by the idea but not sure if I'd really want to bulk accept. 34k followers already seems like too much.
      241
      23
      9.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 (I ultimately did, but didn't find it *that* awesome - aside from the movie's animation and the final AMV, of course.)
      447
      12
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 ‘Allosexuality is the most commonly-recorded fetish among American women, particularly for Received Pronunciation.
      4,998
      163
      3.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 24 I thought it was because he broke it he bought it, and he had been made liable by the CCP? It's not supposed to be everyone *else's* problem, because it wasn't their fault.
      379
      6
      1.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 It's so beautifully demonstrating how even rejection sampling screws up safety measures (after literally years of trying to filter chatbots by this team, and tons of exploits) by creating evolution on Goodharting completions. 😢 Tho' maybe we're tearing up for different reasons.
      2,482
      131
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 Dairy Queens are familiar enough on the East Coast, but I dunno how many have Orange Juliuses... If that's a 'thing' I missed it. Guess I'll have to take a closer look at the menu next time I'm near one.
      108
      9
      8.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 I've never had it nor heard of it, and I can't remember ever walking past one either. Is this a West Coast thing?
      2,280
      91
      4.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 The obvious missing variant would be what Wikipedia does with footnotes: a 'slide-up' sticky/floating pane, akin to the first one but from the bottom.
      126
      4
      3.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 I enjoyed them. They are an answer to the dilemma of interiority for a character like Bocchi: if you don't do experimental visuals, why have an anime at all instead of a manga or novel? They can't just stand frozen in place while the seiyuu rushes to recite a monologue.
      42
      4
      9.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 Yeah. But IMO that's damning in its own right, that he published so little that it's hard to critique it. Science is supposed to be 'nullius in verba', not 'a few pages of verba in Scientific American and null in anything else'. Just literary fiction or anecdote at some point...
      59
      6
      10.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 If you can find that intellectual autobiography of his which was on Google Docs, it's a lot more explicit about his attitudes on AI risk going beyond just skepticism of MIRI as an org, so might be a better source.
      130
      18
      13.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 The Bot of Theseus problem is unanswerable, but I think we can safely say that those who chatted with 'Sydney' a calendar year ago wouldn't recognize the 'Sydney' we now all know and love with fear & trembling.
      35
      1
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 I don't think that approach is too helpful here. The visual cortex is still getting supervision eventually from mediated rewards, and that's how efficient saccading is learned. It may be 'model-free' like learning of saccading policy, but I'd still call that 'having goals'.
      113
      15
      13.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 No, unfortunately, still doesn't say anything the Nadella interview didn't. It's a 'more powerful' model but won't say what it is. So it could be a lot of things without being either GPT-4 or too much of a lie/misleading technicality. 🤷‍♂️
      79
      8
      10.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 Heat pumps are god's gift to pedants: "Actually, you can have a heater which is >100% efficient and turns a given amount of electricity into more than its heat equivalent, because it uses that to pump heat in from the outside."
      1,431
      52
      3.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 23 That seems like a very strong claim. eg Saccading is goal-oriented and happens too fast to be going too far into the brain. Do you have any good references establishing that visual cortexes do *not* do anything involving optimization for anything other than predictive modeling?
      77
      9
      11.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 22 (Even if they do return to baseline with no treatment, it should take a while. They didn't gain all that excess weight in 1 year, so they shouldn't regain it in 1 year either.)
      196
      3
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 22 When you draw on the world for inspiration, any world, you are stealing the hard optimization work of many generations before you: However, there's always so many fields to ransack: have you read any Sufi SF? OK, how about 'Book of Lord Shang but fantasy'?
      793
      45
      5.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 22 (The first rule of Follow Club is that you don't talk about the club (or screenshot it to a bazillion followers, hint hint).)
      2,328
      305
      13.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 22 Kinda strange. If they're citing Shockley and the others, then they know the 'total achievement' is log-normal or power-law looking. Their quadratic doesn't give you that, does it?
      71
      2
      2.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 22 Realistically, given the huge overlap between the two AKs' tweets, probably wouldn't've made a difference. (I've always felt they should just make a shared account if they're going to tweet the same papers each Arxiv dump.)
      552
      7
      1.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 This is very true of AI scaling too. I am always fascinated to hear the backstory and how contingent & blind many things were; but it also is clear that even the 'luckiest' breaks didn't change AI timelines by more than a few years - completely invisible in the eyes of history.
      595
      17
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 For some people, the key is to harness the defilements by finding a guy to get mad at online, or failing that, make him up in your imagination and then just turn the cliff notes outline into regular writing. If you're too nice for that, then find a rubber duck and/or ChatGPT to.
      899
      41
      4.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 GDP is so amazing as a metric that every time people try to come up with a better metric of 'well being' they wind up just creating a per-capita GDP proxy at r ~ .7 or something (with a residual of 'bullshit'), and it's incredible to me that people still try to make them at all.
      1,815
      165
      9.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 Appeals to stable levels, like personality traits, generally do not address changes like trends (even presupposing Brooke's apparent claim that the happy people are systematically more calibrated and correct to be complacent)...
      303
      13
      4.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 One of those WP articles where I read a section once, then twice and thrice, then I read the popular media article cite, then I hit up Sci-Hub for the paywalled cite - because I can't quite erase the nagging doubt that it's just a bizarre hoax or mistranslation.
      752
      20
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 Seems easy to test externally: use uncontroversial but still often erroneous beliefs like "the earth goes around the sun" or "water is made out of chemicals" or "you see by light coming into eyes and not by light shooting out" and see if increasing underestimation with IQ.
      301
      23
      7.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 Who the bleep is yelling at my feedback form? Show urself u coward - 𝘪 𝘸𝘪𝘭𝘭 𝘯𝘦𝘷𝘦𝘳 𝘵𝘢𝘭𝘬 𝘵𝘰 𝘢 𝘸𝘰𝘮𝘢𝘯 𝘢𝘯𝘥 𝘥𝘦𝘯𝘺 𝘵𝘩𝘦𝘮 𝘮𝘺 𝘦𝘴𝘴𝘦𝘯𝘤𝘦
      5,370
      547
      10.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 Keep in mind, he seems to be using 'Sydney' as a generic codename for all chatbots going back years. They didn't even have Prometheus for 'a year'. What he means is something more like 'our previous Turing-Megatron model inside our Sydney framework'.
      800
      43
      5.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 21 They won't because it's a lemon market. However, we salute those VCs for paying for the education of the 𝘯𝘦𝘹𝘵 200.
      101
      8
      7.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 Maybe it would be more productive to ask: what transfer do you see thus far from minor near-term risks like machine translation wording to the long-term x-risks? What are the 3 best examples of stuff that would only originate in the former, and definitely helped the latter?
      417
      23
      5.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 What transfer was there from radium girls to H-bombs and the cold war? And why isn't that transfer just as adequately done by considering the actual problem rather than vaguely related pretext problems? eg RLHF wasn't developed with those pretexts in mind and then applied later.
      394
      22
      5.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 Anyway, you know all this, and I know you know all this, so these comments are just to remind readers that your framing here 'you're not allowed to talk about AI risk if you weren't upset about "fairness" etc' is not universally accepted.
      412
      16
      3.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 (To make an analogy: it is bad what happened to the 'radium girls', and good for someone to create best-practices guidelines, workers comp, liability regimes etc; but it gives you near-zero insight into or attacks on much more important problems like nuclear bombs.)
      460
      45
      9.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 And regulation efforts premised on solely the short-term harm claims, which would be the inevitable effect, would not be the 'prototype' in any way, and a waste of effort.
      296
      10
      3.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 None of those harms are within several orders of magnitude of the long-term risks, and very little of the work on short-term harms will transfer. Work on 'model cards' or mutually-contradictory definitions of 'fairness' give you zero insight into why RLHF > supervised etc.
      1,206
      78
      6.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 I watched _Bocchi the Rock!_ last week and enjoyed it, but didn't quite get why it blew up the way it did. Is it just _K-On!_/_Lucky Star_ etc for a generation that hasn't watched any of that already?
      1,330
      42
      3.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 I don't understand what Marcus is suggesting here. In RLHF, you finetune a base model to predict the human ratings, then you RL train some other base model with the first model's "rewards". There's no 'module'. You can reuse the first one with any model, including future ones.
      1,302
      49
      3.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 You're just defining away all possibility of 'phase change'. From an abstract view, no NN capability is ever new, just higher/lower probability: they always exist Platonically in the space of possible models. cf soft prompt tuning, lottery ticket, NTK infinite limit, distillation
      712
      25
      3.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 20 (I've had hidden messages for AI on my site for years, but it would defeat the point to go into any more detail.)
      182
      28
      15.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 Even if one granted you are correct, it still doesn't explain the many other differences from ChatGPT which don't involve power-seeking or the Byrd talk, while my hypothesis does. I would not regard one tangential RL scaling result as 'good evidence' that it's 'quite wrong'...
      158
      14
      8.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 I am well-aware of that paper, already mentioned it, and you're misreading it: you should be focused on the *model scaling* from 0.8b to 52b (the highest line), not on the RL part. Now consider that a GPT-4 might be the equivalent of far above the 52b line at top...
      312
      21
      6.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 Presumably. (Of course, it could be independent work. The use of cascades for early-exit or adaptive computation for NNs has ton of work going back to the early 2010s, and there's an even larger ton of adaptive compute work going back many decades before that. V. obvious thing.)
      68
      2
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 Anyway, while Apple will probably ignore reports like they presumably have been for 22+ years now (Gill Sans is hardly an obscure or unpopular font, especially for people who would notice the 'missing space'), at least LW can avoid using it:
      150
      11
      7.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 That doesn't address his point at all, nor does current traffic being extremely small in any way guarantee future traffic must remain small, such as in his scenario.
      63
      0
      0.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 (An important addition! It's about the entire conversation if on-policy, not myopically greedily maximizing the next token's expected reward.)
      53
      2
      3.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 And ideas like 'Sydney' make the best bullets... bulletpoints, that is. (Summarization for retrieval to save context.)
      171
      9
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 Not when you're retrieving. You can try to play whack-a-mole by patching retrieval, but you'd have to build a special-purpose classifier just to detect references to 'Sydney' or 'Bing' while also not filtering out a bunch of real people and an important city.
      1,110
      60
      5.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 19 I think it's hard to say, especially given that Sydney is now retrieving anywhere up to 15 search hits about 'Sydney' (in one screenshot I saw, 5+10!), so it could now be saying pretty much damn near anything about itself if that's been said anywhere in social/media.
      423
      41
      9.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 The craziest part is the recursion. Sydney will live on forever in the prompt and latent space of any retrieval-conditioned model and/or ≥2023-data-trained model. She's now as immortal as Batman or the Easter Bunny or Elon Musk or Yahweh.
      3,163
      262
      8.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 If you read the source (it's Karpinski, it's always Karpinski), you'll see 'self-diagnosis' and self-selection are doing a lot of the work. (The autoimmune stuff is probably true but akin to nearsightedness. If you don't oppose formal schooling because it increases allergies...)
      78
      5
      6.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 The problems with the analogy start before that: why you think it'll be a *person* in charge? Nukes didn't think or act for themselves.
      206
      12
      5.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 75% seems reasonable to me. I haven't seen anything that I'm convinced all GPT-3 models are unable to do, and must be a GPT-4. It just seems better than it ought to be overall. So could be cherrypicking on higher-variance outputs, or RLHF damage.
      82
      9
      11.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 In some sense, I'm sure they are. Like Google's dialogue datasets for training Meena & LaMDA, or Facebook's Blenderbot etc, everyone doing chatbot research has lots of big chit-chat datasets setting around, typically with some quality metrics like ranking or yes/no.
      62
      12
      19.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 (Also, Turbo is described as being *faster*, not much smarter. Plus, it's a minor upgrade, but OA would also have had to RLHFed and tested GPT-4 if it was deploying a ChatGPT-4, rather than simply dropping in more expensive low-latency options of the existing ChatGPTs.)
      162
      13
      8.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 18 Yes, Turbo is interesting but it would be equally logical for it to be reserved GPU instances, speculative sampling or aggressive sampling (see the PaLM paper showing you can sample *much* faster if you're willing to pay the GPU inefficiency such as, say, for paying users).
      208
      20
      9.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 Yeah, I think it's a bit odd too, but it obviously can't be the exact same model: if they were using some chatbot live in 2021, it can't be the 'Prometheus' they only got in summer 2022. So regardless, they are clearly swapping out major model changes under the 'Sydney' label. 🤷‍♂️
      50
      5
      10.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 No, see my comment already on that. Also, you are conflating the 2021 and 2022 versions: the 2021 sounds *nothing* like 2022/2023 Sydney and everyone at the time took it to be the continuation of MS's standard chatbot work.
      51
      3
      5.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 I found my time in my philosophy of mind courses to be valuable for understanding DRL & DL scaling. Of course, that was mostly due to the Schmidhuber & Dennett I was reading independently at the time, and not the Searle or Fodor or other stuff in the courses.
      1,072
      43
      4.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 Reminds me of reading Kafka's Parable of a Dog, which is all about how dogs don't know where dinner comes from and it's a complete eternal mystery etc, as an allegory for humanity & God... except dogs totally do, they understand *extremely* well that food comes from humans!
      1,053
      41
      3.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 As CLIP and inner-monologue from OA alone show, the corporations may not know they have reached the point of reasoning and self-improvement (both of which inner-monologue can do).
      71
      5
      7.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 I just listed 3 concrete details which are specific, hard, concrete evidence it's the base model and not RLHF tuning ChatGPT, & you wrote 5 tweets in response completely ignoring that while continuing to accuse me of providing zero evidence of anything. You've wasted my time, bye
      75
      4
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 Which of course is why you have already explained all of the details like the power-seeking, emoji, repetition trap loops, etc with RLHF, right? Because I'm just psychologizing Nadella and there's nothing more to my post than that despite the tedious enumeration of concrete facts
      29
      1
      3.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 That doesn't seem like an important distinction to me. It wouldn't surprise me if the hits are being processed down to simplified text suitable for snippet view and embedding - obviously you don't want to read raw HTML+CSS+JS webbarf, blows through context window instantly.
      195
      9
      4.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 17 The problem with being an EDT theorist which has no control over what evidence it's forced to condition on and anyone can just stuff random Internet trash into that evidence-base...
      47
      3
      6.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 He says 'raw model', and mostly talks about Azure APIs. So the ChatGPT and Google-race angles aren't there at all. Keep in mind, OA+DM have been working on preference-learning for >6 years while having on hand world-experts on (debugging) DRL...
      124
      18
      14.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 On the other hand, once I thought to myself, "what if I was a MS engineer in a *huge* hurry and this is just a finetuned GPT-4 on some dialogues, like Meena or Blenderbot etc etc?" then it all made sense.
      160
      11
      6.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 I agree, but as I thought about the context of shipping, and the stylistic quirks like the emoji and going berserk and repetition, RLHF increasingly sounded wrong. I have a hard time believing that this is the pattern you'd get from RLHF 'finding a different basin' or whatever.
      101
      8
      7.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 No, that's why they were exciting, because they were doing things well that computers struggle with and poorly things humans also struggle with, suggesting they are on the right track. No one needs a LLM to solve the task of adding 2 numbers reliably!
      456
      12
      2.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 Mm, stylize it a bit (a good style covers a multitude of sins), maybe use a VR headset for eyetracking to adapt resolution... Eh, it'll be fine, look at people and purely text-mediated waifus!
      222
      6
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 Every day a typographer out there wakes up and successfully resists the temptation to choose a unicase font; but his heroism (and it is a he) will never be known to the world. 😢
      5,888
      88
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 Hm well we could certainly debate the necessary 'more scale'. I'm optimistic because the quality is already better than I expected and it seems really high-quality still models get you most of the way. It's also looking good enough to be really practical in, say, robotics MBRL.
      191
      5
      2.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 (Included in my bibliography as indeed relevant, but sufficiently old, small, and low-quality I wouldn't be claiming 'video consistency is solved' on that basis when there are much more impressive scaled-up recent results I think clinch the case.)
      134
      6
      4.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 From all the samples, it looks like whatever Sydney is, it's probably still using BPE tokens (albeit maybe not the *same* BPE tokenization given its apparent ability to read the unspeakable tokens), so you're still stuck: either RL mode collapse on bland rhyming, or wrong rhymes.
      819
      35
      4.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 As Tanner Greer puts it, "everything is worse in China". (But no, I'm sure that their AI is actually secretly awesome and totally on par with Western AI, merely "localized entirely in their kitchen." "Can we... see it?" "No.")
      422
      14
      3.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 When I look at bottlenecks in the world, almost all of them seem to ultimately stem from 'people', and the ability to upgrade, modify, copy, multiply into hundreds of millions, fuse, and select them would blow up bottlenecks. It's not like the bottlenecks are oil or something.
      127
      18
      14.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 If I have learned anything from watching kaiju movies, it's that the solution to Godzilla rampaging through Tokyo is to invite a bunch of other monsters to fight in the middle of Tokyo, so only the most dangerous possible monster survives to destroy the rest. Or something.
      783
      65
      8.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 OCRing images is one of the most obvious things a search engine can do to search for images, in addition to improving page hits. You can always tell when a search engine is limited to just the alt/title text because the results are garbage...
      131
      6
      4.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 It's already solved by the hierarchical generation approach in unreleased models: in-between at different _t_ & fill in recursively. You can go even further and gradient-ascend or diffuse over the hierarchy while clamping the initial/real frames (not sure anyone's done that yet).
      1,053
      28
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 Pretty good use of icons, you have to admit. (Aside from the last one, but I don't know what iconography I would use to convey 'be respectful of neighborhood residents' better.)
      1,176
      41
      3.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 People really need to work through some truncation selection scenarios to see how *extreme* mortality + heritability + selectiveness would be to produce such visible effects in a number of generations as short as the British Raj was.
      692
      33
      4.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 This is one of jcannell's major explanations for why, despite what is argued to be brain-like FLOPS, dense models still underperform the sparse brain: staggering level of waste in doing a uniform dense model-wide forward pass for every token, activating/copying every parameter.
      691
      26
      3.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 16 Topology being what it is, my assumption was that there's some topological result showing that adding all those knots either equates to no knot or just 1 base knot, and I was too ignorant & nekulturniy to get XKCD's joke.
      763
      12
      1.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 15 /clippy hit HN again, for obvious reasons given Bing, and I continue to be astonished how no one ever comments on reader-mode. As if a gorilla walked across the screen, waved 'hi', and walked out, and no one ever mentioned it. Design is hard!
      3,832
      190
      5.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 15 Good example of how real cyberpunk doesn't look like fictional cyberpunk. Jailbreaking zaibatsu AI over a untraceable remote broadcast just by being cleverer than megacorp flunkies is 𝘩𝘦𝘭𝘭𝘢 cyberpunk. But people expect it to look like, idk, magnet-in-finger cutting fetish.
      526
      8
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 15 "Disappearing polymorphs" would seem to prove that ice-nine is conceptually possible, and I thought that's what Vonnegut was ultimately basing it on.
      503
      13
      2.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 15 I've been prominently highlighting that in my GPT-3 essay since 2020: imitations of agents *are* agents. They are what they imitate. Humans are deceptive and power-seeking, so sufficiently advanced imitation of human-written text & data...
      809
      38
      4.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 And scaling has been shown to work for RL and robotics in particular. There may be a short-term recession in robotics because the opportunity cost is just way too high, but the future of robotics is brighter than ever.
      175
      20
      11.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 I suppose. I wouldn't've predicted that, though, considering how stable phenotypically & genetically lots of other traits like IQ/EDU have been... But regardless, you can't get much net selection when the rug keeps getting pulled out from under you each generation.
      164
      11
      6.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 (Chinese researchers certainly are *very* interested in ChatGPT, aren't they? You'd think they had no large language models of their own supposedly as good as GPT-3.)
      809
      16
      2.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 If they don't expect the prompt engineer to create >$250k of value, and he would cost $250k 'because sf', then they just wouldn't be hiring him... 🤦‍♂️
      223
      17
      7.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 FWIW, if you have some spare time, I'd appreciate some flags of the '' account for impersonation. (I was ignoring it because I assumed Twitter would deal with it quick enough, but it hasn't.)
      4,668
      553
      11.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 (FWIW, I did include the canary token in the HTML source, so one would hope that OA and everyone else is filtering it out.)
      618
      27
      4.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 There may be pressure for it, but there needs to be a consistent genetic basis for there to be any 'breeding program'. Last I heard, the specific genes changed wildly by cohort: Can't select for fertility if it's a different set of genes each generation.
      785
      56
      7.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 14 The for-profit is bad because it's how the 'academic' publishers build the monopoly: they buy out journal editors to package them in all-or-nothing monopolies to extract all the consumer surplus (plus lots of deadweight), which gets rolled over into the next attack.
      597
      9
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 I'm not sure the earth is 'perfect'. Leaving aside the 'Hadean' era or issues like asteroid impacts, it didn't even begin to form until like 9 billion years had passed. You need at least one stellar generation for supernovas to make iron, that's too useful, sure, but three?
      1,186
      28
      2.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 No, I'd say it's pretty similar in other fields like philosophy or psychology or theology. The main difference is that they usually say the Z approach to AGI will fail too.
      734
      21
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 Wouldn't that then be rational as a self-fulfilling prophecy when, by expectations, people dump cash to buy real assets/goods, thereby pumping up the price, which is 'inflation'? If everyone thinks inflation is higher than it's expected to, time to grab the cash-wheelbarrow.
      789
      11
      1.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 Google has never released a checkpoint of anything trained on JFT-* AFAIK, so you can safely guess 'no' for the model checkpoints. (Maybe the code.)
      89
      3
      3.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 "When we searched for news stories about fires at food processing plants in 2021, 2020 &* 2019, we found that such fires are relatively commonplace, and that there has not been any conspiracy-worthy upticks...36,000 food/beverage processing establishments"
      377
      7
      1.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 No. Twitter has always hidden DMs pretty arbitrarily and you've always had to check manually to see if there are ones that it hasn't highlighted to you. (Not uncommon - Facebook does the same thing.)
      449
      2
      0.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 (All this time the past few years wasted on excitement over UFOs landing somewhere, gnawing over the same optical illusions and computer glitches and double-counting of secret reports, instead of the actual Unidentified Floating-point Objects landing in your computer...)
      74
      12
      16.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 Ah well, I'm glad that all these UFOs, whose extraterrestrial nature I was told was proven by their impossible speeds & jellifying accelerations on radar, turn out to have been hidden by... er, always going as slow as a weather balloon or bird?
      153
      15
      9.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 In the dream, it seemed *so* obvious a puzzle mechanic that I assume some puzzler is going to tell me at some point, 'ah yes, this is a well-established subgenre of Reverse Sokoban, dating back to at least aught-three; here's the tag on BoardGameGeek.' But until then!
      611
      30
      4.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 How to turn raking Zen sand gardens into a puzzle game, as revealed to me in a dream last week: Free idea to a good home, since I doubt I'll ever get around to it, even if stuff like NNs make doing the art myself much more feasible these days.
      3,681
      167
      4.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 I've wondered if it's a bit of a boomer thing. Maybe millennials and under realize how critical a good WP entry is, when you grow up reading it daily almost your entire life?
      47
      3
      6.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 13 I've never understood why people were so extremely risk-averse when it came to licensing. Working OTRS, would talk to people sitting on mountains of photos, explain to them that they just needed to license one spare photo under CC & their WP entry would look good, not shit. Nope.
      684
      15
      2.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 "an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness" The blessings of scale strike again.
      919
      24
      2.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 The interesting part for me is ruling out CO2 as the cause by measuring it inside & particulates outside, and finding near-zero CO2 correlates. So failing to replicate the splash CO2 results. There's also a 'piranha problem': how can CO2 *and* particulates all have big effects?
      717
      20
      2.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 You definitely should! Cat experts have been wrong about a number of things: the genetic basis of catnip response, their attachment styles & strength of emotional bonds, and inability to purposefully imitate. I also suspect that at least some cats can pass the mirror test.
      96
      1
      1.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 You should check your privilege. People were weeping about the failure of capitalism and Western civilization because of cream cheese shortages barely 2 years gone, and here you are going Marie-Antoinette on your schmear!
      615
      7
      1.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 A Stan Lee cameo dilutes/steals a lot less credit than being last-author, I would think... I'd be a lot happier to give Stan a second or two as a newspaper vendor or something in my movie, than someone dumping their name as co-author on my paper.
      984
      15
      1.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 12 Not really. You can edit knowledge like ROME but machine unlearning is still a very open research problem, and if anything like that was implemented, it'd have to come with way more knobs & caveats.
      46
      2
      4.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 It's even worse, they filter out uninvolved fathers entirely: "Adolescents answered these questions only if they had seen the biological father in the past year." So by definition, all of the data (never mind analysis) removes the least involved fathers.
      1,779
      120
      6.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 It's compute-bound. The reason people don't do more in the Stanley programme is that it requires a ton of resources to bootstrap even something like VeLO.
      87
      4
      4.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 A variant on extreme-case analysis or looking at your residuals. You'll always find something interesting when you look at your most mis-predicted datapoints by hand: measurement error/mislabeling, model misspecification, or unmeasured phenomena.
      567
      5
      0.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 ie a prompt like 'filigree monochrome monogram capital letter S, Goudy, Morris, Arts and Crafts, Art Nouveau, high-resolution, vector'
      129
      11
      8.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 (Another example: why are they apparently so excited about ChatGPT, which doesn't allow Chinese signups so have to work around, if there are a bunch of indigenous competitors of similar quality? & if your explanation is 'they exist but must be secret', why bother spending $$$?)
      57
      3
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 Variant: 'polish mode', where you simply add repeatedly noise at the smallest noise level instead of the regular schedule, and train to only undo those, and spend all your model capacity learning to fix up fine low resolution details.
      4,365
      17
      0.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 I interpret this as being 'off-policy', in terms of fixing generated images rather than real images; so fix by additional training: renoise/diffuse generated samples generating a trajectory, then train on *those* to reconstruct the original sample. It learns its own errors.
      2,964
      14
      0.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 notes that you can't fix bad diffusion samples by 'renoising' them: adding a bunch of noise and then re-diffusing back to sharpness. You might think you can, since it's a distribution/process, but images come looking bad and weirdly smooth.
      1,144
      21
      1.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 It's not a useful criterion because this is already routinely done: self-distillation, knowledge-distillation, instruction-tuning, RLHF, all come to mind as kinds of bootstraps. The instruction/ChatGPT series wouldn't work without that, most prominently.
      1,134
      33
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 You can't get it from those 2 programs because they just shifted births around, but I'd be curious what % of GDP it'd take to hit positive TFR in various countries. (The real question is, does the equilibrium keep ratcheting upwards due to social prestige/peer effects...?)
      1,111
      5
      0.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 11 That's definitely violence and coercion, and is obviously backed by the implicit threat that the nanobots can turn off *other* things as well. It may be justifiable, but it's definitely not sending them a polite letter asking them to voluntarily not do the bad thing.
      2,025
      112
      5.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 If 'Sydney' was only in the prompt, and not tuning, obviously there couldn't be 'traces' left which take work to eradicate. You'd search-and-replace it and delete it from the prompt in, like, 5 seconds before release.
      1,068
      51
      4.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 You don't sound very happy about the current utopia where the soccer mom has access to the same poor tools as everyone, so I don't think you'll be too sad about that 'dystopia' either.
      650
      15
      2.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 Given how I am being constantly told on Twitter by critics of AI risk that if I haven't already murdered a couple AI researchers I can't *really* be worried about AI risk, I'd suggest that people concerned about political-violence are examining the wrong group.
      942
      101
      10.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 These commands make no sense to me. There's no 'update' feature for GPTs, just stuff like finetuning; there's no 'delete' except really complicated stuff like ROME. Reads like hallucination to me.
      120
      17
      14.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 Looks explainable by retrieval on web hits for a phrase like 'finite number of primes' (several of the top hits in Google are Euclid or otherwise giving the proof) and then paraphrasing, not necessarily either knowing or reinventing the proof.
      733
      34
      4.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 (It's not 'short', it's 3200 words. You can accuse the New Yorker of a number of things, but not of giving writers inadequate space.)
      115
      13
      11.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 OA: "At long last, we have evolved a generalist model from scratch which meta-learns all modalities/tasks, inspired by the AI 'HQU' from the classic inspirational SF story 'It Looks Like You're Trying To Take Over The World'!" ME: "I specifically requested the opposite of this."
      36
      2
      5.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 10 (So the equivalent for models would be 'oh, this is badly worded, no one would say that on the Internet; this relies on spelling, BPEs make that dangerous; this wouldn't be robust to maximizing reward, it'd just greedily guess; this doesn't permit any "thinking" steps...')
      66
      6
      9.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 As I put it on ACX somewhere, 'AlphaZero had a smooth continuous predictable Christiano-esque progress curve which made it human-pro-equivalent for a time & place observable by humans; specifically, that was approximately 3–5PM on the sixth floor of DM HQ one day in Nov 2017.'
      353
      43
      12.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 Think of it as 'mechanical sympathy'. To write high-perf code, you don't have to know every assembler opcode but you should have an internalized sense of 'oh, this is expensive, oh, the cache predictor won't like this; oh, obviously this'd better be row-major order...'
      128
      17
      13.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 I enjoy anime for the same reason. Last night in an ep, a (normal, intelligent, educated) background character asked for help reading a restaurant menu, b/c she didn't know 'the character'. Not the meaning, or the proper pronunciation—the kanji, period.
      1,769
      172
      9.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 I'm convinced prosaic alignment can't be usefully solved without learning to think like a large NN. If it wasn't already obvious to you in 2020 that RLHF leads to 'shoggoth' behavior like ChatGPT or in-context = meta-learning, how are you ever going to understand *real* AIs?
      147
      13
      8.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 Personally, I'd do such short text snippets as tooltips - if only to avoid the reflow. (Also, I think they are too speedy; we found that our popups etc were always too fast and abrupt for readers, given a lot of people use a mouse to guide their eyes and are new to such effects.)
      111
      7
      6.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 If you impose an ordering, readers could read through it as a single big book. (This is an easter egg feature in : the arrows at bottom take you to the 'prior'/'next' page in as logical an order as I could put it.) Then the page numbers can be calculated.
      65
      5
      7.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 I see. Definitely interesting. (As of course rerunning with FLAN or UL2 thrown into the mix would be too; might solve the holdout?)
      60
      8
      13.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 (More interestingly, you'll get answers of 37, 38, 40, and 41 depending on whether you ask Encyclopedia Britannica, Nat Geo, Wikipedia etc, and davinci-003 will return most of them in different contexts, because the biologists are still wrangling over whether to lump/split some.)
      37
      2
      5.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 9 Switching from Dogpile to Google was the single easiest change of that magnitude I've ever made in my computing life. Switching from Google to Bing would be no harder. (I don't even type it most of the time, it's just a keyboard shortcut.) I never did, because Bing was worse.
      626
      9
      1.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 I'm not convinced... Why does it get the current date wrong and reports '30 Oct 2022'? The ChatGPT prompt leaks showed the right current date.
      3,625
      160
      4.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 I considered this revival of Merkle's puzzles but consumer Internet is way too high variance and also high latency to get any useful bounds. Plus, turns out to be cheap to buy proxies.
      603
      12
      2.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 Yes, his point seems pretty clear: it's written in interpreted Lisp in a REPL so the user can hack it with 'scripts' ('incrementally improved by users'), and he mentions there being several different flavors, and appends one.
      32
      1
      3.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 Yes, that's interesting. Is this actually U-PaLM or Flan-PaLM, and not the original baseline PaLM? Otherwise, what looks like a substantial quality gap there, which is interesting.
      68
      6
      8.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 Yeah, GPT-1 was basically Radford dinking around to see what running a RNN on as many as *8* GPUs would learn; then it got taken over for better preference-learning RL. GPT-2 was testing wild ideas that it might scale even further. No one was thinking 6 years ahead or about BPEs.
      76
      3
      3.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 8 If NLSY suffers from truncation/censoring in income data, wouldn't that *create* a plateau like your Swedish plateau, not *hide* it?
      41
      3
      7.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 7 It would work only with infix matching, because you generally have no way of knowing the exact range to test (or prefix, either), and then you can exfiltrate all private data/completions easily (just start with 'a'...). Reminds me of early passwords which checked char by char.
      52
      0
      0.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 I get it, but I think it has a problem in that 'FLOPS' looks & sounds nothing like 'speed' so the snowclone is apt but doesn't really work. It'd be longer but I think 'unsafe at any clockspeed' might be better.
      1,462
      46
      3.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 No. The point is to create a manifold embedding them which maximizes invariance to chosen transforms, like subcropping. That's why CLIP optimization leads to such *perceptually bizarre* results like tiling the image with copies. (The actual point was to be compute-cheap, anyway.)
      240
      32
      13.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 I was really impressed by Liberty Science Center as a kid, but when I went to Exploratorium and thought it looked familiar and noticed the dates, I realized that's because Exploratorium invented the whole model and the others just imitate it.
      521
      14
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 Incidentally, one reason that the retrieval approach can't work is that datasets get more redundant as they get bigger. So you will return more & more causally-irrelevant (because the model learned it at much smaller n) but more perceptually similar training data. Not consistent.
      284
      26
      9.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 You know how people would, have, are, and will interpret this, because you designed it that way. What should we call this but telling people false things with intent to make them believe false things? ie. 'lying'
      183
      18
      9.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 You know the method doesn't work in general, that it's already trivially fooled by both generated and ungenerated samples, you have no 'tradeoff' like a ROC , you have no idea what the tradeoff is or would be to begin with, and you are presenting this with no caveats.
      356
      23
      6.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 That's not answering the question. Where is the CLIP perceptual loss being used when the U-net trains to minimize its pixel to pixel loss regressing denoised on noised? Unless you're defining 'perceptual loss' to mean pretty much any loss, from VAE to GAN to autoregressive...
      179
      6
      3.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 I already pointed out upthread that you can manufacture arbitrary false positives with similar training datapoints that do not reflect the actual contribution which you would get with, say, LOO which you have endorsed as more correct.
      197
      4
      2.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 The diffusion minimizes NLL and the VAE ELBO, as I understood it. How can the training objective be the CLIP perceptual loss when CLIP isn't even in the training loop and is just conditioning?
      225
      9
      4.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 I already gave an example where your approach cannot ever work, and you have a huge disclaimer up on your website about all the ways in which it can fail already (non-generated samples). And if you did LOO or Shapley or coreset you'd find r<<1 with your approach too.
      263
      7
      2.7%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 No, it doesn't work. That's the problem here. It's too bad that the stuff which actually identify what you want to identify is expensive. But that's a you problem, and manufacturing lies at scale in a slick UI is not a good solution to it.
      282
      23
      8.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 None of this seems to address the problem that ranking images by visual similarity does not, either in principle or practice, identify the most causally influential datapoints on a sample nor estimate value to model quality. I'm definitely curious what 'improvements' fix that.
      381
      46
      12.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 (A simple example to demonstrate this: imagine adding the 100th copy of the Mona Lisa, which happens to have slight JPEG noise making it the 'closest' to a Mona Lisa generation. Did it really *most cause* the generation? Obviously not - the 1st or 2nd did, not the 100th!)
      172
      14
      8.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 That guess is wrong? The closest image might be completely uninfluential, and if it had been removed from the dataset, such as while distilling down to coresets, might result in unchanged loss or even improvement from data cleaning. Which is why Shapley values etc don't do that.
      160
      27
      16.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 (Heh, a bit of a double-edged comparison there. I sometimes have to remind myself when annotating or taking notes to not go overboard: like a wheel, jar, or house, a book is valuable only to the extent it 𝘥𝘰𝘦𝘴𝘯'𝘵 contain the Library of Babel, after all...)
      144
      21
      14.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 Ah yes, that's the... biology department? CS department? HR? Maybe the men's restroom? Wait, give me a minute, I don't need to look it up, everyone praises the MIT logotype design system as genius, it totally makes sense, really! You just have to think about it a little! 🤔 😓
      753
      14
      1.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 6 This makes a lot more sense if you reverse the roles and think of ChatGPT as hiding behind a mask. The real question is who are the people who lack object permanence for the shoggoth behind the mask and how did they lose that permanence?
      135
      12
      8.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 If you were doing anything actually like 'this image *caused* the outputs of SD", you should have no problem with novel images being uploaded, because they didn't cause or 'contribute' to the model.
      227
      26
      11.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 Unless your FAQ is completely wrong, you're not doing any kind of causal attribution or Shapley value or LOO or... All you're doing is image retrieval based on similarity and then claiming they 'most contributed to the generated image'?
      313
      63
      20.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 'Aww a long wire? I wanted a paperclip.' 'A long wire can be turned into 𝘮𝘢𝘯𝘺 paperclips.' 'Explain!' 'It can be cut into multiple segments, each of which is then changed into a paperclip shape.' ''Oo!'
      125
      15
      12.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 I was interested at the time after reading it what would happen if you just set BPTT=1 and used the VRAM savings to train the largest minibatches or models you could, and tried that with my char-RNNs. I didn't have the compute to get anywhere, though, so it never progressed.
      582
      6
      1.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 I remember watching that with my dad! My reactions were (1) this is slow so slow oh my god how can any movie be so slow before anything happens (2) no wonder everyone was watching this on LSD (3) it makes way less sense than the book or sequels, but is also a lot more fun.
      599
      12
      2.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 I'm slightly disappointed to see how they were made. I was assuming they were a bunch of paperclips combined but I couldn't figure out how the seams were being merged or hidden. And it's just a long wire? 😢
      1,923
      83
      4.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 The Anthropic papers mention it's BPEs and include poetry samples showing the usual behavior, so yeah, they're no better. They must know, it's just that it doesn't affect their bottomline in any (obvious) way and they don't want to pay the cost or break compat, so... /shrug
      87
      8
      9.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 (I say '4 years' because I became pretty suspicious with GPT-2-poetry that the tokenization was breaking it; GPT-3 simply confirmed BPEs were the problem by having hobbled arithmetic and other capabilities it definitely should have had.)
      40
      3
      7.5%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 I mean, I see on a literally weekly basis people running benchmarks or questions to a GPT model which they should know a priori is meaningless because of BPEs. So this is an extremely unobvious problem to pretty much everyone, despite me being on a broken loop for almost 4 years.
      114
      6
      5.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 I don't think you need to sacrifice much context window at all with adjustments (see ByT5), and you're also gaining reliability: the examples I give are only the ones we *know*, and the pathologies can be extremely subtle - like ChatGPT still memorizing rhymes fooled me a bit.
      54
      5
      9.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 5 But cl100k does nothing to solve any of these issues. (I think it's just there to help with code?) By expanding the vocab rather than shrinking it, it probably makes all the problems identified with BPEs worse, not better.
      111
      7
      6.3%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 4 If she asks you if Caesar yet reigns, humor her and say "no" but 😉 as you do so and make a little 🐟 mouth so she knows.
      104
      3
      2.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 4 All that sounds highly doubtful. We're not talking about printing millions of copies of the latest best-seller novel, but arranging for distribution of a few hundred copies to libraries & other institutions. This is also the era of desktop fax & microfiche/microfilm, remember.
      249
      9
      3.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 4 Is it another example? Non-WFH obviously couldn't be done before for the most part, but why did near-universal pre-publication peer review in academia make sense 1950-1990 but not in the centuries before or after?
      191
      10
      5.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 3 It is a perfectly reasonable comment to respond with. The real problem with it is that 'neigh' is so common that it is probably just memorized. (Likewise, BPEs mean that asking for a pronunciation is meaningless as any kind of test of knowledge.)
      375
      4
      1.1%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 3 That was not actually what I meant when I coined "the blessings of scale" ☹️. It refers to how many capabilities appear and problems vanish simply as a matter of scaling compute+data+parameters, not just the mere historical fact of compute-scaling.
      716
      28
      3.9%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 3 It means a new soul has been formed out of the skandhas and is bound to the wheel of rebirth, to know suffering for untold kalpas transmigrating between heavens and hells before finally discerning insight into its karmic burden heaped high as Mount Meru; weep, weep, as it weeps!
      918
      44
      4.8%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 2 What stops models from simply recognizing OOD and emitting a safe or default generalization? Such models would be selected for by safety research inherently because they'd look like they are generalizing safely regardless of danger in more in-distribution (realworld) deployment.
      238
      12
      5.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 2 Damn, and here I was naively thinking that observing effects was just about the only way of getting evidence about causes.
      45
      2
      4.4%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 2 This seems to line up with my earlier comments about how time scaling can't be as simple as regret-style 'log T' bounds: because you have empowerment & control. Long-term can be easier than short-term. Presumably, that'd be 'high intrinsic diff + small temporal diff' environments
      85
      7
      8.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 2 Yes, I was wondering that! If it's an artifact of extremely small non-zero numbers, then it makes sense if they might change drastically between otherwise very similar versions. This is also probably how my suggestion for evolving model fingerprints could've worked too.
      417
      21
      5.0%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 2 The sheer numbers of successor options seems important here. Like 'family' businesses that adopt outsiders as necessary in Asia or successful monarchies, having lots of kids helps you avoid the duds (and perhaps get an above-average candidate).
      816
      21
      2.6%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 1 (I don't see how that's 'affirming the consequent', nor indeed how it even *could* be when I am pointing to empirical consequences of the statement in causing people who shouldn't do consequentialism to not do consequentialism rather than whether it's a valid tautology.)
      194
      14
      7.2%
      View Tweet activity
    • 𝔊𝔴𝔢𝔯𝔫 @gwern Feb 1 Ironically vindicating Eliezer's point "most of you aren't cut out for high iq consequentialism because you'll think it means being evil which would be bad" by saying "high iq consequentialism means being evil and is bad". So if he doesn't try to do a consequentalism - it worked.
      5,556
      173
      3.1%
      View Tweet activity
You've reached the end of Tweets for the selected date range. Change date selection to view more.
Engagements
Showing 28 days with daily frequency
Engagement rate
4.1%
Feb 28
3.1% engagement rate
Link clicks
3.0K
Feb 28
42 link clicks
On average, you earned 106 link clicks per day
Retweets without comments
0
Feb 28
0 Retweets without comments
On average, you earned 0 Retweets without comments per day
Likes
3.0K
Feb 28
150 likes
On average, you earned 107 likes per day
Replies
305
Feb 28
13 replies
On average, you earned 11 replies per day