It's astonishing to see Apple settle for DALL-E 3 (or worse?! these remind me of the Bing samples) for the image generator part. Hasn't the incredible extent of mode-collapse and the horrible DALL-E 3 style become universally known and disliked yet?
35,337 karma · joined April 8, 2009
About me: https://gwern.net/me
What I've written recently: https://gwern.net/changelog
FWIW, I don't think CSS/JS have been much of a problem in terms of breaking. Before Said Achmiz got involved, while gwern.net was still a fairly simple static website, there were pretty much no serious instances of browsers breaking the CSS/JS. Everything ran fine. (Browsers are good at backwards compatibility, as long as you aren't pushing features too hard, as my site was not.)
The big problem was changing expectations and demands. 2010-era Gwern.net would continue to look fine and render fine in a 2025 web browser... as long as you were on a desktop/laptop. If you were laying in bed at night on a smartphone or a tablet, it would be both blinding and unreadable. This wasn't a problem in 2010, when <5% of my traffic was 'mobile' and mobile was very hard to design for even if you wanted to bother, but it is a problem in 2025, when >50% of my traffic is mobile. You can't write off >50% of your traffic.
People now assume that a website will render reasonably on a smartphone, and they increasingly assume there will be a dark-mode. But, even if you aren't using a legacy website based on Flash or tables, and are creating a clean greenfield HTML5 website, these are actually quite complex, demanding features!
You can get 80% of the way with some relatively simple tweaks, but then the rest can be almost arbitrarily difficult. Note that aspects of dark-mode can be subtle - while most websites have settled on the body class approach which avoids the 'blinding flash of white' problem, they continue to screw up dark-mode by treating it as a binary toggle between light vs dark, instead of a three-way toggle between auto vs light vs dark, and most dark-modes settle for fading out images instead of inverting vs fading them with a heuristic like https://invertornot.com/ . (We're still iterating on both features. Most recently, in November 2024: moving the scrollspy header to the bottom of mobile screens instead of top, and taking another dive into color theory in order to have colored link-icons which look good on mobile dark-mode.)
Compared to the burden of features like a mobile mode which is actually good, the regular level of web browser JS/CSS breakage is scarcely noticeable. (It tends to look something like, 'oh no, now in Chrome 123, there's a 1px gap in the dropcaps and the spacing constants have to be adjusted, THANKS GOOGLE'. Annoying, but hardly a major burden.)
Of course, you can quickly strangle yourself with your dependency stack, particularly if you get involved in NPM or you listen to the wrong web gurus, and your website will break overnight, and JS is definitely very fragile and rapidly bitrotting if you define 'JS' as node et al. But if you are being sensible about it and using plain JS (as we do) or well-chosen stable frameworks, of web dev's woes, 'web browsers keep breaking old working CSS/JS' is thankfully far down the list.
The big problems are self-inflicted ecosystem problems like bitrot/bitcreep (https://gwern.net/holy-war#bitrot), and the ever greater complexity expected of websites compared to the old 'preview in the only web browser anyone uses (ie. IE6), make sure it looks fine in the standard monitor dimension 800x600, export from MS Frontpage, ftp up to your shell account, done' days.
I have emailed the journal and told them about the bug.
In the mean time, you can read it in the PDF version with images intact at https://cdn.lanl.gov/files/arq-2023-1_14f30.pdf#page=37 (mirrored: https://gwern.net/doc/statistics/probability/2023-summerscal... ).
Which seems like it would be roughly equivalent to throwing away gradients if they are inconsistent...?
The narrator (and all other dogs) are blind to non-dogs as entities like them, so when the narrator sees a record with the logo prominently on it (https://en.wikipedia.org/wiki/File:His_Master%27s_Voice_EG_9...) being lifted or displayed, he (or rather the other dogs, whose owners are more musical aficionados than him) sees a soaring dog moving through the air, or being fixed in place on a wall or shelf, without visible effort or mechanism. The HMV dog on a record would indeed be about the size of a small dog's head, and given the perspective of the painting, the dog's legs are even tinier than they would be normally. And of course, printed images neither sow nor reap in any sense, despite their voluble yammering (of human or musical nonsense).
(Note that this further parallels the musical dogs and movie theater - a movie theater naturally suggests other sound-emitting technologies Kafka could be satirizing, like radios or... record players. These are dogs, in a musical context, 'performing' in a sense, but a cleverly different sense, so Kafka is not repeating himself.)
> Or again, the enigma of nourishment is easily solved when one understands that the dogs are being fed by an invisible hand, throwing scraps to hungry hounds. Likewise, the Lufthunde or air dogs are the pampered lapdogs of the bourgeoisie, toted around in well-to-do ladies’ arms, or nowadays in designer pooch purses.
While that is of course correct about the food, I think Kafka was not repeating himself, because the soaring dogs come with some hard-to-explain features like being highly voluble utterers of nonsense or the narrator saying he's never seen them. Why would a purse dog utter nonsense or the narrator never see one? It seems like because of the length of the soaring dogs passage, it has to be doing something more interesting than 'unseen human hands' again and have some more clever solution. Given that the whole point of the allegory/roman a clef is the narrator misunderstanding parts of the human world involving dogs and the challenge to the reader to figure out what is really being described, my belief is that the air dogs might be https://en.wikipedia.org/wiki/His_Master%27s_Voice satire - too well known to need any explanation back then, but now utterly obscure.
Note this is also used by Stanislaw Lem in a similar way in his SF novel _His Master's Voice_: the dog is confused by and unable to understand "his master's voice" coming from a machine, because such technology is beyond the capacity of a simple dog to understand in full, no matter how long the dog investigates the gramophone, even if some fragments of the sound make sense.
This seems to me to be a very Kafkaesque point to make.
So DPO vs RLHF is missing the point: the interesting thing here is how they are (presumably) generating the inner-monologue to fill in the gap between the Q and the A that you provide them, and then training on that augmented dataset of Q->solving->A datapoints.
Whether they are using simple finetuning on that dataset, or DPO, or RLHF, or something else, seems less interesting than the broader questions of, "does that work? and are there many important or economically datasets where o1 can 'fill in the gaps', creating a better annotated dataset, and bootstrap itself to be much more intelligent on that dataset?"
I can say that on my website, which is certainly 'richer' than most Markdown-oriented websites, we scarcely ever think about Markdown as a limitation to implementing a feature. Nor was it a serious hindrance at any point. There have been issues, sure, but pretty much never do we tear our hair out going, 'how can we possibly implement transclusion properly, when the source is written in Markdown rather than Asciidoc??? Woe is us!' That's just not a thing.
When we have issues, it's almost always related to something about the Pandoc internal API design choices (essentially unrelated to any syntactic sugar and would be equally true if we were starting from a different input type) or the sheer difficulty of implementing anything well & reliably cross-platform/device/mode. The Markdown part of our discussions usually boils down to a debate over the naming of a class. Because setting the class on whatever element is a non-problem, already solved, and we move on to the real challenges to adding a rich feature.
To whatever extent his imaginary gap exists in reality, I would bet it has far more to do with dynamics around websites which happen to use Markdown, like trying to cater to the lowest-common denominator and banning HTML fragments, and stuff like that.
I've been experimenting on my website with automatically adding to all years or dates a little subscript like '5ya'. While I thought it was clunky at first, I'm finding it helpful to be regularly reminded that when I mention something that happened in '2003', that really was '21 years ago', and I feel a little less like Billy Pilgrim.
So what I take TUOT as being is an exploration of the Idealism idea, where Borges puts a twist on it: the (dialectical) beliefs of the communalistic idealists of Tlön turn out to be true, on a certain level, because sufficiently compelling ideas and totalizing ideologies make their claims true. In that way, 'perception' becomes 'reality'. Only that which the ideology or state can perceive is real, and everyone is required to see like a state. (As much as he loved Idealism & Platonism, Borges always seemed to accept them only on a literary level, as applying to fiction and literature - there is indeed 'Man' in fiction, but there is not an actual Man in a Platonic region of forms, there is only a term 'man' we nominalistically apply to entities as convenient.)
That is, idealism is correct, in a sense, and the artifacts of Tlön become real because the savants of the conspiracy 'perceive' them (in their minds) and create them. And as Tlön takes over the world and gains power, it gains more realness and more of its artifacts come into existence - or people just lie about them or pretend they exist and falsify documents to accord with the new party line, and doublethink their way to 'seeing' the new labyrinthine reality forged by their fellow humans.
One might say that _hrönir_, especially, are a savage Orwellian parody of how things go in totalitarian dictatorships: the description of the experiments with the prisoners could as easily be set in Stalinist Russia or Maoist China, where the real story is that on the fourth try, after turning up only the equivalent of fishing for a muddy boot, everyone has figured out that, to satisfy the decrees from above, they need to buy or forge some ancient artifacts of unconvincing antiquity (and so no counter-revolutionary skeptics can be permitted near) and that is how _hrönir_ are discovered. The same way Lysenko manufactured agricultural miracles or innumerable falsifications like https://en.wikipedia.org/wiki/Learn_from_Dazhai_in_agricultu... became official policy, doubted only on pain of death.
Those who disagree and wish to maintain their integrity, can only retreat into quietism or 'internal exile', and spend their time on topics with as little political relevance as possible and avoid even publishing (except as samizdat), and let "a scattered dynasty of recluses take over", as it is too late to stop the Tlön revolution, and "the [whole] world [will] be Tlön".
Not really so naive when you look into 'robotic process automation' https://en.wikipedia.org/wiki/Robotic_process_automation . Huge industry, and can be just moving a mouse around on a screen and triggering keyboard inputs. If robot hands were cheap and dexterous enough...
'Chess' is not a standard LLM benchmark worth Goodharting; OA has generally tried to solve problems the right way rather than by shortcuts & cheating, and the GPTs have not heavily overfit on the standard benchmarks or counterexamples that they so easily could which would be so much more valuable PR (imagine how trivial it would be to train on, say, 'the strawberry problem'?), whereas some other LLM providers do see their scores drop much more in anti-memorization papers; they have a clear research use of their own in that very paper mentioning the dataset; and there is some interest in chess as a model organism of supervision and world-modeling in LLMs because we have access to oracles (and it's less boring than many things you could analyze), which explains why they would be doing some research (if not a whole lot). Like the bullet chess LLM paper from Deepmind - they aren't doing that as part of a cunning plan to make Gemini cheat on chess skills and help GCP marketing!
Regardless, I am puzzled how OP got this URL in the first place. He wasn't supposed to, he was supposed to get the canonical Arxiv PDF link. Because this is one of the cache mirrors/local archives†, rather than a regular hosted document. We block everything in /doc/www/ in robots.txt & HTTP no-archive/crawl/mirror/etc headers, and we use JS to swap out the local URL for the original URL whenever the reader clicks or mouse-overs or interacts with a link to the URL in a web page (and that is the only place they should be publicly listed or accessible). If OP read it on gwern.net by seeing a link to it, and he wanted to copy the URL elsewhere, he should have just gotten the canonical "https://arxiv.org/pdf/2108.07686#page=85"... But somehow he didn't.
OP, do you remember how exactly you grabbed this URL? Is this an old link from before our URL swapping was implemented, or did you deliberately work around it, or did you find some place we forgot to swap, or what?
(If anyone is wondering why I mirror Arxiv PDFs like this in the first place: it's for the PDF preview feature in the popups. Because Arxiv blocks itself from being loaded in iframes we need local mirrors for PDF preview to work at all; local mirrors save a new domain lookup and speeds up the PDF preview a lot because we compress the PDF more thoroughly and Arxiv servers are always overloaded; and because readers can potentially pop up many Arxiv PDFs easily, it saves Arxiv a lot of bandwidth and avoid burdening their servers further, so it's just the responsible thing to do.)
So there is a natural way to not just use a minimal bit or byte level tokenization, but every tokenization simultaneously: simply define your dataset to be a bunch of datapoints which are 'start-of-data token, then the byte encoding of a datapoint followed by the BPE encoding of that followed by the WordPiece encoding followed by ... until the end-of-data token'.
You need not actually store any of this on disk, you can compute it on the fly. So you can start by training only on the byte encoded parts, and then gradually switch to training only on the BPE indices, and then gradually switch to the WordPiece, and so on over the course of training. At no point do you need to change the tokenization or tokenizer (as far as the AUNN knows) and you can always switch back and forth or introduce new vocabularies on the fly, or whatever you want. (This means you can do many crazy things if you want. You could turn all documents into screenshots or PDFs, and feed in image tokens once in a while. Or why not video narrations? All it does is take up virtual indices, you don't have to ever train on them...)
Could you explain further what part of https://gwern.net/newton you thought didn't support my description of Newton's view?
I thought the large second blockquote in https://gwern.net/newton#excerpts , which very prominently links to https://www.newtonproject.ox.ac.uk/view/texts/normalized/THE... , fully justified my claims, which are taken directly from Newton's statements to his son-in-law, and also closely parallel other historical statements, like Lucretius, which I also present with clear references and specific blockquotes.
I'm a little mystified that you could describe any of this as not supporting it at all, and I'm wondering if you are looking at the wrong page or something?
If you only need a few terabytes, you can rent from Hetzner for more like <$20/month. Maybe <$10 at this point, if you are patient. And you can use the server for other things too, since it's a fixed cost, like your website or encrypted backups or other files you host. I spent a while figuring this out before I decided to do Danbooru20xx, to make sure it was very cheap. Given how my website has grown in size, and how absurdly exorbitant cloud bandwidth and object/hard drive space is, the Danbooru dataset was practically free at the margin!
> Also, I thought these lesswrong folks were all about "effective altruism" and "earning to give" and that stuff.
Some people are, some people aren't. I'm not. (I have intellectual sympathy with EA, but not emotional.)
You only have to go as far as 'CQK' before there is, AFAICT and based on Wikipedia entries, essentially no even slightly important use of a TLA. There are then another >2k more or less unused, and if you go to four-letters, you've got ~390k to play with.
It's a pity that the gold leaf usage probably looks far better in person, though, compared to a digitization. Gold tends to be a material that doesn't translate onto computer screens well, like Rothkos.