gwern
Hard to say. Oyster larvae are highly mobile and move their bodies around extensively both to eat and to find places to eventually anchor to, but I don’t know how I would compare that to spores or seeds, say, or to lifetime movement; and oysters “move their bodies around” and are not purely static—they would die if they couldn’t open and close their shells or pump water. (And all the muscle they use to do that is why we eat them.)
Anthropic doesn’t delete weights of released models
How do you know that? Because OpenAI has done that.
Fulltext link: https://gwern.net/doc/statistics/bayes/1963-mosteller.pdf since it doesn’t turn up in GS. (One of many papers I host with clean metadata and easily scraped and listed in the sitemap.xml for a long time which still doesn’t turn up for unknown reasons.)
Isn’t that just mode-collapse?
“Tilakkhana”, Gwern [poem]
Plants have many ways of moving their bodies like roots and phototropism, in addition to an infinite variety of dispersal & reproductive mechanisms which arguably are how plants ‘move around’. (Consider computer programs: they ‘move’ almost solely by copying themselves and deleting the original. It is rare to move a program by physically carrying around RAM sticks or hard drives.) Fungi likewise often have flagellum or grow in addition to all their sporulation and their famous networks.
There’s no reason whatsoever to think that they are independent of each other. The very fact that you can classify them systematically as ‘positive’ or ‘negative’ valence indicates they are not and you don’t know what ‘raw chance’ here yields. It might be quite probable.
I don’t know what the samples have to do with it. I mean simply that the parameters of the model are initialized randomly and from the start all tokens will cause slightly different model behaviors, even if the tokens are never seen in the training data, and this will remain true no matter how long it’s trained.
Or it could just be randomness from initialization which hasn’t been washed away by any training. “Subliminal” encoding immediately comes to mind.
but I don’t think it would extend lifespan indefinitely.
I think the big question about head transplants for me is how it interacts with the process of aging, given that body systems are heavily intercorrelated and damage in one system will damage other systems (and vice versa with healing).
Intuitively, I would expect it to work statistically in one of two ways. In the pessimistic case, the brain keeps aging on the normal Gompertz schedule, and even a perfect head transplant can never buy you more than a decade or two before any of your head-located diseases kill you; it corresponds to shifting the aging mortality spike down somewhat, but it is still going to spike soon. (In this scenario, the trauma and cost of a head transplant might rule it out for a lot of elderly people: you might die on the operating table or afterwards, and it doesn’t buy you much time.) In the optimistic case, the near-reset of the body to the pink of youth will eliminate all of the aging acceleration from the rest of the aging body and heal the brain and by changing the exponent, potentially buy you drastic increases in lifespan by flattening the spike, and 50 or 100 years might suddenly be statistically plausible. At which point you might as well consider it an indefinite lifespan extension because who knows where medical technology will be in 100+ years?
In Accelerando, it’s abundance of physical compute which carries transhuman society as a whole beyond human comprehension. In A Fire Upon the Deep, it’s some kind of algorithm which makes the difference—when that algorithm runs, unstoppable superintelligence is created.
I think that’s not quite right? The Blight can only run when the Straumli researchers break open the sealed archive because it’s located in the almost-highest-physical-computation-possible zone, the Low Transcend. The Blight can reach into the lower zones but only after running in the Low Transcend ie. there is an abundance of physical compute which carries the Blight beyond human society-level comprehension.
they’re wearing things which extras with no lines would wear in the background in a movie. The things I wear look like, at minimum, a named character.
This is true.
The first time I mentioned “John Wentworth” to someone at Lighthaven, they asked me: “who’s that?” I said, “you’ve seen him around, he’s that guy who looks like The Blues Brothers.” “Oh.”
Is that frame-shift error or those ~6 (?) SNPs previously reported in the literature for anything, or do they seem to be de novos? Also, what WGS depth did your service use? (Depending on how widely you cast your net, some of those could be spurious sequencing errors.)
‘Repetition’ is certainly a drawback to the ChatGPT style: we have lost em dashes and tricolons for a generation. But it can’t in its own right explain the reaction to the SD image, because… ‘German Expressionist linocut’ just doesn’t describe a default, or even a common, output style of any image generative model ever. (That’s part of why I like to use ‘linocut’ as a keyword, and for better or worse, people who might reach for ‘German Expressionist’ these days typically reach for Corporate Memphis instead.)
It could however be a kneejerk reaction: “oh no, this is a generated image, therefore it is exhaustingly overused and boring [even if it isn’t actually]”.
So, most LLM RL training would be expected to exacerbate this issue?
I think that it’s a pity if people write off my SD page because they failed to understand the meaningful illustration I put effort into creating and didn’t, say, check the alt text to see if they were missing something or wonder why such an unusual website would have “AI slop”; and I agree that this may be a case of “things you can’t countersignal”.
However, I refuse to submit to the tyranny of the lowest common denominator and dumb down my writings or illustrations. I don’t usually write for such readers, and I definitely do not write my Gene Wolfe essays for them!
So unless people can point to something actually bad about the illustration, which makes it fail to satisfy my intent—as opposed to something bad about the readers like being dumb and ignorant and writing it off as “AI slop” when it’s not—then I decline to change it.
The image borders are exposed, and the white fade does not work. Compare to the fully functional light mode:
To be honest, I don’t like the light-mode example either. I think it’s bad to have your text visibly overlapping like that. (If I made an image which I had put
.float-right .outline-noton to get the same effect on gwernnet and I saw your ‘good’ white-mode version, I would be immediately complaining to Obormot about his CSS being busted.) So in this example, isn’t a lot of the problem that the UI elements are overlapping so the text is spilling over onto the scales and blocking the wires etc, and the dark-mode merely exacerbates the problem and fixing the core problem would also fix the dark-mode issue?
FWIW, my impression is that while the gwern.net dark mode was a ton of work to create, due to requiring a total refactoring of the CSS and deal with the flash-of-white issue and figure out the correct three-mode behavior, and create tooling like the color conversion script or needing to create InvertOrNot.com for acceptable images, or tweak the syntax highlighting colors to look right in dark-mode… But once we finished paying all of that, the maintenance burden has been relatively small. I have to patch up the occasional image not inverting right and we have to occasionally special case some Wikipedia popups CSS, but it’s on net less effort than much of the website (eg. Apple device support, or the local archive system, are perennial hassles).
The gwernnet dark-mode is pretty good engineering work overall. We should’ve done a design-graveyard writeup to explain not just how it works (non-obvious) but the problems with the more obvious approaches and what problems we had etc. I fear it might be too late to write it now even if we had the time...
I don’t believe it is “AI slop”, much less that it is “pretty nasty”. I consider AI slop to be low-meaning and low-effort generative media which adds little or nothing to the experience
I assume you are referring to the German Expressionism, alluding to Nosferatu (which is highly relevant for at least two reasons), image illustrating the narrator’s childhood iceskating in a New England Protestant town in decline due to Dracula taking it over; I generated it in MJ after cracking SD, to sum up the horrifying reality of my solution. I put several hours of thought and effort into the concept and creating it, and got what I wanted, so I think this is just a case of de gustibus non est disputandum. I felt it cleverly visually encapsulated the mood of the horror that Gene Wolfe meant to lurk underneath the harmless nearly-bucolic appearance of SD and enhanced the experience.
So I think it satisfies my 3 criteria: it is not low-meaning, was not low-effort, and adds something. But I don’t think this is a good place to discuss it, so I have added a more detailed discussion of that image’s process & meaning to my image slop blog post as an example of how I think I get good image samples.
EDIT: I would be curious about the disagrees. What, exactly, are you disagreeing with? Do you think I am lying about the creation process, the prompt, or the meaning? (I would point out that there was already a short version of this description in the alt text, and has been since I added it in the first place c. November 2023.) Do you disagree that the high concept reflects my SD interpretation? Or what?
The genre here is psychological horror fiction, and the style is first-person short story; so it’s reminiscent of Edgar Allan Poe or Ted Chiang; but it’s not clearly condensed or tightly edited the way those tend to be, and the narrator’s style is prolix and euphuistic. From an editing perspective, I think the question I would have is to what extent this is a lack of editing & killing-your-darlings, and a deliberate unreliable-narrator stylistic choice in which the prose is trying to mirror the narrator’s brute-force piano style or perhaps the dystonia-inducing music itself (because it can be hard to distinguish between deliberately flawed writing and just flawed writing—especially in the vastness of the Internet where there is so much flawed writing). I expect the latter given Tomas’s other stories, but I think it’s not sufficiently well done if I have to even ask the question. So that could probably be improve by more precise writing, or more ostentatiously musical structure, or some carefully chosen formal game; but also one could imagine making much more drastic revisions and escalating the horror in a more linguistic way, like a variant in which the narrator suffers an aphasia or dyslexia from reading a pessimized piece of text (revenge?) and the writing itself disintegrates towards the end, say.