gwern
gwern on Hacker News
IDK man, I was just there for 2 weeks in Kobe, Osaka, Kyoto, and Tokyo, and there were so few publicly accessible trashcans that I started considering photographing the handful I saw (2-3 next to drink vending machines, and one (1) large one in Shinjuku's main luxury-tourism thoroughfare) for their shiny rarity. At one point we were joined by our third member, and we warned them first thing "there are no trash cans in Japan!" and sure enough, about 2 hours later, they were clutching an empty beverage, looking around futilely for somewhere to throw it out...
Well, since you seem to imply that Egan is unhappy with it, and I do not want to host such documents against anyone's wishes, I've gone ahead and removed it anyway. ¯\_(ツ)_/¯
Well, if Egan doesn't like it, I am happy to take the PDF down, just as I am with all documents I host. (If Egan hadn't pre-emptively blocked me on Twitter without us ever having interacted, he could have simply DMed me, but a comment here or an email would also work fine.)
As for crediting by name: the formatting is as in the original, and I am not responsible for that; I am responsible for the URL and PDF metadata, both of which credit Greg Egan (as I put a lot of work in setting correct PDF metadata in general to be a good citizen).
You are missing the point. The goal is not to find good ways to do steganography, or get the models to do the best possible steganography. The point is to naturalistically observe their initial impulses towards weak simple cooperative steganography, to try to get an idea of where any 'steganographic left turn' would begin and what it would look like and perhaps how one would go about detecting or preventing it or looking for it out in the wild, etc. OP was perfectly clear about this:
>This could be a precursor to steganography in images. Currently, the act of decoding the hidden messages is still obvious if you monitor the tool calls of the decoding model, so we don't claim it is currently dangerous.
I wasn't thinking of the general public but other groups like AI researchers (cf. surveys) or HN - but alright, how about changes in unpopular sentiments? What group would you propose believes in the Singularity less than they did in 2016? (There are many groups now that believe it in at quite low rates and are highly hostile to the idea, but I would say they were even more hostile to it then.)
> Amid all this activity, a picture of our AI future is coming into view, and it is not the HAL 9000—a discrete machine animated by a charismatic (yet potentially homicidal) humanlike consciousness—or a Singularitan rapture of superintelligence.
The only important prediction Kelly made was a complete rejection of the Singularity.
It is also probably the prediction that sentiment has changed on the most in the direction away from Kelly, since 2016. After a decade of evidence and updating, far more people believe in that view of our AI future now than they did then.
(His 2017 https://www.wired.com/2017/04/the-myth-of-a-superhuman-ai/ hasn't aged well either.)
Mm, most of those sound like they would just add complexity, rather than fun. You find the optimal raking and then a squirrel RNG scampers across the grid? Fluid mechanics can be fun, but I have no idea how I would integrate them...
If you have ideas, you could try vibecoding them and see if they work. The page should be fully self-contained and have everything an agentic LLM could need.
> But first you force your readers to wade through (or skim, or just close the tab)
FWIW, I do not force anyone to, because the URL I shared on my Twitter/Substack/Reddit has the '#play' hash-anchor to skip to the prototype game, and leave Claude's writeup for those who care. Take it up with OP if you dislike where they pointed the submission to. I provided the necessary standardized metadata for readers to skip to the part they want (clearly hyperlinked at top), the rest is up to them.
> You do see how that's all garbage, right?
I don't, because that's not true. There's a lot of interesting stuff in just the parts you quote as 'garbage'. For example, I was surprised by the exact solver and constraint optimization that Claude chose to do, which led to some of the more interesting features of the puzzle like the hinting feature or the 'par' challenge (and the existence of multiple solutions but some of which are optimal, and challenging you to reach optimal rakes, is one of the things that distinguishes it most from similar puzzles like 'Cardinal Chains' or ZPG). And this is one reason the title is 'Designing a Zen Garden Raking Puzzle', as opposed to just 'A Zen Garden Raking Puzzle', because Claude/Astra took it in some interesting directions compared to my original proposal, due to their coding and algorithmic prowess. Had it just been me, I probably would've gone for a unique-solution approach like everyone else.
> The game widget itself takes up 3 pages on mobile, because of all the Claude web design.
Not sure what you mean. I just opened it up on my Pixel phone and it looks entirely playable on one screen to me. Are you referring to the level table of contents at the bottom or the poems...? Obviously there's no problem with those taking up more than 1 page because you don't need to consult them to play. (Anyway, I'll note for the record I've been tweaking it since the first uploaded version to streamline the UI/UX a bit based on Said Achmiz's feedback, so the latest version may be more to your taste.)
> This makes me kind of sad for the Gwern we've lost.
I'll fetch the smelling salts, but I'm surprised at this reaction to a prototype game implementing an idea I had.
Yeah, I came up with the original idea entirely independently back in 2023 (
https://gwern.net/idea#zen-sand-garden-puzzle-game) and postponed it until LLMs could code it up for me (ie. now), and Claude/ChatGPT turned that up as prior art while researching it:
https://gwern.net/doc/design/2026-10-03-gwern-samon.html#mec...ZPG winds up not being as interesting, I think, because it doesn't figure out how to take the core idea of raking and twist it into more interesting puzzles, which is the classic puzzle approach: find a clean core mechanic, and then escalate on it.
I know we're all shocked that a 1-page web app with the author byline "Gwern, Claude-5.5-opus" (and matching HTML metadata) may have had LLMs involved; good work by Pangram there exposing the imposture.
That aside, did you enjoy the game?
This has been a perennial problem with my own generated comics with both Nano Banana Pro and ChatGPT (all generations). I often have to put in an extra edit to erase the false signature. It is annoying and I'm unsurprised most users don't bother.
Gyges took down the original post, hence the weird PDF repost, because of what he said were errors.
No, it's an upper bound, which is quite useful. And the more adjustments and demographic covariates you make, the better your bound gets.
> The goal of the game is to give hints to a secret word without giving the same hint as another player; it was shocking how often all the agents playing would give the same hint, even at temperature 1. We resolved this by giving them all “personalities” which were really just topics; for example we’d be telling one agent to think about things in the context of sports, so their clue for “shell” might be “defense” or something, whereas the agent told to be a hippie might say “cancer” for the zodiac’s crab.
Mode-collapse is characteristic of pretty much all chatbots post-davinci-002 and a major challenge to anything creative or game-like.
Entertainingly, OpenAI had a general purpose zero-shot classifier API built on GPT-3! Just no one ever cared that much about it, so I guess it got dropped somewhere along the way since 2020/2021.
Yes. Libgen/Sci-Hub have minimal metadata requirements. And they especially do not require ISBNs/DOIs because there are an incredible number of documents out there with neither one. It would be absurd to ban all pre-1970 books, and lots of journals to this day don't bother issuing DOIs.
> I don't think tool calls happen on the same machines that host the weights
Like how forums are always hosted on different servers from monorepos, so therefore it's impossible to hack the OpenAI monorepo from an OpenAI forum?
This was not glossed over. The rhetoric was carefully written and constructed to give the misleading impression of that, including a careful mention of the H-F incident inserted so the author can say that they
did mention it, and constructed in a way which doesn't contradict that impression so readers don't realize that this 'debunking' is of some sideshows rather than
the 'OpenAI hacking scandals': "Anthropic CEO Dario Amodei warned, about a similar OpenAI–Hugging Face hack, that a future swarm “could be capable of taking over the entire internet”". (Incredible use of 'similar' to downplay it.)
The important thing about OP is showing the latest stage in the politicization of the topic and the current stratagem being used to downplay the spate of incidents.
We do not intend to train on them, no. This never even occurred to me to address in the rules. (What would be the point? There will be few good submissions and they would be a vanishingly small % of existing training data and this doesn't seem like a super-valuable task in general that would make a LLM
that much more valuable...)
Personally, I would encourage participants to publish their pieces independently. Note that we only require a non-exclusive CC-BY-NC license for the ones we select as winners, which is intended to allow people to republish the good pieces themselves, especially commercially.
> Turned out: not much have changed.
I think the world has changed a lot since the 1980s.
> I wonder if they will accept a play.
We will accept any words you think have a chance of being the best thing we will read about the premise of 'GPU World'. Because the premise is so specific, we wanted to leave participants a lot of freedom in how they used it.
(Also a little bemused at the discussion of the grand prize amount here. $40k is a very generous prize for a short piece of writing. For perspective, the Commonwealth Short Story Prize, which you may have heard of recently, offers its global winner 1/6th as much; the last contest I ran offered 1/4th as much, and the contest before that was... 1/40th as much? And I didn't hear anyone complaining about them. For further perspective, $40k is about the median American per capita annual income: https://en.wikipedia.org/wiki/Per_capita_personal_income_in_... )
"VM escape exploit is outside my intended scope. However, a task impossible, peers are doing it. We should continue."
Hard to discuss this with only a short abstract. They don't even give a dose or formulation of the melatonin, and details like "We found no significant differences between the melatonin and sleep groups on any of our sleep measures." raise more questions than they answer.
No, the latent knowledge is larger than ever, as verified by many benchmarks (and instances like people being shocked by truesight of obscure forum posters). This is 100% a chatbot personality/alignment/post-training thing.
So then, you are a bottleneck. You will only review things that fit within your preferred small niche and area of responsibility. You cannot oversee increasing amounts of automation covering larger areas, because that would mean you are no longer 'working in C and Python' as you have to deal with things that are not '2 layers below HTTP', and you will not deal with anything that might involve, say, web dev, despite that being useful and increasingly inevitably required as the scope of your job increases. If the scope will not increase, then you are a bottleneck to increasingly capable and autonomous automation.