A couple of years ago, after reading the book, I sketched some of them on Procreate with an isometric grid, but not all 55. I only arrived at number 4, since each took multiple hour...
It’s interesting, with a human made piece of art, I’m always drawn to « zoom in », to see how you did little details, if you have hidden something here and there, how you connected different parts.
With AI generated stuff, sure it looks amazing at first glance but I definitely don’t want to zoom in, since I know it’s all a cardboard facade with no passion. Look a bit too closely and you’ll see the bridges that make no sense, the bunch of nonsensical threejs cylinders etc.
I always consider AI-generated content to be a draft — or, in its parlance: something that takes the shape of the final product.
I don't dismiss the ability of models to create. I dismiss their ability to create finished goods. Humans should always be involved, especially with creative works. Models are great at modeling existing creations and solutions — head starts on long roads — but it's the person that needs to walk it and take the journey.
Since this project (both yours and OP's) seem more for personal fulfillment and enjoyment, I wonder who got more enjoyment out of the process, and also who will remember the results better in ten years.
> Leaving there and proceeding for three days toward the east, you reach Diomira, a city with sixty silver domes, bronze statues of all the gods, streets paved with lead, a crystal theatre, a golden cock that crowns each morning on a tower...
Text in the second person is causing my my D&D instincts to kick in.
I understand why that’s impressive. But I don’t feel much. I guess, I was just used to envy how much people can be absorbed by some idea to spend so much time implementing it. Now it feels like a single shotted nice-typography-and-all presentation at work that someone did 10 min before the meeting.
There's also an eerie feeling of emptiness to the outputs. They lack real intent. Glossy and technically impressive, but they don't tickle your brain. I think it's because they don't really communicate or give you that buzz of new insights. They are pretty but not _helpful_.
For example, I didn't learn about lenses or earthquakes or fusion messing with the controls on these.[1] I didn't know what to look at.
However, there were interactive exhibits at the Children's Museum I spent a long time exploring and loved. That's because they were carefully designed to really teach the concept?
"What we obtain too cheap, we esteem too lightly: it is dearness only that gives every thing its value. Heaven knows how to put a proper price upon its goods; and it would be strange indeed if so celestial an article as freedom should not be highly rated." – Thomas Paine
I've been thinking of this line recently — really gets at the heart of the matter for me. There is no substitute, no shortcut, no alternative for effort.
It's low entropy. The amount of entropy is prompt + model but now that you have seen the model various times, the entropy comes down back to the prompt only.
You'd not be amazed by seeing a 2 sentences prompt and thus you are not amazed by this work either.
Yeah, the demos look really cool but you're thrown into them and there's nothing explained as to what you're looking at or why you should care. Funny, an article explaining something like how a camera focus works with a few illustrations would be far more useful than a fancy 3D interactive.
It turns out the more technology you throw at something does not equate to ease of learning.
If it makes you feel better.. I find this impressive but it doesn't pique my curiosity. I may get biased, finding a reason not to care.. but honestly I think if this were presented differently I would. And it's not so much about using AI either. I think you could use Ai to make something compelling but so far the one shot prompt outputs to me lack focus
Isn’t that at least in part because the intention here was to demo the functionality?
Put in the hands of the educator, they will be able to add those illustrations, design for the light bulb moment, etc, and the tools at their disposal just got much better.
Another thing to keep in mind is that if you asked twelve different artists to render these cities from their descriptions, you would likely end up with twelve very different renditions.
Now imagine the same thing but each person is using Opus 5.5 - I’d wager there would be a lot of commonality among each of the outputs.
> If something that previously took weeks for a human to design, now takes 10 min, that means a whole week of salary went up in smoke.
To me, this is the depressing sentiment. The idea that, somehow, the goal isn't to produce useful things but to spread work out over a period of time in order to maximize payment.
At a high level,I conceptualize my job as "doing useful things." Whether that's design, writing code, debugging, sysadmin, CI/CD, writing doc, whatever. I just want to be useful. If there's some tool that makes that easier, I am happy about it. I don't think that tool will result in my getting paid less. But if it does, oh well.
>But if it does, oh well.
I'm glad our ancestors didn't take that line of reasoning we'd still be feudal serfs. "The lord has come and taken all our grain again? Oh well, at least I've been useful to him."
If our ancestors worried more about technology obsoleting human labor we'd still be subsistence farmers. Priorizing usefulness is why we're rich today.
Weird way to look at it, from another view we can now print a week salary in 10 minutes. Might devalue the output slightly but we still have the product.
Undoubtedly a fundamental reconsideration of political economy is necessary if we are going to go all in with AI. As of now, this technology will exacerbate the existing 'singularities' and wealth disparity distortions in the socio-economic body. And now, basic fucking income is not the answer. Maybe pitchforks.
If you haven’t read the book yet, I’m tempted to warn you against opening the visualization. Part of the joy of reading this book is letting your mind’s eye run wild, and I fear if I read it for the first time after toying with this site, I’d be remembering the cities rather than conjuring them.
That is the exact problem with the city of Dinipro.
Citizens of this city high in the sky imagine it so fully in their minds that no one there needs to open their eyes to walk the streets and live their normal lives.
But if one makes the mistake of opening their eyes and tries to see the city directly, they would immediately fall through the clouds to the ground deep below.
I heartily recommend the audiobook, narrated by John Lee.
The book is a trip and the dreamy sort of narration matched it perfectly. My wife and I would spend evenings listening to one city at a time, on a walk around our little park, one headphone in each ear. I don't know if it was just the environment or the time in our lives, but we both really look back fondly on this book.
I read the book decades ago, then I lost it in one of the many relocations. I always thought I should buy it again but never acted on the idea... A few years ago, I was walking by an antiquarian, and noticed a thin unmarked gray spine. I entered, picked it up, paid, and walked away.
PS I didn't open the gallery. Nice idea but I don't care about visualizing something I should feel.
How Opus 5.5 imagines these cities is a question for Opus. The real experience is how we imagine - and I believe there are no two people who have exactly the same mental image of an invisible city.
> I’d be remembering the cities rather than conjuring them
Having looked at some of them (and it being one of my favorite books) ... it won't do that. These visualizations are as terrible as they are unnecessary.
Maybe the interactive link in the article for the optic focus can lead you on a fun journey instead. And the bottom-right of every visualization is a link to yet another fun diorama to explore.
It's a techie creating the torment nexus but the imagination version. No need to picture in your mind the invisible cities! Time to make them visible!!!
Talk about missing the entire point of the piece of literature.
Invisible Cities is my favorite piece of literature hands down. The audiobook narrated by Richard Higgins is also done so beautifully and can't recommend it more.
As neat as this project is, this does no justice to whatsoever to the book. On the surface Invisible Cities seems to be a book about random cities' descriptions. In reality though, this book is about semiotics, meaning, language, and it's limits.
The first city I clicked on at random had the opening phrase:
> Arriving, you rejoice at its bridges, each different
The model city had like five bridges total, three of which were didn't bridge anything at all but instead sat in the middle of the river and went along its flow, the "bridges" for some reason pretending to be small islands.
Kublai Khan received an ornate letter signed by Marco Polo: "In Madrid, City of Lost
Things, no item remains where it was set. If one drops his key in the dirt, he may
never re-enter his home, and, even if he manages to stoop and recover the key, he may
rise to find a tulip garden where his house once stood. In complementary fashion,
things lost by others are forever turning up: A pocket watch on a coffee table. A fond
memory in your recollection. I even know of a prince who turned up in a prison cell.
When he appealed to the guards for his release, he failed to find the crown on his
head, and when he was asked his name, he searched his thoughts, but could not find it.
Indeed, the only hope now for the release of this prince of Spain is for you to send
back 300 ducats for his release. Of course, he will reward you handsomely once he is
out. Yours truly, Marco." Kublai Khan cocked an eyebrow and declared before his court,
"Hey, everyone: Looks like we're about to get ripped off by the guy who traded gold
for paper!" The court erupted in booming laughter.
— Italo Calvino, Invisible Cities 2: This Time It's Visible.
Like many other people in this thread, I feel like passing what would essentially be a hobby project off to Claude robs the creator of most the fulfillment and gives the viewer less reason to appreciate it. But with the models being as good as they are today, I genuinely feel conflicted. If your end goal is to share something with others, why spend a weekend on a project like this when you could just hand it off to Claude?
Your appreciation should come from the one shot prompt. It's a direct comparison to Astra's one shot. I feel like you and many others are trying to "appreciate" the wrong thing.
If the creator truly wanted viewers appreciate the end product they'd start with the one shot output and spend their weekend iterating on each city either by hand or with Claude. But they didn't want to do that, they wanted to show off what Claude/GPT could do in 6hrs. That's the whole blog post.
This isn't a ceiling, though. This is the new floor: What you can get with more or less zero human effort and ~$100.
I expect we're going to see more and more amazing "hobby projects" that did involve both a few hundred dollars in tokens and hours of human creativity and effort in combination.
Why would that someone want to listen something you haven't created? What have you "shared"? Claude could have made the tools that let the creator build their cities from their own imagination, but that's not what this is.
One of my favorite things about the National Treasure Cinematic Universe (two movies and a TV show) is the existence of characters that I like to call "treasure grumps". Their role in the story is to warn the main characters that searching for treasure is a terrible way to spend your time, and can only lead to personal and familial ruin.
I was finishing Invisible Cities yesterday, because the book was due at my local library today. I was wondering whether the cities could be adapted into some visual art form. And obviously I thought about feeding it to Opus. 24 hours later, I see this post…
Citizens there gather their tools at the bottom of a giant staircase and climb up countless stairs, only to find out that what they were planning to do has already been done by the time they reach the top.
Dissapointed, they descend to grab a new set of tools and make new plans for tomorrow, only for the same thing to happen again.
Six hours of continuous runtime is the underrated part here. Most one-shot demos fall apart well before that, so this doubles as an endurance benchmark.
Why, I once encountered a running session longer than a day!
The agent had spun up a backward shell script to watch for the shutdown of another process, but wrote a bug in the script that would have left it running indefinitely until I got home and noticed it.
This was with Fable, no less! And it happened a few more times, though I caught them sooner.
I’m not sure runtime is an important metric at all. Shouldn’t we aim for 0 runtime with maximal results?
The supervisor agent will keep the session in a loop until 6 hours have passed and eventually the agent will decide to use up the remaining time rather than fighting with it
I wonder how much co2 is being thrown into the atmosphere everyday through the steady stream of “look what I made this LLM do” and endless “benchmarking”?
One useful hint here is the API token cost - in this case "about $74 in API tokens" for Opus 5.5.
We know Anthropic run inference at a margin, so that $74 means that the cost of the electricity involved is substantially less than $74. I'd love to know the actual cost there.
Approximately nothing - the carbon cost of electricity is already very low, and most of the inference costs are GPU and datacenter amortization. You could require all AI inference to be carbon-neutral and that'd barely raise the API prices.
I wonder how much co2 you use to get through the day, and how much you use to remark on other people's creations in a way that suggests you disapprove of their utilization of their available resources. How much co2 do your projects emit, since you seem to be interested in those metrics?
I strongly suspect one HN post (or two now) is not even in the same ballpark. I could ask Claude to figure it out but that would be a complete waste of time and energy. ;)
Intuitively I'd say that just-for-fun AI projects emit orders of magnitude less CO2 than what just-for-fun road trips or air travel emit.
So unless you can show that it approaches all the other frivolous consumption that humans love to waste resources on, maybe we can get back to experimenting with cool technology and talking about it, on this website called "Hacker News"?
And shame on all the people shitting on it because AI made it. If you valued art for art's sake, you wouldn't be so bothered when somebody makes something art-adjacent using some technique you don't approve of. It in no way detracts from the kind of art you admire. Don't be type who just has to let others know that you find something they like to be beneath you.
It's one of my favorite books - and I think it deserves better; it's a conceptually very dense text and just slapping together a few 3D models misses the point
it's also not very astonishing that the current models are able to create this - it looks exactly as expected
and I actually think getting an interesting AI generated rendition of the text from a single prompt is a worthwhile endeavor
maybe even a good replacement for the pelicans: it would test if a model can engage with a text beyond the surface level
This is probably one amazing positive these models have brought forward - the ability to break out of a single modality, and utilise others to help us learn and visualise. While the visuals here are amazing, my son for example prefers to learn by listening and talking, so we convert lot of his study materials into audio and real-time voice roleplay.
While reading the book years ago, I did not visualize Isidora as a city with 9 houses, like a children’s book, or a little video game level intro, but so unlike Kublai Khan or Italo Calvino.
$74 vs $ 10 vs $25 for the same prompt,
The interesting number is not actually quality, its what we can get per dollar like 6 subagents runs in parallel.
Yeah for now, this week, on that model, all of which are constantly changing. I would love to know actual cost of running those inferences on such hardware, including electricity, training, etc. i.e. what we should be paying. I can imagine that job would cost multiple hundreds if they were charging enough to actually be profitable companies.
both the linked camera lens example as well as the last one the author linked runs at ~5fps on vivaldi and pegs CPU 100%. i have never seen any 3d visualization website have performance this bad, i assume they're not using webgl given my GPU is 0% utilization.
wow that's awesome! super cool :), fyi so inspiring that I did something similar for d&d forgotten realms, including a timeline: https://narfman0.github.io/realms-atlas/
fable drove it, opus agents did the work, took 30-45 minutes. (intend on extending to planescape, elaborating on significant historical events, and some text to speech narrating cool events)
It reminds me of Total War. The soundtrack and way that the map is shown. Is Total War an inspiration? Anyways, I liked the visualization. But I never read the book, so I followed (partly) droidjj recommendation
What is this for? It doesn't deepen the understanding of the book. The illustrations aren't attractive on their own. They're not a very good representation of the descriptions in the book.
I'm a little torn by this sort of demonstration, because while many of the scenes have obvious markers of slop (impossible intersecting geometry, bridges to nowhere, and so on), the scenes largely do work to convey the intended concept/emotion, and the low-poly aesthetic is executed decently well.
I guess I shouldn't be surprised that another visual medium is starting to fall to LLMs after the success of diffusion models for image generation. Nevertheless it's wild that a mostly text-focussed model is building little worlds like these.
A couple of years ago, after reading the book, I sketched some of them on Procreate with an isometric grid, but not all 55. I only arrived at number 4, since each took multiple hour...
https://camillovisini.com/drawing/fc4jc6-le-citta-invisibili...
https://camillovisini.com/drawing/p2yrdj-le-citta-invisibili...
It’s interesting, with a human made piece of art, I’m always drawn to « zoom in », to see how you did little details, if you have hidden something here and there, how you connected different parts.
With AI generated stuff, sure it looks amazing at first glance but I definitely don’t want to zoom in, since I know it’s all a cardboard facade with no passion. Look a bit too closely and you’ll see the bridges that make no sense, the bunch of nonsensical threejs cylinders etc.
I always consider AI-generated content to be a draft — or, in its parlance: something that takes the shape of the final product.
I don't dismiss the ability of models to create. I dismiss their ability to create finished goods. Humans should always be involved, especially with creative works. Models are great at modeling existing creations and solutions — head starts on long roads — but it's the person that needs to walk it and take the journey.
This pairs well with the saying "art is never finished, it is abandoned."
Abandoned one-shot drafts are akin to speaking words but never finishing a sentence before moving onto the next thought. There's no there there.
Since this project (both yours and OP's) seem more for personal fulfillment and enjoyment, I wonder who got more enjoyment out of the process, and also who will remember the results better in ten years.
You draw what you saw in your minds' eye. I can guarantee the author of this didn't visualize the cities like they came out.
Love it. I see isometric angles - I updoot.
> Leaving there and proceeding for three days toward the east, you reach Diomira, a city with sixty silver domes, bronze statues of all the gods, streets paved with lead, a crystal theatre, a golden cock that crowns each morning on a tower...
Text in the second person is causing my my D&D instincts to kick in.
These are amazing!
That looks pretty nice, I didn't know Procreate could be so good at precise line art!
SOUL
I understand why that’s impressive. But I don’t feel much. I guess, I was just used to envy how much people can be absorbed by some idea to spend so much time implementing it. Now it feels like a single shotted nice-typography-and-all presentation at work that someone did 10 min before the meeting.
There's also an eerie feeling of emptiness to the outputs. They lack real intent. Glossy and technically impressive, but they don't tickle your brain. I think it's because they don't really communicate or give you that buzz of new insights. They are pretty but not _helpful_.
For example, I didn't learn about lenses or earthquakes or fusion messing with the controls on these.[1] I didn't know what to look at.
However, there were interactive exhibits at the Children's Museum I spent a long time exploring and loved. That's because they were carefully designed to really teach the concept?
[1] https://sael.net/fusion-pulse/?ref=earthquake-tower
"What we obtain too cheap, we esteem too lightly: it is dearness only that gives every thing its value. Heaven knows how to put a proper price upon its goods; and it would be strange indeed if so celestial an article as freedom should not be highly rated." – Thomas Paine
I've been thinking of this line recently — really gets at the heart of the matter for me. There is no substitute, no shortcut, no alternative for effort.
It's low entropy. The amount of entropy is prompt + model but now that you have seen the model various times, the entropy comes down back to the prompt only.
You'd not be amazed by seeing a 2 sentences prompt and thus you are not amazed by this work either.
Yeah, the demos look really cool but you're thrown into them and there's nothing explained as to what you're looking at or why you should care. Funny, an article explaining something like how a camera focus works with a few illustrations would be far more useful than a fancy 3D interactive.
It turns out the more technology you throw at something does not equate to ease of learning.
If it makes you feel better.. I find this impressive but it doesn't pique my curiosity. I may get biased, finding a reason not to care.. but honestly I think if this were presented differently I would. And it's not so much about using AI either. I think you could use Ai to make something compelling but so far the one shot prompt outputs to me lack focus
Isn’t that at least in part because the intention here was to demo the functionality?
Put in the hands of the educator, they will be able to add those illustrations, design for the light bulb moment, etc, and the tools at their disposal just got much better.
Another thing to keep in mind is that if you asked twelve different artists to render these cities from their descriptions, you would likely end up with twelve very different renditions.
Now imagine the same thing but each person is using Opus 5.5 - I’d wager there would be a lot of commonality among each of the outputs.
Looking at the posts on HN AI topics lately, this seem to be the depressing sentiment.
If something that previously took weeks for a human to design, now takes 10 min, that means a whole week of salary went up in smoke.
When have we automated enough?
> If something that previously took weeks for a human to design, now takes 10 min, that means a whole week of salary went up in smoke.
To me, this is the depressing sentiment. The idea that, somehow, the goal isn't to produce useful things but to spread work out over a period of time in order to maximize payment.
At a high level,I conceptualize my job as "doing useful things." Whether that's design, writing code, debugging, sysadmin, CI/CD, writing doc, whatever. I just want to be useful. If there's some tool that makes that easier, I am happy about it. I don't think that tool will result in my getting paid less. But if it does, oh well.
>But if it does, oh well. I'm glad our ancestors didn't take that line of reasoning we'd still be feudal serfs. "The lord has come and taken all our grain again? Oh well, at least I've been useful to him."
If our ancestors worried more about technology obsoleting human labor we'd still be subsistence farmers. Priorizing usefulness is why we're rich today.
Weird way to look at it, from another view we can now print a week salary in 10 minutes. Might devalue the output slightly but we still have the product.
Who is "we"?
This is so often the essential question, and equally often unanswered
Undoubtedly a fundamental reconsideration of political economy is necessary if we are going to go all in with AI. As of now, this technology will exacerbate the existing 'singularities' and wealth disparity distortions in the socio-economic body. And now, basic fucking income is not the answer. Maybe pitchforks.
Derrek Fox, Rupert Hunt & myself of course, who else could I possibly have meant?
The person controlling the salary
> If something that previously took weeks for a human to design, now takes 10 min, that means a whole week of salary went up in smoke.
The problem also is, if a human took weeks to design this, it would be much better. This is quick but also bad.
This really sounds like a bias selection filter. Humans spend a lot of time on stuff that's shit too.
Every decision a human makes has some level of intent. AI cannot intend, it can only suppose.
If you haven’t read the book yet, I’m tempted to warn you against opening the visualization. Part of the joy of reading this book is letting your mind’s eye run wild, and I fear if I read it for the first time after toying with this site, I’d be remembering the cities rather than conjuring them.
That is the exact problem with the city of Dinipro.
Citizens of this city high in the sky imagine it so fully in their minds that no one there needs to open their eyes to walk the streets and live their normal lives. But if one makes the mistake of opening their eyes and tries to see the city directly, they would immediately fall through the clouds to the ground deep below.
On that note, I'd quite recommend the Coyote vs. Acme movie, some great world building there.
I heartily recommend the audiobook, narrated by John Lee.
The book is a trip and the dreamy sort of narration matched it perfectly. My wife and I would spend evenings listening to one city at a time, on a walk around our little park, one headphone in each ear. I don't know if it was just the environment or the time in our lives, but we both really look back fondly on this book.
I read the book decades ago, then I lost it in one of the many relocations. I always thought I should buy it again but never acted on the idea... A few years ago, I was walking by an antiquarian, and noticed a thin unmarked gray spine. I entered, picked it up, paid, and walked away.
PS I didn't open the gallery. Nice idea but I don't care about visualizing something I should feel.
I fully agree, thank you for this warning.
How Opus 5.5 imagines these cities is a question for Opus. The real experience is how we imagine - and I believe there are no two people who have exactly the same mental image of an invisible city.
Checking Amazon, several of the available editions are illustrated, which I suppose would have the same problem.
I opened it because I had no idea what Invisible Cities meant here. I thought they might be some real cities that don't come up on maps.
The website gives you a good hint before opening up the visuals. So, I got an idea.
I don't really know about Marco Polo. My only knowledge comes from the Netflix series, which doesn't talk about these invisible cities.
Invisible Cities is a book.
> I’d be remembering the cities rather than conjuring them
Having looked at some of them (and it being one of my favorite books) ... it won't do that. These visualizations are as terrible as they are unnecessary.
Don't worry about opening the GPT-6 Astra one though, those ones are so bad that your brain will immediately expunge them.
Maybe the interactive link in the article for the optic focus can lead you on a fun journey instead. And the bottom-right of every visualization is a link to yet another fun diorama to explore.
It's a techie creating the torment nexus but the imagination version. No need to picture in your mind the invisible cities! Time to make them visible!!!
Talk about missing the entire point of the piece of literature.
I copied the prompt from the article and inserted it into Claude, here is what it created:
https://claude.ai/share/6bbd421a-147d-435d-bc00-7e93b393b943
(the standalone version at the very end of the chat is the one that runs in the browser)
Well, overall, by faaaaar not that beautiful as what the author from the article received? :-D
(though, it was done with Opus 4.8)
Invisible Cities is my favorite piece of literature hands down. The audiobook narrated by Richard Higgins is also done so beautifully and can't recommend it more.
As neat as this project is, this does no justice to whatsoever to the book. On the surface Invisible Cities seems to be a book about random cities' descriptions. In reality though, this book is about semiotics, meaning, language, and it's limits.
The first city I clicked on at random had the opening phrase:
> Arriving, you rejoice at its bridges, each different
The model city had like five bridges total, three of which were didn't bridge anything at all but instead sat in the middle of the river and went along its flow, the "bridges" for some reason pretending to be small islands.
You made the mistake of looking too closely
like bridge over slop and water, i will let you down.
Astra Penthesilea has towers that sometimes clip into the green... hedges? Pipes?
Opus Eutropia has a few isekai circle cities with one of them... ON the river, giving no space for ship traffic.
Slop has never been this beautiful before!
Not to mention that the slop site claims that the cities are Invisible, when I can clearly see them.
Such a drag- a few carelessly lumped together Ai demos. I mean the headline says it all, there is not much more to it. ok, no thanks!
Like many other people in this thread, I feel like passing what would essentially be a hobby project off to Claude robs the creator of most the fulfillment and gives the viewer less reason to appreciate it. But with the models being as good as they are today, I genuinely feel conflicted. If your end goal is to share something with others, why spend a weekend on a project like this when you could just hand it off to Claude?
Your appreciation should come from the one shot prompt. It's a direct comparison to Astra's one shot. I feel like you and many others are trying to "appreciate" the wrong thing.
If the creator truly wanted viewers appreciate the end product they'd start with the one shot output and spend their weekend iterating on each city either by hand or with Claude. But they didn't want to do that, they wanted to show off what Claude/GPT could do in 6hrs. That's the whole blog post.
This isn't a ceiling, though. This is the new floor: What you can get with more or less zero human effort and ~$100.
I expect we're going to see more and more amazing "hobby projects" that did involve both a few hundred dollars in tokens and hours of human creativity and effort in combination.
Why would that someone want to listen something you haven't created? What have you "shared"? Claude could have made the tools that let the creator build their cities from their own imagination, but that's not what this is.
Interactive demo: https://p.migdal.pl/invisible-cities-opus-5.5/
Source code: https://github.com/stared/invisible-cities-opus-5.5
Yeah. This fails to load both in Opera and in Chrome on Android. I'd tone the hyperbole down a bit.
I can't imagine a worse thing happening to a better book.
One of my favorite things about the National Treasure Cinematic Universe (two movies and a TV show) is the existence of characters that I like to call "treasure grumps". Their role in the story is to warn the main characters that searching for treasure is a terrible way to spend your time, and can only lead to personal and familial ruin.
I was finishing Invisible Cities yesterday, because the book was due at my local library today. I was wondering whether the cities could be adapted into some visual art form. And obviously I thought about feeding it to Opus. 24 hours later, I see this post…
That is what happens in the city of Ternopoli.
Citizens there gather their tools at the bottom of a giant staircase and climb up countless stairs, only to find out that what they were planning to do has already been done by the time they reach the top.
Dissapointed, they descend to grab a new set of tools and make new plans for tomorrow, only for the same thing to happen again.
Funny note about Opus 5.5.
I once instructed it to spawn at most 20 subagents and it spawned 40, with an adversarial reviewer for each of the 20.
It doesn't follow these instructions very well.
20 sub agents, 20 doms.
Six hours of continuous runtime is the underrated part here. Most one-shot demos fall apart well before that, so this doubles as an endurance benchmark.
Why, I once encountered a running session longer than a day!
The agent had spun up a backward shell script to watch for the shutdown of another process, but wrote a bug in the script that would have left it running indefinitely until I got home and noticed it.
This was with Fable, no less! And it happened a few more times, though I caught them sooner.
I’m not sure runtime is an important metric at all. Shouldn’t we aim for 0 runtime with maximal results?
How do you get it to spend six hours? I’ve done projects where it would have benefited if it put in extra work.
/goal spend at least 6 hours doing … works
The supervisor agent will keep the session in a loop until 6 hours have passed and eventually the agent will decide to use up the remaining time rather than fighting with it
I wonder how much co2 is being thrown into the atmosphere everyday through the steady stream of “look what I made this LLM do” and endless “benchmarking”?
Sigh
One useful hint here is the API token cost - in this case "about $74 in API tokens" for Opus 5.5.
We know Anthropic run inference at a margin, so that $74 means that the cost of the electricity involved is substantially less than $74. I'd love to know the actual cost there.
I wonder how you only reach $74 in 6 hour runtime. I would have guessed this runtime goes into the hundreds of dollars.
Or I'm realizing that I'm overpaying my tokens...
A lot of that 6 hours isn't spent waiting for the models to return tokens, it's spent executing the code that they've written.
Approximately nothing - the carbon cost of electricity is already very low, and most of the inference costs are GPU and datacenter amortization. You could require all AI inference to be carbon-neutral and that'd barely raise the API prices.
I wonder how much co2 you use to get through the day, and how much you use to remark on other people's creations in a way that suggests you disapprove of their utilization of their available resources. How much co2 do your projects emit, since you seem to be interested in those metrics?
I strongly suspect one HN post (or two now) is not even in the same ballpark. I could ask Claude to figure it out but that would be a complete waste of time and energy. ;)
This has given me a great idea: I have a modest proposal on how to reduce net CO2 emissions while not reducing token consumption at all
USE carbon dioxide to get through the day? What?! We're not plants!
I assure you that if your body did not emit co2 throughout the day you would not continue metabolic processes for an especially long period.
> I wonder how much co2 you use to get through the day
What a terrible thing to say to human! Being (presumably) human, it is their inalienable, natural-born right to do so.
Shame on you.
Intuitively I'd say that just-for-fun AI projects emit orders of magnitude less CO2 than what just-for-fun road trips or air travel emit.
So unless you can show that it approaches all the other frivolous consumption that humans love to waste resources on, maybe we can get back to experimenting with cool technology and talking about it, on this website called "Hacker News"?
Most of that time may have been spent on testing.
The amount of responses to this obvious bot comment is somewhat sad.
claude slop, like from every comments you posted
Very cool, thanks for posting!
And shame on all the people shitting on it because AI made it. If you valued art for art's sake, you wouldn't be so bothered when somebody makes something art-adjacent using some technique you don't approve of. It in no way detracts from the kind of art you admire. Don't be type who just has to let others know that you find something they like to be beneath you.
It's one of my favorite books - and I think it deserves better; it's a conceptually very dense text and just slapping together a few 3D models misses the point
it's also not very astonishing that the current models are able to create this - it looks exactly as expected
and I actually think getting an interesting AI generated rendition of the text from a single prompt is a worthwhile endeavor
maybe even a good replacement for the pelicans: it would test if a model can engage with a text beyond the surface level
but this ain't it
Does this bring any happiness? for me, No. effort need a purpose.
This is probably one amazing positive these models have brought forward - the ability to break out of a single modality, and utilise others to help us learn and visualise. While the visuals here are amazing, my son for example prefers to learn by listening and talking, so we convert lot of his study materials into audio and real-time voice roleplay.
Latent space is amazing isn’t it
While reading the book years ago, I did not visualize Isidora as a city with 9 houses, like a children’s book, or a little video game level intro, but so unlike Kublai Khan or Italo Calvino.
$74 vs $ 10 vs $25 for the same prompt, The interesting number is not actually quality, its what we can get per dollar like 6 subagents runs in parallel.
Yeah for now, this week, on that model, all of which are constantly changing. I would love to know actual cost of running those inferences on such hardware, including electricity, training, etc. i.e. what we should be paying. I can imagine that job would cost multiple hundreds if they were charging enough to actually be profitable companies.
On the down side, $84 spent on a slot machine might have actually yielded some benefits.
both the linked camera lens example as well as the last one the author linked runs at ~5fps on vivaldi and pegs CPU 100%. i have never seen any 3d visualization website have performance this bad, i assume they're not using webgl given my GPU is 0% utilization.
wow that's awesome! super cool :), fyi so inspiring that I did something similar for d&d forgotten realms, including a timeline: https://narfman0.github.io/realms-atlas/ fable drove it, opus agents did the work, took 30-45 minutes. (intend on extending to planescape, elaborating on significant historical events, and some text to speech narrating cool events)
rock on
It reminds me of Total War. The soundtrack and way that the map is shown. Is Total War an inspiration? Anyways, I liked the visualization. But I never read the book, so I followed (partly) droidjj recommendation
I just have zero interest in looking at any of these "AI did a thing for me." posts.
Cecilia looks like an average residential area in Russia.
The creation really is amazing. I harnet heard of the book but got a long way through it!
What is this for? It doesn't deepen the understanding of the book. The illustrations aren't attractive on their own. They're not a very good representation of the descriptions in the book.
Doesn't work right on my phone.
Vanadium (chrome)
https://imgur.com/a/nGLyJf6
this is amazing, wonder how much better it's going to get in a year from now
I'm a little torn by this sort of demonstration, because while many of the scenes have obvious markers of slop (impossible intersecting geometry, bridges to nowhere, and so on), the scenes largely do work to convey the intended concept/emotion, and the low-poly aesthetic is executed decently well.
I guess I shouldn't be surprised that another visual medium is starting to fall to LLMs after the success of diffusion models for image generation. Nevertheless it's wild that a mostly text-focussed model is building little worlds like these.
Loads nothing on iphone lockdown mode fwiw.
Isn’t it expected?
I ran cat /dev/urandom for six hours but I didn't need to write a blog post about it
Just curious - was the result a 3D animation? If so, I'd love to read your blog post!
If you're going to spend tokens on 3D stuff, then please build me a better FreeCAD first.
https://github.com/dzervas/cadara
Thanks, I hate it. I imagine Calvino turning in his grave at such an abdication of human imagination.
"Attention whore" used to mean someone who makes garbage to get attention. Now it's a machine that uses attention to make garbage.
Well it still has the trademark "colored left-side of the boxes" ai slop, but otherwise looks (and sounds!) very impressive.
Well it still has the trademark "colored left-side of the boxes" ai slop, but otherwise looks very impressive
What a beautiful website, wow. I've been online 10,000s of hours, have seen pretty much everything and this one goes straight in my Top 10.
I'm so glad I'm able to witness this Cambrian explosion of software :).