Загрузка видео...
Не удалось загрузить видео
I had access to the new Claude Projects and was able to do some very complex work. Here, I asked it to go through all the images, videos and records about Umberto Eco's famous 33,000 book library & try to reconstruct it, including book locations, in 3D.
57,271 просмотров • 20 дней назад •via X (Twitter)
Комментарии: 16
Are you doing this on the $100 or $200 plan? Did you hit any limits?
Nice
Eco called this an "antilibrary" books you haven't read matter more than ones you have. In 1974, George Miller's chunking work showed experts don't store facts, they store retrieval paths through a space like this one. Your reconstruction is literally rebuilding his mental index, not the shelves
How is it different? I’m sure that I’ll get it when I play with it, but it’s not making sense as described?
I have not been able to crack image generation in Claude. I know it’s parameters are more strict than ChatGPT, for example, but everything I’ve promoted looks like garbage.
Reconstructing a 33,000-book library in 3D from scattered photos and video isn't really a coding task, it's a synthesis task across formats an agent doesn't normally get asked to unify. That's a better stress test than most benchmark demos use.
Reconciling conflicting fragments is what most real work actually looks like and this is such a good test case for 'complex work'
33k books is a catalog problem. shelf placement is the hard one. did one Projects thread do camera-path layout and the others fill titles, or did each thread guess shelves on its own?
Holy shit
really cool, as Italian I've always loved the idea of Umberto Eco's library. Can you please share what type of model is it?
ok that's amazing
Eco's whole point was that the unread books were the valuable ones. We just built a machine that can find any of them in three seconds. Sorry about that.
A perfect 3D library is still a demo until the P&L moves. Claude rebuilt Eco's 33,000 books — MIT NANDA still finds ~95% of enterprise AI pilots with zero bottom-line impact.
with a library reconstruction, you want to know why a book is on that shelf. keeping sources and certainty visible gives you a way to separate what the records show from what the model inferred.
The 'Guesses only' button deserves to escape this demo and become a standard feature.
We need to publicly shame these "I had Early Access" posts
![Dave Portnoy says he paid $10,000 for an AI clone to narrate his audiobook for him — and it was “actually scary” how dead-on it sounded. His parents couldn’t even tell the difference. Then Audible shut it down, and he had to record every word himself. “I could only do 15 minutes [of recording a day] because it was that hard for me to read it. I hated it. I tried to get out of it. I paid $10,000 to do AI. I had this AI program for it.” “My parents couldn’t tell the difference. And it could literally just scan the book and go word-for-word.” “It was actually scary how it was dead-on. And it could have literally just done every single word in that book.” “[Audible] wouldn’t let me use it... I was told, ‘No, they don’t let you do that.’ So I had to [record the whole book myself]. I hated it. Hated it!”](https://image.24vids.com/tw-2077262361967112580/amplify_video_thumb/2077260350420840448/img/2dYUr5eFtutFpAwL.jpg)
