Enough talk—let’s get straight to today’s daily sit (doge)!
Just now, the expert Kappasi announced that "we" at Anthropic have begun enhancing large models in an entirely new way—
The Lord of the Rings.
According to Kapsis, this "Lord of the Rings benchmark" is replacing the once-popular "Pelican Biking" SVG test.
Specifically, Kappasi directly fed the opening of The Lord of the Rings to Opus 5 and had the model generate an entire "Middle-earth" in real time using Three.js.
As for the final result, it’s somewhat reminiscent of Chinese 3D animated series from the late 1990s and early 2000s: rough and highly abstract.
But objectively speaking, if you look closely for a while and happen to be familiar with Hobbiton, the filming location of The Lord of the Rings in New Zealand, you might indeed notice a certain resemblance~
Yet, this seemingly unpolished thing actually cost Opus 5: 1 million tokens, 2 hours, and 5,500 lines of code.
Of course, Claude wasn't the only one shining on stage.
The funniest thing, though, is the new dish served up by netizens. DeepSeek V4 Flash version.
Directly show the entire scene of "Chinese people can fly," with all characters floating.
Faced with these currently obvious inconsistencies, Capasi has also been quite reasonable.
He noted that Opus 5 currently cannot truly "enter" the world it generates; instead, it can only take screenshots at different time points and gradually check where issues occur.
And this precisely reveals a significant shortcoming of current large models:
They can already write code, create scenarios, and generate games, but they still cannot truly understand videos or play the games they create themselves.
In two hours, build a Middle-earth from scratch.
Let’s first look at how this “Lord of the Rings” test is conducted.
If taken literally from Kappaci’s tweet, the primary prompt entered into Opus 5 should be the opening of Chapter One, “A Long-Expected Party,” from The Lord of the Rings.
Bilbo Baggins of Bag End announced he would be hosting a grand birthday party for his 111th birthday, and immediately, the Shire buzzed with chatter...
Friends who are interested can try it out themselves.
In addition, Kappaci specified Three.js in the prompt, which is a JavaScript library for creating 3D scenes using code.
Thus, Opus 5’s task is to first understand the opening text of The Lord of the Rings and then translate it into a 3D world that can run in real time within a browser.
Throughout the process, Opus 5 must assemble characters, buildings, and props using polygons, place them individually into the x, y, z coordinate system, and then set up the cameras, lighting, and animations.
When characters move, where the camera turns, how the lighting changes, and how objects in the scene interact—all must be defined by the model through code.
Although the final visuals weren't polished and often suffered from issues like clipping and floating, such as characters' bodies and heads separating, the entire scene was at least fully constructed.
It should be noted that this is not the same as a video model generating frames pixel by pixel.
Three.js requires first creating characters, objects, and scenes as 3D objects, then dynamically calculating their positions, angles, and motion states based on code.
Therefore, what we see in the demo is not a regular AI-generated video, but a screen recording of a 3D web scene running in real time.
However, Kappasi also noted in the comments that programmatic 3D and video generation are not mutually exclusive.
Some netizens suggested submitting this rough Three.js screen recording to Seedance as a reference video, then having the video model re-render it at higher quality.
Kapasi quickly agreed.
According to his vision, programmatic code could handle storyboarding and control, first establishing the framework—such as where characters stand, how the camera moves, and how the plot progresses.
Next, feed the screen recording into the Video-to-Video model to add textures, lighting, and details, fully enhancing the visual appeal of the entire Middle-earth world.
As for the audio, Opus 5 did not include it this time.
Kapasi stated that, due to personal requirements for audio quality, ElevenLabs was ultimately used.
Thus, the opening segment of the Lord of the Rings demo—with its visuals, camera shots, and narration—suddenly came into being.
Kapasi has also open-sourced the entire project; the specific link can be found at the end of this article.
Netizens are much more interesting than Picapasi.
Shortly after Kappasi released the demo, many netizens came to test it out.
Overall, all we can say is:
This generation of internet users has far more imagination than Kappa.
Someone has launched an "Earth Online" project using Claude, aiming to model the entire Earth with a network of AI agents. Click here Built.
Currently, the Agent has not yet created the entire Earth, only constructing a low-poly-style waterfront area of San Francisco. However, based on the existing visuals, the building proportions, human scale, and overall style remain remarkably consistent.
This effect is somewhat like an animated movie set in a LEGO world—the art style isn't complex, but the characters, environments, and movements all operate smoothly within the same universe.
Continue in this direction with simple-styled animations and lightweight games; a prototype of an AI-native approach is already taking shape.
Some netizens have built a 3D digital model of New York City using Fable 5 and GPT-5.6 Sol, integrating real-time data into it.
The model must not only correctly place New York’s streets and buildings but also maintain the spatial relationships between different areas. The author therefore suggests that this task of generating virtual worlds could serve as a new benchmark for spatial reasoning.
Even more impressive is yet to come.
Some netizens who couldn't get tickets to Kanye West's concert simply used AI to stage their own concert in the browser.
The entire project is still built using Three.js. There is only one HTML file, with no external 3D models used—all scenery, characters, and lighting are generated entirely through code.
The entire concert features 14 songs, each with its own lighting design and a unique spherical stage visual.
Regular users can listen to clips; after connecting to Spotify Premium, they can play full tracks and entire performances.
Didn’t get a ticket? Just create your own private screening—that’s such a waste!
Not over yet!!!
Kapasi also envisioned at the end of the original post that future updates could introduce gameplay elements to these 3D worlds, allowing players to enter as observers, NPCs, or even story characters.
Before Kapaxi could arrange it, netizens had already created the game version.
This project also uses Opus 5 and Three.js, enabling not only execution and interaction, but also including audio.
More 3D design examples are continually emerging. From architectural scale and spatial layout to scene style, Opus 5 maintains relatively consistent results even in fairly large-scale projects.
There are many similar examples, which we won't list individually here.
Objectively speaking, these works are still far from being truly mature games.
It is still unclear whether the complex gameplay is enjoyable, whether it runs stably over long periods, or whether players will truly be willing to stay for 15 minutes.
However, the barrier to creating real-time 3D content and playable prototypes has been significantly lowered.
Moreover, wouldn't it be wonderful to have a new way to test model capabilities that is also fun and engaging?
Pelican, ride's over
So why suddenly have a large model generate The Lord of the Rings?
This goes back to the popular "Pelican on a Bicycle" test from a few years ago.
This question originally came from developer Simon Willison and must be answered in one sentence:
Generate an SVG image of a pelican riding a bicycle.
Although this question may seem simple, it actually tests the model quite rigorously.
Because SVG appears to be an image but is actually a string of code underneath.
The model must not only know what a pelican and a bicycle look like, but also break them down into lines, circles, and polygons, then precisely position each component using coordinates.
More importantly, pelicans and bicycles simply don't go well together.
The frame, tires, and pedals of a bicycle must maintain proper geometry; pelicans, on the other hand, have large beaks, short legs, and a physique that hardly seems suited for cycling.
When combined, it’s almost immediately clear whether the model understands spatial relationships:
Whether the wheel is crooked, whether the feet are on the pedals, or whether the bird is riding the bike—or being dismembered by the frame.
Take a look at the demos from the end of 2024 above~
For this reason, "the pelican rides a bicycle" once became a classic test for assessing large models' spatial understanding, object combination, and code generation capabilities.
But as the model grows stronger, this pelican is quickly reaching the end of its ride.
Look below DeepSeek R1 and DeepSeek From V4's performance, it's clear that today's models can solve this problem quite well.
More importantly, a single SVG can only evaluate the model's one-time output.
It cannot measure whether a model can plan a complex project, work continuously for hours, and repeatedly review and correct its own errors within thousands of lines of code.
So, Kappasi turned a simple “draw a picture” task into a massive undertaking—building a world—
The pelican can get off; Middle-earth has officially taken over.
Kapasi's Pelican
Soon, the news that Kappasi was preparing to change the questions for the "Pelican Test" spread across various communities.
The top-rated comment on Hacker News suggests that although the demo visuals are of low quality, this precisely demonstrates that we need a new test harder than generating a single image.
The new benchmark shouldn't only assess whether a model can draw accurate images, but also whether it can understand a world.
After all, "pelican riding a bicycle" has been used for too long.
Models continuously achieve new benchmarks on these tasks, making it increasingly difficult to distinguish differences between them.
In contrast, generating a complete 3D world requires the model to understand spatial relationships between characters and objects, and to handle camera angles, movements, and scene transitions—rather than simply invoking a video generation model to output a sequence of frames.
Of course, some people have raised objections.
The pelican test is sufficiently simple, low-cost, and yields results that are easy to compare.
Using so many tokens to generate an entire 3D world just to test the model feels a bit like using computing power for fireworks.
Meanwhile, whether Three.js itself can evaluate the comprehensive capabilities of large models has also been questioned.
Some believe that such demos can only demonstrate that Anthropic has trained well on Three.js code, but do not prove that the model truly understands space and the physical world.
But soon, netizens countered:
Convert a piece of abstract, ambiguous literary text into a 3D animation, requiring the model to simultaneously handle spatial relationships, physical laws, everyday objects, and mathematical problems in 3D transformations and computer graphics.
If this still only counts as “knowing how to write Three.js,” it’s a bit of an understatement for these 5,500 lines of code.
Another netizen raised a more open-ended question:
Is "spatial reasoning" truly different from the reasoning large models typically use when processing text and code?
One perspective holds that whether the image is valid depends on whether the model understands spatial relationships such as front/back, inside/outside, near/far, and occlusion, as well as the distance, angle, and relative size of objects from different viewpoints.
However, another perspective holds that whether the model is processing “next to a stone” or “inside an array,” it may be doing the same thing: generating tokens one by one based on context, thereby completing reasoning in the process.
If that's the case, then what Opus 5 demonstrates is not just a sudden emergence of "spatial capability."
More likely, the general reasoning capabilities of large language models, originally designed to understand text and code, are naturally extending into the three-dimensional world.
From drawing a pelican to building a Middle-earth, the topic may seem to have changed, but what’s being tested behind the scenes might still be the same question:
Can a model transform its understanding of the world into a truly functional structure?
And finally, there may be an even more outrageous question—
If general large models can already write code to build 3D worlds and call APIs like Seedance and ElevenLabs to generate visuals and audio, how necessary is it for users to manually open a dedicated video generation product and input prompts?
Reference link
[1]https://karpathy.ai/lotr-movie/
[2]https://simonwillison.net/2025/Jun/6/six-months-in-llms/
[3] https://x.com/wizardbrainz/status/2083012159341203708
[4]https://x.com/aniketjart/status/2083645765097033845
[5]https://x.com/davidfromkansas/status/2075691129899528254
[6]https://x.com/MindaugasLT/status/2083488027343470939
This article is from the WeChat public account "Quantum Bit," authored by Henry.