[Introduction] Fable 5 is back online. In a video, Gostev from Arena.ai unleashed 63 3D worlds, nearly all generated in one go. After watching the video, even Karpathy, who just joined Anthropic's pre-training team, exclaimed he never expected this.
A bear stands by a river, snapping up a leaping salmon.
The fish struggles desperately in its mouth.
This scene completely stunned Andrej Karpathy.
After viewing this set of 3D videos, he wrote in a post:
It's absolutely incredible. I hadn't realized before that models could create such incredible, rich, and playable worlds.
He also coined a new term: fablemaxxing.
This is so beautiful, it's peak fablemaxxing! Every new model scale, there's always something qualitatively new that leaps forward and surprises.
Karpathy just joined Anthropic in May, specifically their pre-training team.
In June, Fable 5 debuted, was shut down by US export controls three days later, and was just reinstated on July 1st. The video's creator is Peter Gostev from the AI evaluation platform Arena.ai.
Using Fable 5, he ran 63 high-difficulty 3D generations in one go, almost all in Three.js, with the majority succeeding on the first try.
Just over a month ago, getting AI to create a playable 3D world you could walk into in one attempt was nearly a pipe dream. Now, Fable 5 can churn out 63 at once.
Thus, Gostev couldn't help but marvel in the video: even the people in the labs haven't fully grasped what it can do yet.
1600 Lines of Code, a Living Underwater Manhattan
Let's start with the most impressive one in the video.
The entire Manhattan is submerged underwater. Central Park, skyscrapers, street textures — the details are unbelievably rich.
Gostev looked at the code: 1600 lines. These 1600 lines support a living underwater city.
Fable 5 generated Manhattan, showing the entire island from Battery to Inwood, with thousands of buildings each with distinct outlines, all powered by 1600 lines of code. (Source: Peter Gostev video)
In total, Gostev created 63 such worlds, categorized into 6 major themes.
Among the giant 3D worlds, Istanbul spanning the European and Asian shores, London over 2000 years, magnificent pyramids, the eruption of Pompeii, and the Golden Gate Bridge with flowing traffic were all constructed.
Istanbul generated by Fable 5 in one go, spanning Europe and Asia, layered into the sea. (Source: Peter Gostev video)
The Edible Kingdom generated by Fable 5, a chocolate factory made of candy. (Source: Peter Gostev video)
For playable scenarios, there's parkour on New York rooftops, a physics playground where you can tear down an entire city, and a flight simulator with a fully equipped dashboard.
The most unbelievable category is stepping into world-famous paintings. Van Gogh's Starry Night, Monet's Water Lilies—each spread out into a traversable 3D world.
Another category is impossible perspectives: a one-millimeter-tall ant observing a garden in the rain, opening its eyes to a mountain of gold.
Fable 5 compresses the perspective to one millimeter high, with blades of grass towering like majestic peaks. (Source: Peter Gostev video)
Among natural wonders, Niagara Falls, a forest where fireflies synchronize their flashes, and the bear catching a salmon that stunned Karpathy are in this category.
Finally, there are elements and the universe: the Red Sea parting, a volcanic island being born, a space elevator piercing the clouds.
Fable 5 generates the Red Sea parting, with people walking between the two walls of water. (Source: Peter Gostev video)
In Gostev's view, whether a single scene looks good has never been the hard part. The difficulty lies in whether the model can manage vast numbers of elements simultaneously and make them coordinate with each other.
Weaker models often do the first 80% decently, then completely collapse in the final 20%, leaving even longer debugging time.
You Can't Simply Copy a Painting; It Has to Understand First
If building cities still follows some logic, then the next batch of paintings is what made Gostev say "unbelievable" several times.
Van Gogh's The Starry Night.
You can't simply copy such a painting, with thick paint smeared all over the canvas, brushstrokes swirling into vortices.
Fable 5's approach is to use individual lines to reassemble that starry sky in three-dimensional space, then let you fly into it, navigating between the swirls.
Fable 5 deconstructs Van Gogh's The Starry Night into individual 3D lines, reassembling them in space, allowing you to fly amidst the swirling stars and vortices. (Source: Peter Gostev video)
Monet's Water Lilies and Hokusai's Great Wave were similarly "opened up" into worlds you can roam.
In Gostev's opinion, Fable this time is particularly good at explaining things clearly, able to demonstrate while clarifying how things work. This explanatory power is something he didn't anticipate before.
Of course, these worlds weren't conjured with just "one sentence".
Gostev fed the model long, detailed specification documents. Most cases succeeded on the first try, with some requiring one or two rounds of fine-tuning.
The real watershed moment is this: things that used to take dozens of rounds to polish out can now mostly be generated in one go, with a single HTML file.
Behind the Magic, There Are Also Shortcomings
Gostev admits his showcased cases were curated. He generated about 20% more than he displayed, filtering out those with obvious bugs, leaving these 63.
Games are its weak spot.
He admits those few playable scenarios "get boring after 30 seconds." There was a Roman Empire version he thought looked too cartoonish.
More subtly, he feels the model sometimes seems lazy, holding back its capabilities, and you have to repeatedly push it to be more ambitious before it shows its true prowess.
When Fable 5 launched, it topped the Agent Arena leaderboard on Arena.ai's own platform with the largest lead in history.
Arena.ai states that this leaderboard measures a model's ability to genuinely accomplish tasks across millions of real-world assignments.
This Exploration Has Only Just Begun
Returning to the scene in the video of the bear catching the salmon.
What struck Karpathy as strange is that the fish struggles when caught—a detail no one requested. How does a large model, trained only from the internet, know this?
And knowing it, how does it translate that understanding into xyz coordinates, meshes, transformations, animations, effects, interactions, even micro-narratives?
This question moves beyond "whether the AI draws well" to "how much does it actually understand."
Karpathy also mused: adding one to three more model scales on top of this, what could be created is truly unimaginable.
Gostev is equally amazed by the rapid AI progress, "Don't be constrained by models from six months ago. You must have something that past models couldn't do, but today's model can. Go try it."
Even the creators of the models haven't yet found its limits.
This exploration has only just begun.
References:
https://github.com/petergpt/3d-prompt-collection#prompt-01
https://x.com/karpathy/status/2073505440479293773
https://www.youtube.com/watch?v=rTc2_-1KuRE&t=14s
This article is from the WeChat public account "Xinzhiyuan" (新智元), author: ASI启示录; edited by: 元宇
