Globally exclusive first test of Genie 3! The details of the laboratory are exposed, which are extremely shocking. The last piece of the AGI puzzle has been completed.
Last night, the "Third World War" officially broke out.
On the eve of the release of GPT-5, three major model manufacturers joined the fray. August 5, 2025, should be a day that will be inscribed in the annals of AI history.
Amid the intense competition, Google DeepMind's world model Genie 3 is like a powerful bomb, representing a new frontier in world models.
It can be said that the leap from static videos to interactive worlds marks a turning point in the development of world models and AGI.
You know, Genie 2 looked like this just a year ago, but in just one year, Genie 3 has evolved into what you see on the right...
You know, Genie 2 was not real-time and required a few seconds of waiting, but Genie 3 is completely real-time.
Moreover, Genie could support about 10 seconds of generation, Genie 2 could support 20 seconds, and with Genie 3, it can simulate an interactive environment for several minutes.
It can be said that Genie 3 has changed everything.
A YouTuber visited Google DeepMind's London headquarters in advance and conducted the world's exclusive first test of Genie 3. The 30 - minute video released revealed more astonishing details.
Internal test by a former Google employee: It will forever revolutionize the gaming industry!
Without the need to pre - build 3D models, Genie 3 can generate several minutes of consistent videos at 720p resolution just through text descriptions.
The "promptable world events" feature is even more astonishing. Just through text commands, new objects can be added and characters can be generated, opening up new possibilities for training AI agents.
Just now, former Google DeepMind employee Tejas Kulkarni also shared his initial experience of using Genie 3.
Here is his exclusive test demo.
His evaluation can be summed up in four words - "Unbelievable!"
In summary, this is the first neural game engine, or world model, he has tried that performs so well and has long - term world consistency.
He believes that the birth of Genie 3 will completely revolutionize the gaming industry. It can be said that it is the last piece of the puzzle before we achieve full AGI.
In many ways, it is more like ASI than AGI. Since its fidelity and generalization ability have reached human levels and will quickly surpass humans, it can be combined with 3D artificial intelligence and LLMs to completely revolutionize AAA games.
According to this former employee, the highlights of Genie 3 can be summarized as follows.
Truly general, with a fast startup time, and can be extended to other industrial and real - world scenarios.
It can learn physical knowledge. It can learn game engines and non - rigid body physics without an underlying engine. It is very effective for stylized environments where characters move around.
It is much more interesting than video models.
Realistic roaming, and the drone shooting effect is excellent.
The global lighting and lighting effects are great.
The visual memory is very powerful.
Of course, there are still some unresolved issues.
Physics is difficult. (It failed when trying the classic intuitive physics experiment with a tower of blocks.)
Social and multi - agent interactions are difficult. 1v1 combat games don't work.
Long - term instruction following and simple combinatorial game logic fail (e.g., collecting some points/keys, walking to the door, unlocking, etc.).
The action space is limited.
It is far from being a real game engine, but it gives us a glimpse of the future.
Moreover, Kulkarni also specifically mentioned a major highlight of Genie 3 that was officially mentioned - the memory function.
Even after 20 - 30 seconds, something you see will remain the same.
Unveiling the birth of Genie 3: The world's exclusive first test, and the lab details are extremely shocking
As soon as Genie 3 was released, YouTuber "Machine Learning Street Talk" immediately released an interview video with the behind - the - scenes team.
They visited the laboratory on - site and revealed the birth process of Genie 3.
During the process, the host kept exclaiming: "This is the most amazing technology I've ever seen!"
After trying out Genie 3 at Google DeepMind's London headquarters, he said: "This technology will become the next trillion - dollar industry and may even become a killer use case for VR."
The guests of this episode are the masterminds behind Genie 3 - two researchers from Google DeepMind, Shlomi Fuchter and Jack Parker Holder.
Interestingly, unlike previous interviews, this time they were very secretive about the key technical details of the Genie 3 architecture.
The host commented: "It's understandable. After all, Zuckerberg is like a truffle - hunting dog, searching everywhere. But he advised Zuckerberg not to do so because these researchers are doing 'god - like work'. If Zuckerberg really wants it, he should make one himself. (Just kidding)"
The world's exclusive first test
It can be said that one of the most impressive features of Genie 3 is its consistency.
The world it creates has reliable memory. If we look away from an object and then look back, it will still be there.
Surprisingly, the two researchers explained that this consistency is not explicitly programmed; it is a surprising "spontaneous" ability that emerges in powerful AI models.
Moreover, it represents a huge leap. Genie 2 was already a significant leap before, but its speed was not sufficient for real - time interaction, and the resolution was much lower.
This time, Genie 3 has a resolution of up to 720p, is interactive, and has photo - realistic fidelity, and can run smoothly for several minutes each time.
Moreover, Genie 3 represents a killer application for training robots.
The team believes that Genie 3 will completely change the landscape of AI training. Instead of training self - driving cars or robots in the real world (which is both slow and dangerous), we can create infinite simulated environments.
You can even trigger some rare events, such as a deer running across the road, to teach AI how to safely handle unexpected situations.
Genie 3 is different from traditional game engines or simulators and is not like a video - generating model, but it does have the characteristics of all three.
In essence, it is an interactive world model and video generator.
This is a major step forward in technology. You know, in the 1996 Quake engine, it still required explicit programming of physics, rules, and interactions.
However, the new - generation AI represented by Genie 3 can directly learn the dynamics of the real world from video data.
Moreover, it allows us to control agents in the world in real - time.
This transformation completely gets rid of the limitations of hand - coded simulators. You know, the most advanced platform before, XLAND, was just like a cartoon and far from the real world.
But now, with just a simple prompt, you can generate any interactive world you want to train agents in.
The first version of Genie was trained on 30,000 hours of 2D platform game recordings.
Its core innovations