Google has just released Genie 4: the real world model!

  It's the first time in the history of the world.

Google's version of "The Hacking Empire" is coming true tonight!

Recently, Google DeepMind has released the Dark Knight Rises, officially launching a new generation of universal world models.Genie 4。

This is Google's most realistic, interactive model of the AI world to date.

A suggestion can create a direct message.It's a real-time, life-changing "living universe".！

For the first time, the CEO of Google announced that it was time to move from "generating a video" to "running a world in real time".

This time around, Genie 4 brings the world model straight into a whole new era of "voices, people, brains".

For the first time in the Genie series, native speakers generate voice and improvised dialogue.Simultaneously generating surround sound and soundtrack, it allows the characters to speak spontaneously depending on the emergency situation.

Two minutes of continuous memory, 24 threads of sliding interaction.A single coherent world maintains a time of doubling to about 2 minutes, while a real-time image is stable at about 24 frames per second.

Re-imagining real human and corporeal physics.With the help of Gemini Omni, real-life common sense, personality micro-expressions, mirror reflections and physical collisions have all jumped to their full potential.

The three Gemini combined to create a "mixed world model".The Gemini 3.8 Flash is shown in the background as "Director", and the Gemini 4 Argon dismantles a 22-step precision operation to command the robot to jump the chessboard.

You take a step forward and the world grows an inch in real time; you change a sentence, and the laws of physics and NPC rhetoric are rewritten on the spot.

At a time when video models were just a few seconds shorter than footage, Google had created a parallel universe of infinite trial and error for future general agents and robots.

Now, Google has begun to predict that Genie 4 will be available for Google AI Ultra users to apply for a waiting list.

To demonstrate the dominance of Genie 4, Google has released a series of stunning demos, interspersed styles, changes in detail, and stunning visuals.

And in front of Genie 4, you're no longer the audience, you're the "first-person gamer", the "god of the world".

Can you believe that the carnival on the Hong Kong Bridge below is not the original?

The girl in the back seats just shouted the words, and the players taped two lines: "The background exploded" and "The giant alien hand that destroyed the world".

Soon afterwards, the distant Viphong skyline caught fire and an alien hand fell directly from the sky.

The orange trailing shadows of the street lights, the texture of the woman's skin and the subtle expressions of concern were perfectly rendered in the long shot.

Last year's Genie 3 pioneered real-time interactive video, but there was a deadly shortcut with no sound and a loophole in real-life generation.

Google released a direct comparison of Genie 3 and Genie 4 on the same screen.

On the left, Genie 3 is a five-sensory sticky, rigid, voiceless paper clipper; on the right, Genie 4 is based on the reality of Gemini Omni and is capable of capturing cosmetic light, actors' micro-faces, and even real-time reversals of scenes, as sharp as a real-time long shot.

Ethan Mollick, a professor at the Wharton School of Business who qualified for the first test, unveiled the first test screen.

He only cried out "a neo-Gothic coarse city inhabited by horses and robots", and in less than a minute, the city appeared.

The single-cylinder delay is stable at 23.4 to 25.3 FPS. When you press the "Dragon Raid Air" button with your hand, a giant dragon moves its wings in a matter of seconds.

He added: "A friendly watermelon NPC came to ask a question".

Then, on the corner of the street, a woman in a suit of armor ran out and said in London's language, "Break it up, traveller".

And then he unleashed another divine test: in Genie 4, he's running around destroying DOOM warriors!

He set "going into an office where every computer and every wall is running at DOOM" and then threefold instructions: "Devil colleagues appear", "Gatherlin's machine gun in his hand", "Send me a bonus".

As soon as the sound came out, a demonic colleague really came out, his hands turning to the heavy gaitlin, and the prize was launched!

Justine Moore, a 16z partner, also unveiled the submersible experiment, saying it was a huge leap.

In the depths of the undersea caves, "cursed purple diamond glow" and "electrolyte cluster attack", there is an instant flash of purple light in the depths of the caves, a huge electrode cluster appears.

Once again, the ghost of the pirate captain appears, which is too real.

Bilawal Sidhu, former Google's AR/3D product manager, described Genie 4 as a real-time virtual scene for science fiction:

He summoned a UFO and uttered the phrase "it sucked me in".

According to Google, Genie 4 has improved four key features.

Simulating the world in real time

We can use natural language to modify simulation scenarios in real time, such as changing the weather, moving lenses, or introducing new characters or objects. Gemini will adjust our instructions according to the surrounding scenarios.

Sounds like a dream come true?

There was a man on the island who played the saxophone, knocked on the "tail appeared", and immediately a shark appeared in the water behind him.

Then we let the world start snowing, and the snow fell instantly.

If the lens is pulled to a first-person perspective, the impact of this real-time rewrite of the world will be even greater.

In this image, we are riding a giant dragon at 140 kilometers per hour to jump over the castle, and if the order "the castle fires a fireball", the tower will spark a flame, the fireball will rub over and it will be filled with excitement.

In the following scene, Genie 4 first generated the Antarctic Ice Sheet as instructed by the user, and then generated the first human-looking hot air balloon scene.

Expand the real boundaries between physics and human behavior

This time around, Genie 4 introduces a whole new set of reality standards for interactive simulation.

Based on its super-powerful understanding of world physics, Genie 4 is highly realistic in simulating real-world physical scenarios such as aerodynamics, collisions, etc.

The most significant advances in Genie 4 are in the simulation of characters, which can create roles that move naturally, react to the environment, and can be manipulated in real-time.

In rural Japan in the 1980s, if you give a prompt, the passengers will shout out a hint.

For one more tip, an elderly woman appeared with a suitcase.

The following image shows the extreme realism of Genie 4 physical interaction.

On the platform of the outer courtyard, the half marble was pointed toward the pumpkin.

With the sound of a horn, the pumpkin and the salt water flying from gravitational impact to the trajectory of the wreckage, Genie 4 recovered in real-time interaction.

Genie 4 is the same in terms of mechanics and aerodynamics.

The pink silk on the grass is very real, and when a wind blows, the wind immediately blows up the layer of silk.

The seeds of an oak tree fell to the ground, and the scenes were not realistic.

When the lens is pushed to a distance between millimeters, the wings of an oak tree slide like a helicopter's propeller, slowly stopping between the leaves.

Goodbye to the Silent Mirror, Genie 4 can simulate sound!

Genie 4 is Google's first native audio-generating real-world model.

Each world has its own surrounding sound, like the sound of the leaves of the sand, the waves of the sea beating the sand.

However, the real challenge is human language! Genie 4 introduces real-time interactive dialogue functionality. Characters will engage in real-time conversations and reactions based on your input or events.

In this world of translucent 3D stained glass, a deer is summoned with a hint word. The most fascinating thing is that its footsteps are automatically turned into crisp glass collisions.

Hard-core architecture dismantled: three Gemini teams take turns to build a "mixed world model" circuit

If Genie 4 could only be described as a higher-quality, loudspeaker video generator, it would underestimate Google DeepMind's ambitions.

In the past, real-time models of the world could collapse within a few seconds because pure pixel predictions lacked long-range logical planning, often taking the first step and completely forgetting the next goal.

In order to cure this disease, Google has done a very disruptive thing at this end.The architecture of the hybrid world modelThree Gemini models were shipped in collaboration with the Genie 4 team.

Thus, world simulation adds a whole new level of reasoning, allowing Genesis 4 to simulate complex multi-step events and test how they operate in different environments before introducing AI into the real world.

The first priority is to makeGemini 3.8 FlashHe is the "director-general" who acts as a real-time controller in the circular circuit.

When Genie 4 generates images in front of it, Gemini 3.8 Flash looks at the screen in real time to unravel the multi-step reasoning sequence while verifying whether the current step is complete and following the next step instructions.

With user intervention, Gemini immediately adjusted to create a world of reliable causes, programmer behavior, and multi-step reasoning.

Take a look at the close-up magic below: every time the player closes the box door, the Gemini 3.8 Flash in the upper right corner confirms the step is complete, commanding Genie 4 to turn the peacock into a cat and then turn the cat into a vintage radio!

Even if the play is changed halfway through, the "director-general" can rewrite the script on the spot.

For example, in a frozen, snow-capped land, a meteorite is drawn and then commanded by a gray crown to make a melody before the electronic organ, which finally makes it fly into the sky.

If Gemini 3.8 Flash solves "how the world evolves logically", then what is it?Gemini 4 ArgonwithGemini RoboticsIn addition, it broke the ceiling of physical intelligence and robot training.

When humans arrive at the phrase "Jump the chessboard with the opposite boy and walk the red chessboard one step", the Gemini 4 Argon on the left instantly breaks down the task into a 22-step mm-level action script.

These high-density directions are injected in real-time into the generation stream of Genie 4, where the right-hand-driven Apollo robot walks through the cloud to complete the chess, run, and match moves.

Genie 4 can even go into the lab for precise scientific research.

Even an industrial-grade UR5 six-axis mechanical arm can accurately hold a fluid-displacement tube in the Genie 4 and insert liquid drip into the micro-hole board.

Even if the target is removed in the middle, the model can recalculate the strategy in real time and capture it at a new location.

What is the real "world model"?

It's like you push a door, turn around and leave for 30 seconds, then come back, and the door is still open.

To remember the laws of physics, to maintain space memory, is to divide the video model from the world model!

Now, Google Genie 4 has done it.

Using "mass unlabeled video pre-training + minor movement trajectory tweaking", combined with Gemini's symbol logic and long-range memory, the Genie 4 move pushes the maintenance time of a single coherent world to about 2 minutes.

Since then, building interactive environments has completely put an end to busy manual 3D modeling.

Now that Genie 4 is open to Google AI Ultra users, it will be the first ever model to enter the world of the Gemini API with Google AI Studio, opening its doors to developers.

We can see what Google's ambition is:A real AGI cannot live in a dialog box!

When AI learns that there are tens of thousands of physical possibilities in the parallel universe that it creates, the day when it really opens the door to the real world is no longer far away.