🔍 Read the full analysis: Genie 4 Creates Worlds, But Can It Learn How Reality Works? on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Google DeepMind says Genie 4 can generate realistic environments that may help train AI agents, describing this work as a step toward AGI. But available research finds that video models’ visual realism does not reliably show physical understanding, and Genie 4 has not been tested on the cited physics benchmark.
Google DeepMind is presenting Genie 4’s realistic, generated environments as a way to train AI agents, and calls exploration of such worlds a “key stepping stone” toward AGI. But research cited in a November 1 analysis by Thorsten Meyer finds that video models can look physically convincing without demonstrating reliable physical understanding; Genie 4 itself has not been tested on the benchmark discussed.
Genie 4 generates video environments in response to previous frames and user input, such as a key press, mouse movement or written instruction. Meyer describes it as a model that predicts what the next frame should look and sound like, rather than a conventional simulation with an underlying game engine or stored three-dimensional map. A ball rolling downhill may appear to speed up and bounce, but that appearance alone cannot establish whether the model has learned the physical rules behind the motion.
Google’s stated case is that agents and robots need extensive practice in varied situations, including scenarios that would be slow, costly or dangerous to reproduce in the real world. DeepMind has reported that its SIMA 2 agent learned to navigate and follow instructions in newly generated Genie 3 environments. Waymo said in February 2026 that its driving simulator was built on Genie 3 and could generate unusual events, including a tornado or an encounter with an elephant. Those examples concern earlier versions, and company reports do not by themselves show that Genie 4 understands real-world physics.
The cited counterpoint comes from Physics-IQ, a benchmark published by Google researchers in 2025. According to Meyer’s account, the researchers found video models’ physical understanding to be “severely limited, and unrelated to visual realism.” On Physics-IQ Verified, where real videos score 100, the best image-to-video result reported in the source was 55.1, the best of eight tries. Gemini Omni 1.1 Flash, described in the source as part of the model family anchoring Genie 4, scored 35.3. Neither score is a Genie 4 result; the source says Genie 4 has not been tested on the benchmark.
THE AGI CLAIM, VERSION BY VERSION
Three Genies, one promise
Google DeepMind has tied every recent Genie to the path to AGI. Here is what each version could do, what was claimed, and who said it.
- 4 Dec 2024Genie 23D worlds from one image, consistent for up to about a minute.“…the breadth and generality required to progress towards AGI.”Company claim
- 5 Aug 2025Genie 3First real-time Genie: 720p, 24 frames per second, a few minutes of consistency.“World models are also a key stepping stone on the path to AGI…”Company claim
- 13 Nov 2025SIMA 2 in Genie 3 worldsDeepMind's game agent orients itself and self-improves in newly generated worlds.“…a major milestone toward training general agents across diverse, generated worlds.”Company report
- 6 Feb 2026Waymo World ModelWaymo says its driving simulator is built on Genie 3 and can generate rare events.“…from a tornado to a casual encounter with an elephant.”Company report
- 27 Oct 2026Genie 4Photorealism, native sound, live text steering, Gemini in the loop. Trusted testers only.“…a key stepping stone on the path to AGI.”Company claim
- 31 Oct 2026TechRadar pushes backColumnist Eric Hal Schwartz: realism is not understanding.“Faking plausibility is easier than grasping the underlying rules…”Opinion
The Genie 2 and Genie 3 capability notes are DeepMind’s own descriptions. No Genie post measures progress toward AGI; DeepMind published no benchmark numbers for Genie 4.
Sources: Google DeepMind, Genie 2 (4 Dec 2024), Genie 3 (5 Aug 2025), SIMA 2 (13 Nov 2025), Genie 4 model page (read 1 Nov 2026); Waymo (6 Feb 2026); TechRadar (31 Oct 2026, opinion).
Powered by Thorsten Meyer AI
Why Realism Is Not a Physics Test
The distinction matters because generated worlds may become part of how companies train and assess agents, including systems intended to act in physical settings. If a simulation captures only the appearance of an event, an agent trained there could behave well inside the generated environment yet fail when real objects, forces or unexpected conditions differ. The cited benchmark results do not establish how any particular deployed system would perform, but they show why visual quality alone is a weak basis for judging physical competence.
For businesses considering world-model tools, the practical question is not only whether a demo looks convincing. It is whether the system reproduces relevant physical behavior, how it performs on measurable tests, and whether results transfer from generated settings to real ones. Meyer’s analysis argues that companies should treat these models as tools for sketching and stress-testing ideas, rather than as proof that a robot, vehicle or plan will work in reality. For readers, Google’s “step toward AGI” wording describes a research direction, not a measured finding that Genie 4 has achieved general intelligence or learned the rules of the physical world.
Products Worth Considering
Logitech MX Brio Ultra HD 4K Webcam, Graphite
Canon SELPHY CP1500 Compact Photo Printer (White)
FLEXISPOT 60 × 24-Inch Electric Standing Desk, White
As an affiliate, we earn on qualifying purchases.
Google’s World-Model Research Bet
Google’s framing predates Genie 4. According to Meyer, DeepMind used “a key stepping stone on the path to AGI” in its Genie 3 material in August 2025 and described Genie 2 in December 2024 as pointing toward the breadth and generality needed to progress toward AGI. Across the versions, the argument is that simulated environments could give agents a broad place to practise before deployment. The repeated language signals a continuing research ambition; it is not a benchmark demonstrating progress toward AGI.
The term “world model” can refer to different kinds of systems. Meyer distinguishes a learned model in which an agent imagines possible outcomes and practises, a video-generating system such as Genie 4, and a structured model built from data and rules. These categories should not be treated as interchangeable. A video can depict a bridge, for example, without calculating how much weight it can support. The source also says Genie 4 uses a Gemini model to plan scenes and check frames while Genie generates the visuals, and quotes DeepMind describing hybrid world models as a promising direction.
“The ability to explore an unlimited range of realistic environments is a key stepping stone on the path to AGI.”
— Google DeepMind, on the Genie 4 model page, as quoted by Thorsten Meyer
What Genie 4 Has Not Shown
The source reports no published Physics-IQ score for Genie 4 and no benchmark numbers on its model pages measuring progress toward AGI. Results for other video models, including Gemini Omni 1.1 Flash, cannot be treated as Genie 4 results. The reported SIMA 2 and Waymo examples concern Genie 3, and the source does not establish how well those systems’ capabilities transfer to real-world tasks.
It also remains unclear how reliably Genie 4 represents physical causes rather than reproducing patterns learned from video, or how it performs across different objects, conditions and rare events. The available material supports neither a conclusion that Genie 4 understands reality nor a conclusion that it cannot contribute to useful training. Independent evaluations and direct tests of transfer to physical tasks would be needed to answer those questions.
Tests Needed Beyond Demo Clips
The next meaningful evidence would include a clearly described evaluation of Genie 4 on physical reasoning tasks, with results that can be compared against real footage and other models. Tests would also need to show whether agents trained in Genie-generated environments perform better on tasks outside those environments, and whether gains carry over to robots or other systems acting in the physical world.
Until such results are available, Google’s AGI language should be read as a statement of research aims, not as proof that Genie 4 understands physical laws. Companies assessing the technology can ask vendors for task-specific physics results and transfer tests, rather than relying on the appearance of a short demo. The source material does not specify a date for a Genie 4 benchmark or a further independent evaluation.
Key Questions
Does Genie 4 understand how the physical world works?
The available source does not establish that. Genie 4 can generate convincing environments, but the cited Physics-IQ results concern other models; Genie 4 has not been tested on that benchmark, according to Meyer.
What does Google mean by calling Genie 4 a step toward AGI?
Google’s stated argument is that realistic generated environments could give agents and robots many situations in which to practise. The phrase describes a research direction, not a measurement showing that Genie 4 is close to AGI.
What did the Physics-IQ benchmark find?
As reported by Meyer, Google researchers found video models’ physical understanding “severely limited, and unrelated to visual realism.” The source gives a score of 55.1 for the best image-to-video result across eight tries and 35.3 for Gemini Omni 1.1 Flash on Physics-IQ Verified, where real videos score 100. These are not Genie 4 scores.
Have agents learned in Google-generated worlds?
DeepMind reported that SIMA 2 could navigate and follow instructions in newly generated Genie 3 worlds. Waymo said its driving simulator was built on Genie 3. These reports offer examples of generated environments being used, but they do not establish that Genie 4 has passed physical-understanding tests or that performance transfers to reality.
Source: ThorstenMeyerAI.com
