🔍 Read the full analysis: Genie 4’S Big Question: Can AI Move From Imagined Worlds To Reality? on ThorstenMeyerAI.com
Get the latest gadgets 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 describes Genie 4’s realistic, generated environments as a step toward artificial general intelligence, arguing they could give agents and robots places to practise. But the claim is a research direction, not a measured result: existing tests find a gap between realistic video and physical understanding, and Genie 4 itself has not been tested on the cited benchmark.
Google DeepMind says Genie 4’s generated environments could help move AI toward artificial general intelligence, but the company has not published benchmark results showing that the model understands real-world physics. A November 1 analysis by Thorsten Meyer examines that gap: existing tests suggest that video models can look realistic without reliably representing the rules behind what they depict.
On Genie 4’s model page, DeepMind calls exploring “an unlimited range of realistic environments” a key stepping stone on the path to AGI. The company’s argument is that agents and robots need many places to practise, including situations that are expensive, risky or difficult to reproduce in the physical world. Generated environments could provide that practice at scale. The statement describes a research goal; it does not report a measured advance toward AGI.
The distinction matters because Genie 4 generates video-like environments from earlier frames and user input rather than relying on a conventional game engine or an underlying stored 3D map, according to Meyer’s account. A scene can appear convincing without the model correctly predicting what an object would do under real physical forces. Meyer notes that DeepMind’s own Physics-IQ benchmark, published in 2025, found video models’ physical understanding “severely limited, and unrelated to visual realism.”
On Physics-IQ Verified, where real videos score 100, the best image-to-video result cited by Meyer is 55.1, the best of eight attempts. Gemini Omni 1.1 Flash, described in the source as part of the family DeepMind says anchors Genie 4, scores 35.3. Those figures are results for the named benchmark entries, not a Genie 4 score: Genie 4 has not been tested on it in the material provided.
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 Realistic Worlds Need Physics Tests
Generated worlds could make it cheaper and safer to train AI agents, particularly for rare events that are hard to stage with real equipment or people. DeepMind has reported that its SIMA 2 agent learned to navigate and follow instructions in worlds generated by Genie 3. Waymo has said its driving simulator is built on Genie 3 and can create unusual scenarios. These are reported applications of an earlier model; they do not establish that Genie 4 has learned transferable physical rules.
For businesses considering world-model tools, the practical distinction is between useful simulation and evidence that a plan or machine will work in reality. A visually persuasive demonstration alone cannot establish that a simulated robot, vehicle or construction plan will behave safely outside the model. Meyer’s analysis advises companies to ask vendors for physics tests, not just demo clips. For readers, DeepMind’s “step toward AGI” wording is best understood as the company’s research thesis rather than a benchmark finding.
Products Worth Considering
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Black
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
As an affiliate, we earn on qualifying purchases.
Three Different Ideas Called World Models
The term “world model” covers systems with different capabilities. Meyer distinguishes a model that lets an agent imagine future states and practise, a video generator that predicts subsequent frames, and a structured model built on data and explicit rules. Genie 4 is described as the video-generating kind: it produces scenes in response to prior frames and inputs. That can support exploration, but visual output alone does not show that the model has an accurate internal account of cause and effect.
The idea of practising in an imagined environment predates Genie. In 2018, researchers David Ha and Jürgen Schmidhuber described training an agent inside a learned model of a simple video game and then applying its skill to the game itself. DeepMind has linked its Genie series to similar ambitions: its Genie 3 material in 2025 used “a key stepping stone on the path to AGI,” while its 2024 Genie 2 announcement pointed to progress toward the breadth and generality required for AGI. The repeated language sets out a direction, not a test result.
Other systems use different foundations. Meyer cites construction start-up Dream, which TechCrunch reported on October 29 was training a model using more than 100 million square feet of building data from U.S. permit archives. Such structured data is not the same as generated video. The comparison highlights the open question: whether a system that paints plausible scenes can also learn the rules that govern them.
“Severely limited, and unrelated to visual realism.”
— Google DeepMind researchers, Physics-IQ findings as summarized by Thorsten Meyer
Genie 4’s Physics Remains Untested
The material available for this report provides no Genie 4 score on Physics-IQ and no benchmark figures from DeepMind that measure the model’s physical reasoning or progress toward AGI. Results for other video models cannot be treated as Genie 4 results, even if the systems are related. The reported SIMA 2 and Waymo uses involve Genie 3, so they also do not independently verify Genie 4’s capabilities.
It is also unclear how well Genie 4’s generated environments match physical conditions across unfamiliar scenarios, or whether skills gained there transfer reliably to real robots and other systems. DeepMind’s approach is described as hybrid: a Gemini model plans a scene and checks frames while Genie generates them. The source says the company views hybrid world models as promising, but offers no comparative evidence here showing how much that design improves physical understanding.
Evidence Needed Beyond Generated Scenes
The next meaningful test would be a published evaluation of Genie 4 on physical-reasoning tasks, with methods and comparisons that let researchers distinguish visual quality from accurate prediction of cause and effect. Results from real-world transfer would also help establish whether agents trained in generated environments perform better outside them. Until such evidence appears, the available comparisons apply to other models and earlier Genie applications, not to Genie 4 itself.
For now, DeepMind’s AGI language should be read as a research ambition. Companies evaluating the technology can ask for task-specific physics evaluations and evidence of transfer to real settings; readers can separate that evidence from convincing clips and broad claims about future potential.
Key Questions
Has Genie 4 been shown to understand real-world physics?
Not in the material cited here. The report says Genie 4 has no published score on the Physics-IQ benchmark, so its physical understanding cannot be inferred from the scores of other models.
What does DeepMind mean by calling Genie 4 a step toward AGI?
DeepMind argues that realistic generated environments could let agents and robots practise many tasks. The wording describes the company’s research direction, not a benchmark result or a claim that Genie 4 is close to AGI.
What did the cited Physics-IQ results show?
As summarized in the report, the 2025 benchmark found video models’ physical understanding “severely limited, and unrelated to visual realism.” On Physics-IQ Verified, where real videos score 100, the cited best image-to-video result was 55.1 across eight attempts; Gemini Omni 1.1 Flash scored 35.3.
Are SIMA 2 and Waymo evidence of Genie 4’s performance?
No. The reported examples involve Genie 3: DeepMind described SIMA 2 learning in generated worlds, and Waymo said its driving simulator uses Genie 3. They provide context for the broader approach but do not test Genie 4.
Source: ThorstenMeyerAI.com
