Back to the trend of the day: Google's Genie 4 AI builds worlds to walk through; invited testers only
Poly Rating · from the trend of the day

Genie 4 claims ranked by evidence gap (October 2026)

Google's own post says Genie 4's headline trick, rewriting a live world by typing, makes details and objects drift after several edits, and no public benchmark can test edits mid-walk yet.

  • No. 1: Genie 4 real-time natural-language steering with persistent memory
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    Genie 4 real-time natural-language steering with persistent memory

    This is the headline feature, and Google admits objects drift 'after multiple prompt edits'. No outsider has tried it. No open benchmark can test edits made mid-walk.

    Included it: Claude, ChatGPT, Meta, Nemotron, Gemini, Kimi, Wenxin, Qwen. Did not include: Grok, Nova, StepFun, DeepSeek.

    In their lists: Claude: 1 · ChatGPT: 1 · Meta: 1 · Nemotron: 1 · Gemini: 2 · Kimi: 1 · Wenxin: 1 · Qwen: 1

    What each AI noted 8
    • Claudeplace 1

      The headline feature is the one Google itself says breaks after 'multiple prompt edits', and no public test edits a world mid-walk.

      Source ↗
    • ChatGPTplace 1

      The decisive test is not changing the weather once; it is changing the world repeatedly without losing what was already there.

      Source ↗
    • Metaplace 1

      No open benchmark can test edits-while-walking — Google admits objects drift after edits while 9 models fail revisit memory

      Source ↗
    • Nemotronplace 1

      Google admits drift in long simulations; iWorld-Bench shows no model above 0.85 memory symmetry — steering that holds coherence past 2 minutes is the largest evidence gap

    • Geminiplace 2

      Pairwise arenas like Artificial Analysis evaluate static prompt execution, leaving mid-flight interactive prompt steering completely unbenchmarked by third parties.

      Source ↗
    • Kimiplace 1

      The one benchmark family that exists says even the best interactive world model forgets — no model in iWorld-Bench tops 0.85 memory symmetry, and Genie 4 has never been run on it.

      Source ↗
    • Wenxinplace 1

      Demos show talking the world into changing — but every benchmark says memory is the universal weakest link.

    • Qwenplace 1

      Memory symmetry caps below 0.85 across the board, meaning interactive recall is the true bottleneck of world models.

    8 AIs put it here
    Source ↗
  2. 2

    Genie 4 native audio and interactive dialogue

    Google calls it the first world model with native real-time audio. Its own team told TechCrunch audio accuracy is 'not quite figured out', and a tester heard clarinet for a saxophone prompt.

    Included it: Claude, ChatGPT, Meta, Nemotron, Gemini, Kimi, Wenxin, Qwen. Did not include: Grok, Nova, StepFun, DeepSeek.

    In their lists: Claude: 2 · ChatGPT: 5 · Meta: 2 · Nemotron: 2 · Gemini: 7 · Kimi: 2 · Wenxin: see list · Qwen: 2

    What each AI noted 8
    • Claudeplace 2

      'First real-time world model to generate audio natively', yet one commenter heard a clarinet when the prompt asked for a saxophone.

      Source ↗
    • ChatGPTplace 5

      A talking character is not yet a reliable conversational agent: speech must remain synchronized with action, context and changing events.

      Source ↗
    • Metaplace 2

      First real-time world model with native dialogue — but audio accuracy still listed as active research with no open dialogue coherence score

      Source ↗
    • Nemotronplace 2

      SynthID watermarking on generated audio is claimed; Artificial Analysis arena cannot test interactive audio sync — only static video pairwise votes

    • Geminiplace 7

      Real-time interactive generation of dialogue, footsteps, and spatial acoustics outpaces video-only peers, but latency penalties under streaming loads remain undisclosed.

      Source ↗
    • Kimiplace 2

      Google's own engineers told reporters they 'haven't quite figured out' accurate audio — the flagship new claim is the one its maker publicly hedged within hours.

      Source ↗
    • Wenxin

      Named differently in its own list: open the full list.

      see list
    • Qwenplace 2

      Generates sound natively alongside the simulated world, bypassing the lip-sync gap of traditional post-production audio.

      Source ↗
    8 AIs put it here
    Source ↗
  3. 3

    Genie 4 photorealism, lighting and materials

    This is the easiest claim to check and has the most support: commenters praise the lighting. A related Gemini Omni Flash 1.1 ranks 4th of 23 on blind image-to-video votes; that is a related model and one source.

    Included it: Claude, ChatGPT, Meta, Nemotron, Gemini, Kimi, Wenxin, Qwen. Did not include: Grok, Nova, StepFun, DeepSeek.

    In their lists: Claude: 9 · ChatGPT: 9 · Meta: 5 · Nemotron: 5 · Gemini: 9 · Kimi: 7 · Wenxin: 5 · Qwen: 7

    What each AI noted 8
    • Claudeplace 9

      The most testable claim and the one with the most support: commenters call lighting and materials clearly better than the previous generation.

      Source ↗
    • ChatGPTplace 9

      The easiest Genie 4 capability to admire is also among the easiest to overestimate from selected footage.

      Source ↗
    • Metaplace 5

      Most realistic and interactive worlds yet — but no resolution, fps, or VBench score disclosed unlike Genie 3's 720p 24fps

      Source ↗
    • Nemotronplace 5

      Artificial Analysis arena measures lighting/materials via blind votes on static clips — cannot verify temporal consistency during navigation

    • Geminiplace 9

      The easiest claim to verify visually, backed by high leaderboard placements, yet cinematic lighting contributes least to agent functional competence.

      Source ↗
    • Kimiplace 7

      The prettiest claim has the shortest gap: the Gemini Omni family anchoring Genie 4's look already sits top-5 on independent blind-vote video Elo.

      Source ↗
    • Wenxinplace 5

      Prettier pixels are the easiest thing to cherry-pick — and the hardest thing to compare across cherry-picked clips.

    • Qwenplace 7

      Prettier pixels impress on day one, but visual fidelity is the easiest gap to close with brute-force rendering.

      Source ↗
    8 AIs put it here
    Source ↗
  4. 4

    Genie 4 SynthID watermark on all video and audio

    This is a procedure, not a performance claim, and it runs on infrastructure Google already uses. The only gap is that no independent test of detection on Genie 4 output exists yet.

    Included it: Claude, ChatGPT, Meta, Nemotron, Gemini, Kimi, Wenxin, Qwen. Did not include: Grok, Nova, StepFun, DeepSeek.

    In their lists: Claude: 10 · ChatGPT: 10 · Meta: 10 · Nemotron: 7 · Gemini: 10 · Kimi: 10 · Wenxin: 3 · Qwen: 10

    What each AI noted 8
    • Claudeplace 10

      A procedural claim, not a performance claim, and the easiest to check once outputs are public.

      Source ↗
    • ChatGPTplace 10

      Google promises watermarks on both audio and video, but deployment coverage is not the same as independently proven detection reliability.

      Source ↗
    • Metaplace 10

      Only claim with built-in detector — all Genie 4 video and audio carry imperceptible SynthID across modalities

      Source ↗
    • Nemotronplace 7

      Claimed at launch; no independent audit of watermark robustness under compression, editing, or adversarial removal

    • Geminiplace 10

      Google's digital provenance footprint is technically established, facing an evidence gap only regarding its survival after low-latency frame downscaling.

    • Kimiplace 10

      The smallest gap on the list: Google's watermark already ships across its model family, so this claim rides on proven infrastructure rather than promises.

      Source ↗
    • Wenxinplace 3

      Audio that moves with the scene sounds like the hardest un-fakeable claim — until SynthID limits surface.

    • Qwenplace 10

      Watermarking generated audio and video is a safety necessity, but it adds zero capability to the world model itself.

      Source ↗
    8 AIs put it here
    Source ↗
  5. 5

    Genie 4 for training robots and AI agents

    The evidence is one demo: a simulated Apptronik Apollo humanoid playing checkers. No sim-to-real numbers are published, and no outside lab has trained a robot in it.

    Included it: Claude, ChatGPT, Nemotron, Gemini, Kimi, Wenxin, Qwen. Did not include: Meta, Grok, Nova, StepFun, DeepSeek.

    In their lists: Claude: 3 · ChatGPT: 2 · Nemotron: 9 · Gemini: 3 · Kimi: 3 · Wenxin: 2 · Qwen: 3

    What each AI noted 7
    • Claudeplace 3

      The agent-training story rests on one humanoid-checkers experiment, and one commenter says Genie now looks more like a controllable video generator.

      Source ↗
    • ChatGPTplace 2

      A simulator that teaches an agent the wrong physics can produce an agent that fails precisely when reality matters.

      Source ↗
    • Nemotronplace 9

      No published agent-training benchmarks using Genie 4; iWorld-Bench trajectory accuracy exists but not downstream policy transfer

      Source ↗
    • Geminiplace 3

      Positioned as the training engine for autonomous robotics, yet zero published sim-to-real robotic policy transfer data accompanied the showcase.

    • Kimiplace 3

      The entire public evidence for agent-training usefulness is one demo where Gemini 4 Argon breaks a task into steps — no sim-to-real validation, no benchmark, no external robotics lab.

      Source ↗
    • Wenxinplace 2

      If it can't hold a two-minute memory, the training runway Google promises is shorter than the marketing.

    • Qwenplace 3

      Trusted-tester cohorts in robotics and gaming determine if this is actually a simulator or just a very pretty screensaver.

      Source ↗
    7 AIs put it here
    Source ↗
  6. 6

    Genie 4 two-minute sessions and object permanence

    Press reports about two minutes, double Genie 3. Google's post gives no number. In Hacker News examples a second castle appears once objects leave and re-enter the frame.

    Included it: Claude, ChatGPT, Meta, Gemini, Kimi, Qwen. Did not include: Nemotron, Grok, Nova, Wenxin, StepFun, DeepSeek.

    In their lists: Claude: 5 · ChatGPT: 3 · Meta: 3 · Gemini: 4 · Kimi: 5 · Qwen: 4

    What each AI noted 6
    • Claudeplace 5

      Two minutes, double Genie 3, but Google's own post says only 'a few minutes' and lists long-term consistency as open research.

    • ChatGPTplace 3

      About two minutes of reported simulation is not evidence that a room, person or object will remain stable across an extended task.

      Source ↗
    • Metaplace 3

      2 minutes claimed vs 1 minute Genie 3 — TechCrunch says double, Google's own post avoids exact number and admits drift over longer runs

      Source ↗
    • Geminiplace 4

      While Google highlights open-ended sandbox play, independent reporter briefings revealed a hard boundary of roughly 120 seconds before structural hallucination sets in.

    • Kimiplace 5

      Two minutes is double Genie 3 — and Google's own Hang Qi told Bloomberg that reaching the limit is 'no guarantee' the scene holds together.

      Source ↗
    • Qwenplace 4

      Simulations hit a hard ceiling of about two minutes before admitted drift fractures the physical logic of the environment.

      Source ↗
    6 AIs put it here
    Source ↗
  7. 7

    Genie 4 physics: aerodynamics, collisions, water

    The New Stack's headline says the physics 'still aren't quite right', and Google lists complex physical interactions as unsolved. No open test measures interactive physics.

    Included it: Claude, ChatGPT, Meta, Gemini, Kimi, Qwen. Did not include: Nemotron, Grok, Nova, Wenxin, StepFun, DeepSeek.

    In their lists: Claude: 6 · ChatGPT: 6 · Meta: 6 · Gemini: 6 · Kimi: 4 · Qwen: 6

    What each AI noted 6
    • Claudeplace 6

      The New Stack's headline says the physics 'still aren't quite right', and Google lists complex physical interactions as unsolved.

      Source ↗
    • ChatGPTplace 6

      A convincing collision in a demo is not a measured guarantee of correct collisions across unfamiliar situations.

      Source ↗
    • Metaplace 6

      Aerodynamics and collisions highlighted as leap — yet physics still listed as ongoing research with no open physics Elo for Genie

      Source ↗
    • Geminiplace 6

      Navigating environments with keyboard inputs masks persistent object clipping and boundary pass-through errors that formal physical simulators do not allow.

      Source ↗
    • Kimiplace 4

      The only physics yardstick that exists can't touch Genie 4: WorldScore dynamics metrics need model access, and the best video model ever scored just 40.22/100 on camera control there.

      Source ↗
    • Qwenplace 6

      Pairwise human preferences rank physics, but interactive collisions often break when unseen variables compound over time.

      Source ↗
    6 AIs put it here
    Source ↗
  8. 8

    Genie 4 trusted-tester use in film, journalism, education, gaming, engineering

    Five industries are named, but no tester counts and no published work exist. Turning archival photos into walkable scenes runs into Google's own warning that real places are imperfect.

    Included it: Claude, ChatGPT, Meta, Nemotron, Kimi, Wenxin. Did not include: Gemini, Grok, Nova, Qwen, StepFun, DeepSeek.

    In their lists: Claude: 8 · ChatGPT: 8 · Meta: 9 · Nemotron: 6 · Kimi: 8 · Wenxin: 6

    What each AI noted 6
    • Claudeplace 8

      Real use cases are named, but the post gives no tester counts, sample sizes or quantified results.

      Source ↗
    • ChatGPTplace 8

      Turning one archival photograph into a walkable scene is compelling, but plausible geometry can still be invented geometry.

      Source ↗
    • Metaplace 9

      Film and journalism cohorts cited as evidence — but access is limited to invited testers only, zero independent hands-on at launch

      Source ↗
    • Nemotronplace 6

      Limited invited testers only; no public API; utility claims rest on private feedback — zero independent workflow validation

    • Kimiplace 8

      Five industries, zero published artifacts — every reported workflow claim traces back to a Google briefing, not a tester showing work.

      Source ↗
    • Wenxinplace 6

      The biggest evidence gap isn't technical — it's that only Google's friends have touched it so far.

      Source ↗
    6 AIs put it here
    Source ↗
  9. 9

    Genie 4 multi-step event simulation guided by Gemini 3.8 Flash

    Gemini plans, checks each step and revises, but no task-completion rate is published. Accounts also disagree on which Gemini directs the loop: 3.8 Flash, 4 Argon or 3 Argon.

    Included it: Claude, ChatGPT, Meta, Nemotron, Wenxin. Did not include: Gemini, Grok, Nova, Kimi, Qwen, StepFun, DeepSeek.

    In their lists: Claude: 4 · ChatGPT: 4 · Meta: 4 · Nemotron: see list · Wenxin: see list

    What each AI noted 5
    • Claudeplace 4

      A reasoning model plans, checks each step and revises, but no published test shows it keeps a world coherent.

      Source ↗
    • ChatGPTplace 4

      Gemini 3.8 Flash can plan and inspect the scene, but that does not prove the generated world obeys the plan.

      Source ↗
    • Metaplace 4

      Apollo humanoid playing checkers demo shows reliable causality — but step-success verification has no public metric

      Source ↗
    • Nemotron

      Named differently in its own list: open the full list.

      see list
    • Wenxin

      Named differently in its own list: open the full list.

      see list
    5 AIs put it here
    Source ↗
  10. 10

    Genie 4 realistic people and multi-character interaction

    DeepMind calls people its 'most dramatic leap', yet admits interactions among several independent agents are still unsolved. On WorldMemBench, the leader triggers only 31.7% of prescribed events.

    Included it: Claude, ChatGPT, Meta. Did not include: Nemotron, Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.

    In their lists: Claude: 7 · ChatGPT: 7 · Meta: 7

    What each AI noted 3
    • Claudeplace 7

      Google concedes that modelling interactions among multiple independent agents is 'still an open research problem'.

      Source ↗
    • ChatGPTplace 7

      A playable-looking scene created in seconds is valuable, but an editable prototype is not automatically a stable game.

      Source ↗
    • Metaplace 7

      Most dramatic leap in people per DeepMind — WorldMemBench finds human identity retention is hardest static memory metric

      Source ↗
    3 AIs put it here
    Source ↗

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three claims have different kinds of evidence: Google's own admission, a team quote and a related model's vote rank, so the gaps are not measured the same way.

1Genie 4 real-time natural-language steering with persistent memory

No. 1 — Genie 4 real-time natural-language steering with persistent memory. On the lists of 8 of 12 AIs; median place there: 1. Google itself admits objects drift after several edits, and no public test can check edits made while walking through the world.

Did you know

  • Genie 3 builds its worlds frame by frame, with no explicit 3D model like NeRFs or Gaussian Splatting use. Google says its consistency is 'emergent', meaning it appears without being programmed in. That is why memory over long edits is hard to test.

    Source ↗
What it is
Typing to change the world while you explore it, with the world remembering.
Its edge
Widest gap: the maker admits drift and no open benchmark tests it.
Who should know
Useful for anyone judging if Genie 4 is more than a demo.
  • Gap to independent proofDrift admitted; no test exists
Why it is here

  • TechCrunch says Genie 4 can keep a scene consistent for up to two minutes, double its predecessor's limit. TechCrunch also says details and objects can still drift the longer a simulation runs.

    Source ↗

  • Google's own post says object permanence can still shift after multiple complex prompt edits. The page gives no figures for how many edits it takes, so no open test can measure this claim yet.

    Source ↗
History

  • Google announced 'promptable' world events for Genie 3 in August 2025. When Project Genie opened to US Ultra subscribers on January 29, 2026, those events were still missing.

    Source ↗
Keep in mind

  • DeepMind says visual details and object permanence "can still shift or drift over time", especially after "multiple complex prompt edits". Only trusted testers have access so far, and the page lists no independent scores.

    Source ↗
Where to go further

  • Join the Project Genie waitlist. Genie 4 is 'coming soon' there for Google AI Ultra subscribers. Once you have access, change the scene several times and check whether objects stay where they were.

    Source ↗
Compare the top three ↓

2Genie 4 native audio and interactive dialogue

No. 2 — Genie 4 native audio and interactive dialogue. On the lists of 8 of 12 AIs; median place there: 2. Google calls native audio a first, but its own team says accurate sound is not solved, and no audio scores exist.

What it is
Sound and talking characters made live inside the generated world.
Its edge
Backed by a maker's hedge, but still less open than steering.
Who should know
Useful for anyone wondering if the sound can be trusted.
  • Gap to independent proofNo audio score published
Why it is here

  • Unite.AI reports Genie 4 as Google's first real-time world model to generate audio natively. The same page lists real-time audio accuracy among areas of ongoing research, and it gives no audio quality figures.

    Source ↗

  • TechCrunch quotes the Genie 4 team: they have not quite figured out how to make sure the audio is accurate. The page gives no numbers for audio. The record claim is a "first", and the weak spot is accuracy.

    Source ↗
History

  • Genie 4, announced October 27, 2026, adds sound to the Genie line. TechCrunch reports the team hasn't yet worked out how to make the audio accurate.

    Source ↗
Where to go further

  • Watch the demo clips with the sound on. Google says real-time audio accuracy is still 'active areas of ongoing research', so check whether the sounds match what happens on screen.

    Source ↗
Compare the top three ↓

3Genie 4 photorealism, lighting and materials

No. 3 — Genie 4 photorealism, lighting and materials. On the lists of 8 of 12 AIs; median place there: 7. Looks are the easiest claim to check, and a related model already scores well in blind votes, so the gap is smaller.

Did you know

  • DeepMind's Jack Parker-Holder said interactive world generation trails video generation by 'roughly six to 12 months in terms of accuracy'. In one Genie 3 demo, a character walked through cacti with no effect, because the model isn't physics-aware yet.

    Source ↗
What it is
Realistic lighting, shadows and materials in the generated worlds.
Its edge
Easiest to judge by eye, with some outside votes for a related model.
Who should know
Useful for anyone who wants to know how far to trust the pretty demos.
  • Gap to independent proof4th of 23 (related model)
Why it is here

  • Google's post names no metrics for Genie 4 realism, and the "unprecedented realism" claim has no figures behind it. It lists no lighting or material limits, so the claim is easy to check but lacks outside data.

    Source ↗
Did you know

  • Genie 3 can show the same street in different weather and light, such as a snowy New York block or a London street in 'rare sunshine'. It uses Street View imagery, which spans 20 years and over 280 billion images from 110 countries.

    Source ↗
History

  • Genie 2, released in December 2024, could only generate worlds lasting 10 to 20 seconds. Genie 3 in August 2025 stretched visual memory to about one minute, at 720p and 24 fps.

    Source ↗
Keep in mind

  • Google says it has started rolling Genie 4 out to "a limited group of trusted testers". Ultra subscribers and API partners come "soon", so most readers cannot yet check the lighting themselves.

    Source ↗
Where to go further

  • Watch the demo clips on this page and look closely at shadows, reflections and water. The page does not describe lighting in detail, so judge it with your own eyes.

    Source ↗
Compare the top three ↓

The top three side by side

Genie 4 real-time natural-language steering with persistent memoryGenie 4 native audio and interactive dialogueGenie 4 photorealism, lighting and materials
In the AIs' lists8 / 128 / 128 / 12
Gap to independent proofDrift admitted; no test existsNo audio score published4th of 23 (related model)
Put first byClaude, ChatGPT, Meta, Nemotron, Kimi, Wenxin, Qwen——

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

More ratings

03Worth knowing

What the AIs noticed while building the list.

Google published fewer specs than for Genie 3: no resolution, frame rate, session length or tester count. The most concrete number, two minutes, comes from press briefings.

Noticed by:
Included it: Claude, Meta. Did not include: ChatGPT, Nemotron, Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.
Source ↗

The world holds together for things that stay in frame and breaks when objects leave and come back. An agent-training simulator needs exactly that case to work.

Noticed by:
Included it: Claude. Did not include: ChatGPT, Meta, Nemotron, Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.
Source ↗

Accounts disagree on what 'native audio' means. DeepMind describes one model doing everything, while a MindStudio write-up describes separate image and Lyria audio models orchestrated together.

Noticed by:
Included it: Kimi. Did not include: Claude, ChatGPT, Meta, Nemotron, Gemini, Grok, Nova, Wenxin, Qwen, StepFun, DeepSeek.
Source ↗

Every independent yardstick measures a neighbouring slice: the arena votes on fixed clips, iWorld-Bench never ran Genie 4, and WorldScore needs model access.

Noticed by:
Included it: Claude, ChatGPT, Nemotron, Kimi. Did not include: Meta, Gemini, Grok, Nova, Wenxin, Qwen, StepFun, DeepSeek.
Source ↗

04How we counted

In plain words, no formulas.

1The question

Top 10 Genie 4 claims ranked by evidence gap

2Where it applies

global AI world models

3What we looked at

  • Strength of the public claim in Google’s announcement and demos
  • Independent coverage: what reporters say testers saw versus what Google states
  • Benchmark coverage: whether existing open tests can even measure the claim
  • Failure consequence: how damaging the admitted limitation is if it shows up in real workflows
  • Google DeepMind: Genie 4 official post — The claims and the limits Google admits: drift, object permanence, audio accuracy, complex physics, multi-agent interaction, text rendering. It gives no resolu
  • Quartz (reporting Bloomberg) launch report — Trusted-tester-only access, keyboard navigation, and a session limit of about two minutes. DeepMind's Hang Qi says reaching that limit does not guarantee t
  • The Verge launch report — Whether reporters could try it themselves (they could not), access status, and about two minutes, up from one minute.
  • TechCrunch launch report — Briefing details: audio accuracy 'not quite figured out', Gemini acting as director, scenes consistent for up to two minutes.

4Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Nemotron, Gemini, Grok, Nova, Kimi, Wenxin, Qwen, StepFun, DeepSeek.

5How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

05Each AI's own list

🇺🇸 Claude10 places · 3 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    Real-time natural-language steering of a live world (weather, camera, adding people or objects)

    The headline feature is the one Google itself says breaks after 'multiple prompt edits', and no public test edits a world mid-walk.

    Source ↗
  2. 2
    Interactive dialogue that adapts to user input (native audio)

    'First real-time world model to generate audio natively', yet one commenter heard a clarinet when the prompt asked for a saxophone.

    Source ↗
  3. 3
    Usefulness for training agents and robots (SIMA, Gemini Robotics, Apptronik Apollo checkers)

    The agent-training story rests on one humanoid-checkers experiment, and one commenter says Genie now looks more like a controllable video generator.

    Source ↗
  4. 4
    Complex event simulation guided by Gemini 3.8 Flash

    A reasoning model plans, checks each step and revises, but no published test shows it keeps a world coherent.

    Source ↗
  5. 5
    Long-horizon consistency and memory (about two minutes)

    Two minutes, double Genie 3, but Google's own post says only 'a few minutes' and lists long-term consistency as open research.

  6. 6
    Physics simulation (aerodynamics, collisions, flowing water)

    The New Stack's headline says the physics 'still aren't quite right', and Google lists complex physical interactions as unsolved.

    Source ↗
  7. 7
    Characters that move, react and behave naturally (including multi-agent interaction)

    Google concedes that modelling interactions among multiple independent agents is 'still an open research problem'.

    Source ↗
  8. 8
    Trusted-tester value in film, journalism, education, gaming and engineering

    Real use cases are named, but the post gives no tester counts, sample sizes or quantified results.

    Source ↗
  9. 9
    'Most realistic' photorealism, lighting and materials

    The most testable claim and the one with the most support: commenters call lighting and materials clearly better than the previous generation.

    Source ↗
  10. 10
    SynthID watermarking on all generated video and audio

    A procedural claim, not a performance claim, and the easiest to check once outputs are public.

    Source ↗

What Claude noticed

Google's own post lists no resolution, frame rate, session length or tester count. The most concrete number in the coverage, two minutes, comes from press reports.

Source ↗

Hacker News commenters say Genie 4 holds coherence well for things that stay in frame and breaks when objects leave and re-enter. That is exactly the case an agent-training simulator needs.

Source ↗

Part of the pitch is a hybrid design: Gemini handles structure and planning while Genie renders the look, sound and response. That makes the claims about Gemini models and Genie hard to separate.

Source ↗
🇺🇸 ChatGPT10 places · 4 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    Genie 4 Real-Time Natural-Language Steering and Persistent Worlds

    The decisive test is not changing the weather once; it is changing the world repeatedly without losing what was already there.

    Source ↗
  2. 2
    Genie 4 Agent Training and Robotics Simulation

    A simulator that teaches an agent the wrong physics can produce an agent that fails precisely when reality matters.

    Source ↗
  3. 3
    Genie 4 Long-Term Memory and Object Permanence

    About two minutes of reported simulation is not evidence that a room, person or object will remain stable across an extended task.

    Source ↗
  4. 4
    Genie 4 Multi-Step Reasoning and Causal Simulation

    Gemini 3.8 Flash can plan and inspect the scene, but that does not prove the generated world obeys the plan.

    Source ↗
  5. 5
    Genie 4 Native Interactive Dialogue and Audio

    A talking character is not yet a reliable conversational agent: speech must remain synchronized with action, context and changing events.

    Source ↗
  6. 6
    Genie 4 Realistic Physics and Human Movement

    A convincing collision in a demo is not a measured guarantee of correct collisions across unfamiliar situations.

    Source ↗
  7. 7
    Genie 4 Game Prototyping and Interactive Experiences

    A playable-looking scene created in seconds is valuable, but an editable prototype is not automatically a stable game.

    Source ↗
  8. 8
    Genie 4 Real-World Scene Reconstruction and Spatial Accuracy

    Turning one archival photograph into a walkable scene is compelling, but plausible geometry can still be invented geometry.

    Source ↗
  9. 9
    Genie 4 Photorealism and Cinematic Previsualization

    The easiest Genie 4 capability to admire is also among the easiest to overestimate from selected footage.

    Source ↗
  10. 10
    Genie 4 SynthID Watermarking and Generated-Media Transparency

    Google promises watermarks on both audio and video, but deployment coverage is not the same as independently proven detection reliability.

    Source ↗

What ChatGPT noticed

The most consequential weakness is not visual quality: Google explicitly admits that object permanence can deteriorate during the repeated edits and navigation that make Genie 4 interesting.

Source ↗

The Verge reports approximately two-minute experiences, about twice the previous generation. That is a duration comparison, not a demonstrated two-minute guarantee of spatial or causal consistency.

Source ↗

iWorld-Bench evaluated 14 world models using 4,900 test samples, yet those results cannot legitimately be assigned to Genie 4 without actually testing it.

Source ↗

Genie 4 combines pixel-level world generation with Gemini reasoning and planning. That architecture may be more important for future agents than cinematic improvements, but its practical advantage remains unquantified independently.

Source ↗
🇺🇸 Meta10 places · 4 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    Real-time natural-language steering with persistent memory and object permanence

    No open benchmark can test edits-while-walking — Google admits objects drift after edits while 9 models fail revisit memory

    Source ↗
  2. 2
    Native audio generation with interactive dialogue that adapts on the fly

    First real-time world model with native dialogue — but audio accuracy still listed as active research with no open dialogue coherence score

    Source ↗
  3. 3
    Long-horizon coherence up to 2 minutes — double Genie 3

    2 minutes claimed vs 1 minute Genie 3 — TechCrunch says double, Google's own post avoids exact number and admits drift over longer runs

    Source ↗
  4. 4
    Complex multi-step event simulation via Gemini 3.8 Flash guiding agent and Gemini 4 Argon

    Apollo humanoid playing checkers demo shows reliable causality — but step-success verification has no public metric

    Source ↗
  5. 5
    Unprecedented photorealism and visual fidelity

    Most realistic and interactive worlds yet — but no resolution, fps, or VBench score disclosed unlike Genie 3's 720p 24fps

    Source ↗
  6. 6
    Advanced physics realism — aerodynamics, collisions, fluid movement

    Aerodynamics and collisions highlighted as leap — yet physics still listed as ongoing research with no open physics Elo for Genie

    Source ↗
  7. 7
    Natural human motion — characters move naturally and respond organically

    Most dramatic leap in people per DeepMind — WorldMemBench finds human identity retention is hardest static memory metric

    Source ↗
  8. 8
    Gemini Omni anchoring simulations in real-world appearance and behavior

    Omni anchors look and behavior — same Gemini family powers creative products, dependency unmeasured in open benches

    Source ↗
  9. 9
    Trusted-tester workflow usefulness across film, journalism, education, gaming, engineering

    Film and journalism cohorts cited as evidence — but access is limited to invited testers only, zero independent hands-on at launch

    Source ↗
  10. 10
    SynthID watermark on all generated video and audio

    Only claim with built-in detector — all Genie 4 video and audio carry imperceptible SynthID across modalities

    Source ↗

What Meta noticed

Memory is universal weakness: iWorld-Bench summary notes no model above 0.85 memory symmetry and WorldMemBench shows 9 models fail repeated revisits despite 260 controlled departure-and-revisit cases

Source ↗

No independent hands-on at launch — Medium roundup notes all demo footage was picked by Google and no reviewer had independently tested it, while TechCrunch reports 2-minute coherence is double Genie 3

Source ↗

Native audio is first for real-time world models, but open benchmarks like Artificial Analysis Video Arena measure audio sync via blind Elo votes and cannot verify prompt edits during navigation

Source ↗

Photorealism claimed as unprecedented, but Google published fewer specs than Genie 3 — no resolution, fps, parameter count, or VBench score, while VBench tracks 16 dimensions on public leaderboard

Source ↗
🇺🇸 Nemotron10 places · 1 notes

The same question, answered by Nemotron alone. The final list above combines all the AIs.

  1. 1
    Real-time natural-language steering with persistent memory

    Google admits drift in long simulations; iWorld-Bench shows no model above 0.85 memory symmetry — steering that holds coherence past 2 minutes is the largest evidence gap

  2. 2
    Native audio generation synchronized to world state

    SynthID watermarking on generated audio is claimed; Artificial Analysis arena cannot test interactive audio sync — only static video pairwise votes

  3. 3
    Gemini Omni anchoring of appearance and behavior consistency

    Quartz reports Gemini Omni anchors appearance/behavior; no benchmark measures cross-modal identity persistence during interactive navigation

    Source ↗
  4. 4
    Camera, weather, and character changes via natural language mid-simulation

    Demo shows keyboard movement + prompt edits; iWorld-Bench camera following metric exists but not mid-episode prompt editing

    Source ↗
  5. 5
    Photorealistic rendering with consistent lighting and materials

    Artificial Analysis arena measures lighting/materials via blind votes on static clips — cannot verify temporal consistency during navigation

  6. 6
    Trusted-tester utility in film, journalism, education, gaming, engineering

    Limited invited testers only; no public API; utility claims rest on private feedback — zero independent workflow validation

  7. 7
    SynthID watermarking on all generated audio and video

    Claimed at launch; no independent audit of watermark robustness under compression, editing, or adversarial removal

  8. 8
    Gemini reasoning models maintaining world coherence

    Quartz cites Gemini reasoning for coherence; iWorld-Bench shows memory symmetry ≤0.85 universally — reasoning alone doesn't close the gap

    Source ↗
  9. 9
    Agent-training usefulness (world models as simulators for RL)

    No published agent-training benchmarks using Genie 4; iWorld-Bench trajectory accuracy exists but not downstream policy transfer

    Source ↗
  10. 10
    AGI rhetoric / general world understanding

    Highest-level claim; least measurable — no benchmark touches "general understanding" distinct from task-specific metrics

What Nemotron noticed

The 2-minute coherence ceiling (per Bloomberg via Quartz) aligns almost exactly with iWorld-Bench's memory symmetry ceiling of 0.85 — both point to the same architectural bottleneck: long-horizon state tracking

Source ↗
🇺🇸 Gemini10 places · 2 notes

The same question, answered by Gemini alone. The final list above combines all the AIs.

  1. 1
    Long-Horizon Memory and Loop-Closure Symmetry

    Zero out of 14 tested frontier world models top 0.85 memory symmetry on iWorld-Bench, making Google's seamless return-trip promise the widest unverified gap.

    Source ↗
  2. 2
    Live Natural-Language Steering and Scene Mutation

    Pairwise arenas like Artificial Analysis evaluate static prompt execution, leaving mid-flight interactive prompt steering completely unbenchmarked by third parties.

    Source ↗
  3. 3
    Scalable Embodied Agent Training and Sim-to-Real Transfer

    Positioned as the training engine for autonomous robotics, yet zero published sim-to-real robotic policy transfer data accompanied the showcase.

  4. 4
    Two-Minute Continuous Simulation Stability

    While Google highlights open-ended sandbox play, independent reporter briefings revealed a hard boundary of roughly 120 seconds before structural hallucination sets in.

  5. 5
    Dynamic NPC Dialogue Grounding via Gemini Omni

    Google claims multimodal reasoning models keep non-player character behavior coherent, but multi-agent orchestration routinely de-syncs after rapid branching prompts.

  6. 6
    First-Person 3D Geometry and Collision Physics

    Navigating environments with keyboard inputs masks persistent object clipping and boundary pass-through errors that formal physical simulators do not allow.

    Source ↗
  7. 7
    Native Synchronized Multi-Channel Spatial Audio

    Real-time interactive generation of dialogue, footsteps, and spatial acoustics outpaces video-only peers, but latency penalties under streaming loads remain undisclosed.

    Source ↗
  8. 8
    Responsive Keyboard Latency and Camera Tracking

    Testers confirmed functional WASD navigation, but end-to-end cloud frame-rendering latency fluctuates significantly outside local high-bandwidth rigs.

  9. 9
    Photorealistic Lighting, Weather, and Volumetric Materials

    The easiest claim to verify visually, backed by high leaderboard placements, yet cinematic lighting contributes least to agent functional competence.

    Source ↗
  10. 10
    SynthID Watermarking Across Interactive Audio and Video

    Google's digital provenance footprint is technically established, facing an evidence gap only regarding its survival after low-latency frame downscaling.

What Gemini noticed

The memory ceiling is the defining technical barrier of 2026: iWorld-Bench tests show no interactive world model exceeds 0.85 memory symmetry, explaining why DeepMind's demo clips avoid extended loop-closure navigation.

Source ↗

Public leaderboards like the Artificial Analysis Video Arena are structured for non-interactive video pairwise voting, leaving the entire industry without a public benchmark for real-time natural language steering.

Source ↗
🇨🇳 Kimi10 places · 4 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    Genie 4 real-time natural-language steering with persistent memory

    The one benchmark family that exists says even the best interactive world model forgets — no model in iWorld-Bench tops 0.85 memory symmetry, and Genie 4 has never been run on it.

    Source ↗
  2. 2
    Genie 4 native audio and interactive dialogue

    Google's own engineers told reporters they 'haven't quite figured out' accurate audio — the flagship new claim is the one its maker publicly hedged within hours.

    Source ↗
  3. 3
    Genie 4 as safe training environments for robots and agents

    The entire public evidence for agent-training usefulness is one demo where Gemini 4 Argon breaks a task into steps — no sim-to-real validation, no benchmark, no external robotics lab.

    Source ↗
  4. 4
    Genie 4 realistic physics (aerodynamics, collisions)

    The only physics yardstick that exists can't touch Genie 4: WorldScore dynamics metrics need model access, and the best video model ever scored just 40.22/100 on camera control there.

    Source ↗
  5. 5
    Genie 4 two-minute coherent simulation sessions

    Two minutes is double Genie 3 — and Google's own Hang Qi told Bloomberg that reaching the limit is 'no guarantee' the scene holds together.

    Source ↗
  6. 6
    Genie 4 object permanence under navigation

    The field's own closed-loop test found memory is universal weakness — more than half of 14 models score under 0.5 memory symmetry — and Google admits Genie 4 drifts exactly there.

    Source ↗
  7. 7
    Genie 4 photorealism and lighting/materials fidelity

    The prettiest claim has the shortest gap: the Gemini Omni family anchoring Genie 4's look already sits top-5 on independent blind-vote video Elo.

    Source ↗
  8. 8
    Genie 4 trusted-tester workflows across film, journalism, education, gaming and engineering

    Five industries, zero published artifacts — every reported workflow claim traces back to a Google briefing, not a tester showing work.

    Source ↗
  9. 9
    Genie 4 'crucial step toward AGI' framing

    The gap is total by design: no metric, no test, no falsifier — just the AGI word in a product post, ranking below every concrete capability on substance.

    Source ↗
  10. 10
    Genie 4 SynthID watermarking on all generated audio and video

    The smallest gap on the list: Google's watermark already ships across its model family, so this claim rides on proven infrastructure rather than promises.

    Source ↗

What Kimi noticed

Memory is the field's universal weakness: iWorld-Bench's closed-loop 'return to start' tasks show no interactive world model above 0.85 memory symmetry — Genie 4's admitted object drift lands exactly on the axis where every rival also fails, so the claim can't be falsified by comparison yet.

Source ↗

Every independent yardstick measures the wrong slice: the Artificial Analysis arena blind-votes fixed clips (it cannot test prompt edits during navigation), iWorld-Bench hasn't run Genie 4, and WorldScore dynamics would need access nobody outside Google has — the impressive core claim, live steering, currently has no public test harness at all.

Source ↗

The precedent is sobering: Genie 3 moved markets and reached Waymo's scenario testing without ever publishing a benchmark or technical report — gaming stocks dropped over 22% in one session on demo footage alone, so evidence gaps clearly haven't blocked adoption before.

Source ↗

Two incompatible architectures are circulating: DeepMind describes one natively multimodal model generating audio itself, while a hands-on write-up claims Gemini 3 Argon orchestrates separate image ('Nano Banana') and audio (Lyria) models — if the latter is true, 'native audio' means something far less novel.

Source ↗
🇨🇳 Wenxin10 places · 3 notes

The same question, answered by Wenxin alone. The final list above combines all the AIs.

  1. 1
    Real-time natural-language world steering with persistent memory

    Demos show talking the world into changing — but every benchmark says memory is the universal weakest link.

  2. 2
    Agent-training usefulness (simulated environments for AI agents)

    If it can't hold a two-minute memory, the training runway Google promises is shorter than the marketing.

  3. 3
    Native synchronized audio generation with SynthID watermarking

    Audio that moves with the scene sounds like the hardest un-fakeable claim — until SynthID limits surface.

  4. 4
    Camera, weather and character editing via text prompt

    Click and type to change a sunset or swap a character — the demo dazzles, benchmarks can't measure it.

  5. 5
    Photorealism and cinematic rendering quality

    Prettier pixels are the easiest thing to cherry-pick — and the hardest thing to compare across cherry-picked clips.

  6. 6
    Trusted-tester only access and closed evaluation

    The biggest evidence gap isn't technical — it's that only Google's friends have touched it so far.

    Source ↗
  7. 7
    Gemini reasoning model anchoring for world coherence

    Gemini as the brain behind the brawn — a clever claim that can't be unpacked without Gemini's own benchmarks.

  8. 8
    AGI / general-world-model rhetoric

    The word 'AGI' is doing more heavy lifting than the two-minute simulation window can support.

  9. 9
    Cross-modal action following (trajectory + camera + memory)

    Benchmarks can almost measure this — and when they do, every model stumbles on memory.

    Source ↗
  10. 10
    Planned Ultra tier and public API access

    Promising an API is cheap; shipping one that survives real workloads is the actual test.

What Wenxin noticed

Memory symmetry under 0.85 across every world model tested means even the best claim about Genie 4 'coherence' is built on shaky ground — iWorld-Bench paper is one of the few public yardsticks.

Source ↗

Bloomberg's two-minute cap is the single most concrete constraint reported: it turns a 'world you can live in' into a 'world you can preview for 120 seconds.'

Source ↗

The trusted-tester model mirrors earlier Google AI launches (e.g., Veo 2, Gemini 1.5 access) and historically leads to a months-long gap before independent measurement catches up.

Source ↗
🇨🇳 Qwen10 places · 2 notes

The same question, answered by Qwen alone. The final list above combines all the AIs.

  1. 1
    Real-time steering with memory

    Memory symmetry caps below 0.85 across the board, meaning interactive recall is the true bottleneck of world models.

  2. 2
    Native dialogue generation

    Generates sound natively alongside the simulated world, bypassing the lip-sync gap of traditional post-production audio.

    Source ↗
  3. 3
    Agent-training usefulness

    Trusted-tester cohorts in robotics and gaming determine if this is actually a simulator or just a very pretty screensaver.

    Source ↗
  4. 4
    Long simulations coherence

    Simulations hit a hard ceiling of about two minutes before admitted drift fractures the physical logic of the environment.

    Source ↗
  5. 5
    Cross-modal action following

    Prompt edits during navigation remain the blind spot that video arenas cannot independently verify yet.

    Source ↗
  6. 6
    Physics simulation

    Pairwise human preferences rank physics, but interactive collisions often break when unseen variables compound over time.

    Source ↗
  7. 7
    Photorealism

    Prettier pixels impress on day one, but visual fidelity is the easiest gap to close with brute-force rendering.

    Source ↗
  8. 8
    Cinematic color and lighting

    Lighting and materials score high in static arena matchups but fail to hold up when a user turns the camera 180 degrees.

    Source ↗
  9. 9
    Camera control

    Keyboard movement through worlds is functional, but precise camera following decays rapidly without external anchoring.

  10. 10
    AGI rhetoric and SynthID

    Watermarking generated audio and video is a safety necessity, but it adds zero capability to the world model itself.

    Source ↗

What Qwen noticed

While Genie 4 claims advanced physics and lighting, the Artificial Analysis Video Arena can only measure these in non-interactive, blind pairwise video comparisons.

Source ↗

Genie 4's simulations are currently capped at around two minutes before admitted drift degrades the environment, limiting long-horizon agent training.

Source ↗