In this episode of Denoised, we test Google DeepMind's Genie 4 as a way to turn one location photo into a world you can walk through, then look at what AI is doing to voices, crews and game jams. We break down Block's Dessa, a voice model that sounds natural because it was trained only on audio. We also argue over IATSE's new AI-focused officer and the backlash that led Epic Games to change its AI MegaJam rules.
Jump to:
00:31 Genie 4's two-minute interactive worlds
08:01 Turning a Hearst Castle photo into a walkable space
14:12 Block's Dessa voice model
22:38 IATSE's AI officer and union jobs
31:43 Epic's AI MegaJam backlash
Quick Take
Each development in this episode makes something faster to produce, from a walkable location built from one photo to natural-sounding AI speech to a playable game built in four days alongside AI agents. The open question is who sets the terms for using them, from a world model that forgets what it built to a union deciding what to tell its members.
What We Tested: Genie 4's Two-Minute Worlds You Can Steer With Text Prompts
We spun up a Western saloon live in Google DeepMind's Genie 4 world model, which is still a labs release in early beta. Like Genie 3, it takes an environment prompt and an optional character prompt.
When we tested Genie 3, sessions lasted 60 seconds and held onto details like footprints and tire marks. Genie 4 adds runtime and control.
What changed from Genie 3:
- Twice the runtime. Sessions now run two minutes, up from one.
- Live event prompts. Alongside suggested actions, you can type an event. Our "someone falls through the ceiling" prompt dropped a figure through the saloon roof.
- First-person and third-person modes. In third person you play a character you can also generate.
- Image starts and recordings. You can seed a world with a photo, and each session records a downloadable video.
- An API. Google is rolling out API access for the first time.
Detail took a visible step up. It generates at 720p and chair details read clearly, though Addy noted the crowd looked templated.
Memory is still the weak point. Our cowboy whale prompt put a humpback whale in the saloon, but not as the playable character. When we turned around, the whale and the whole bar were gone.
At a '90s-style Hong Kong pier from a suggested prompt, boats and a person we passed had vanished when we looked back.
Addy's take: memory usage climbs over a session because the model tracks every element and its position, which he suspects keeps texture detail low. He also sees the prompt bar as an opening for an agent to steer the experience and act as the world's logic.
What We Explored: A Single Hearst Castle Photo as a Walkable Location Scout
The more useful test for filmmakers started from a real place. We fed Genie 4 a photo from Hearst Castle and walked through the space it generated.
Areas outside the original frame, like the entryway, looked plausible, and a table held its layout when the camera came back around. A suggested prompt sent books floating off a bookshelf, and we could grab and push a ladder.
Where it broke: one room we walked into had two doors, so we tried to loop through one and out the other. The second door never appeared, and when we walked back out, the space had changed completely.
A Culver Theater photo held together at first, then dissolved into generic streets, and the theater was gone when we came back. Storefront windows reflected a figure holding a phone on a stick, which Joey called a virtual version of himself.
Addy pushed back on dismissing the inconsistencies by comparing the output to a conventional build:
Like you're gonna have to, model every single bookshelf, put a book on every single bookshelf, get the texture pattern of the floor. Like you have, obviously you're gonna get a better result, but the amount of work needed for you to build this, for a 5 to 10-second scene is like a month.
The production use: Joey sees quick previs iteration on locations, and Addy framed it as two minutes of location scouting with no real-world consequences. Because Genie records every session, a walkthrough you like could become source material for a splat.
On why Google is building it, Joey's read is world understanding that could feed robotics and self-driving. Addy speculated it could tie into Google Maps imagery.
What We Questioned: Whether Block's Audio-Trained Dessa Should Sound This Human
Block, which runs Square, previewed Dessa, a voice model that sounds realistic because, as Joey described it, it was trained entirely on audio rather than through the multimodal approach behind many other voice models. Its samples include natural breaths, laughs and pauses, and the model is not open yet.
Addy explained what full duplex means:
- Full duplex is truly two-way, with latency low enough that people can talk over each other, like an analog phone call in the '90s.
- Turn-taking is what most digital calls feel like, which is why Zoom calls end in "oh no, you go first."
- Current chatbots take a second or two to process through a large language model before responding.
Joey's argument: if an AI agent gets you faster service from a business, it's worth asking whether you want it to sound human or want to know you're talking to an AI. HAL and Jarvis have computer voices that make it clear you're talking to a machine.
Addy's counter: he expects agents to get very natural, then for society to pull back toward predictable, synthetic voices to keep a clear line between people and agents. A callback from his internet provider sounded natural but kept talking without leaving him room to respond, and he couldn't tell if it was AI.
We both flagged the scam risk, from telemarketing to voice impersonation aimed at elderly people with limited hearing.
Where it fits production: Addy sees a dubbing use, running AI characters' transcript-based dialogue through a model like Dessa so it sounds more natural. Joey added lip-sync remapping as a pairing, and guessed Block's own interest is AI agents answering phones for Square's retail businesses.
What We Debated: IATSE's New AI Officer and Who Gets Reskilled
IATSE named Jillian Arnold its first labor and technology intelligence officer, a role aimed at what AI means for members, including reskilling and education. Arnold comes from cinematography and has been involved with the union for years.
Joey's argument: the work is translating skills members already have. A cinematographer's lighting sense can direct lighting in image models like Nano Banana Pro, and a production designer's eye carries into styling AI images with reference images.
Addy's counter: Addy said that "for the person to be reskilled and be adoptable in this new era, they have to be willing participants." He argued IATSE needs a culture shift before any upskilling, and Arnold may be the person to influence it.
Arnold's analog workflow point. In an interview with The Hollywood Reporter, Arnold said "our members can approach their employer before employment begins to have a conversation about the use of artificial intelligence." She also said some analog workflows are still considered more economical and can produce a unique copyrightable result.
Joey read that as a nuanced point, since some tasks are faster when someone who knows Blender or Unreal Engine builds them directly, and hybrid VFX pipelines can still need green or blue screen.
Addy asked whether the union's goal is keeping everyone employed or keeping the industry profitable, which are not the same thing. He worries blockers push productions toward non-union hires, and asked what a union does if AI cuts a seven-day shoot to three.
The optimistic version: Joey noted production is leaving LA for economic reasons unrelated to AI, and lower budgets could mean more productions and a best boy moving up to gaffer on another show. Addy added that cheaper risk means more creative projects, with five Obsessions and five Backrooms in place of one Mission: Impossible.
What We Debated: Epic's AI MegaJam Backlash and Why Game Developers Push Back Harder
Per the details Joey read, Epic Games' AI MegaJam lists $50,000, four days, and teams of up to five humans plus as many AI agents as they want, building a playable game in Unreal Engine or Unreal Editor for Fortnite for cash prizes. Game World Observer reported that the jam faced criticism and Epic then changed the rules, which as Joey understood it made AI optional after it had been required.
Addy, an Epic Games fellow, recalled asking about AI in a private Unreal Fest meeting years earlier and getting funny looks. He sees the jam as part of a sharp pivot after Epic's layoffs, which he put at 1,000 people, and a push toward user-generated Fortnite content in the mold of Roblox.
Addy's take: AI should always have been optional, and he expects the team that uses it to make the better game. Joey thinks the rule change is the better outcome because he wants to see what the winning team used.
Why games push back harder than film: Addy pointed to a longer industry decline, with layoffs at EA and Xbox and Netflix shutting down gaming divisions. After years of that, being told AI will do your job lands badly.
Joey countered from the player side, arguing that shorter waits between huge releases like Grand Theft Auto would appeal to consumers. Addy agreed on a sweet spot, while noting a AAA game or tier-one film is a company spun up from scratch, with financing, payroll and project management that AI speeds up in pieces but doesn't change.
Bottom Line: The Tools Are Arriving Faster Than the Rules for Using Them
Each story this episode pairs a speed gain with an unsettled question about who decides how it gets used.
- Genie 4. One photo becomes a two-minute walkable space for previs or scouting, but the model forgets rooms and landmarks, so it suits exploration over continuity.
- Dessa. Audio-only training brings AI voices closer to natural conversation, which could help dubbing and raises the question of whether callers should always know they're talking to an AI.
- IATSE's AI officer. The union now has someone tracking AI, and the debate is whether members embrace reskilling and how to protect hybrid workflows that still work.
- Epic's AI MegaJam. Epic's rule change after backlash, which we understood to make AI optional, lets teams pick their own tools.
The first concrete answer arrives when the MegaJam crowns a winner and we see which tools that team picked.
Links from This Episode
Tools & Platforms
Projects
News & Analysis
WATCH THE FULL EPISODE ON YOUTUBE →
