Every time some new development in AI happens, I'm reminded of Psychonauts creator TIm Schafer's words on the technology from a couple of years ago. Speaking on the potential for AI to replicate something that sounded like Mozart, he said: "I was like—wow, that's a really impressive technical challenge, and it does sound like music, and it does sound like Mozart? But also, who cares, because it's not Mozart."
That same phrasing is echoing in my head in the wake of Google DeepMind's upcoming Genie 3, which boasts that it has the capability to let people play a video just like a videogame. For, er, a few minutes before it starts going wrong. Also it only remembers what you did for one minute.
What if you could not only watch a generated video, but explore it too? 🌐Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt.From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵 pic.twitter.com/P0cwFvf5d2August 5, 2025
Which, in all fairness, is a marked step up from the non-Euclidean nightmares of past experiments, like that "Quake 2" inspired mess Microsoft punted into the public consciousness a while back.
Google boasts that Genie 3 allows live interaction, and that "Environments created remain largely consistent over several minutes, with visual memory extending as far as one minute in the past." You can also prompt the software to add various events, like a guy running in a chicken suit, a jetski, or a dragon.
🔘 Promptable world eventsBeyond navigation, users can insert text prompts to alter the world in real-time - like changing the weather ⛅ or introducing new characters 👤This unlocks a new level of dynamic interaction. pic.twitter.com/vvy9U2KOavAugust 5, 2025
So what are the use cases? It's here that Google's pitch cracks at the edges. The most useful (at a push) suggestion is plonking other AI agents into a simulated Genie 3 world to train them. My understanding is that AI agents need to be trained in highly tuned environments by a human hand to get them consistently solid at doing anything but, hey, I'm not trying to sell Cyberpunk-esque braindances to you.
On the other end of the spectrum, we have Google suggesting you could use these experiences for educational purposes. Imagine being able to see dinosaurs (that may or may not exist) or explore the streets of ancient Greece (that haven't been curated by historical experts in any way).
Or—in perhaps the wildest claim—teaching people about how search and rescues are planned. You know. The operations people do to save lives. Hallucinated by a piece of technology that's historically been shown to make stuff up. Lads, are we really thinking this torment nexus thing through?
I'm not going to sit here and tell you this doesn't look impressive. It does. But, just like Schafer, I'm also left asking myself what the untold millions and incredible energy costs are exactly for, beyond producing what is a neat-looking, extraordinarily costly toy.
It's definitely not for videogames, that's for sure. I don't know if anyone's told ol' Google yet, but we've successfully been making games that remember more than 1 minute of information for decades. And they're probably gonna be cheaper, too.
Unless something huge changes, the cost of wholecloth generation is just prohibitive at-scale. Even that bizarre New York Times article which compared Pac-Man to generative AI noted that, when Replica Studios was running its demo, just 100,000 lines of dialogue was enough to put the startup's operating costs at $1,000 per day. There's a reason they shut down.
Google has more money than we can conceive of, but it still needs to make stuff that returns on its investments eventually. There's a reason why it's not rolling tests of this stuff out to anyone but "a small cohort of academics and creators".
And that's not even counting the fact that, well, videogames are designed by people. People with intent, who want to create something specific. Game design is indescribably complex—any videogame that works is basically a miracle of different disciplines coming together to produce a specific feeling or mechanical loop.
Designers pull their hair out trying to find the exact parameters just to make a parry feel good, do we think the hallucination machine's gonna have that kind of finesse? Really?
Generative AI can only gesture vaguely at specificity. Its actual uses in game development have panned out to be somewhat dull—quickly coded, whiteboxed experiments for certain mechanics, background textures to replace the Lorem Ipsum of old, snazzy procedural generation. Making something people actually wanna play? People need to do that, with technology they have actual fine control over, beyond the ability to prompt someone in a chicken suit to run across the screen.
Or, as Larian CEO Swen Vincke puts it, even if the tech is plentiful, available, and good, then that just means everyone has it, and "you're still gonna wanna create something special on top of that".
In a hypothetical future where Google DeepMind's Genie is fully out of the bottle—energy-efficient, cost effective, with a big ol' memory going back for hundreds of hours—then what? Infinite games that are basically just worse? Cool. I'm gonna go play Baldur's Gate 3 again.
