Field Notes [08.2025]

Hey friends,

Welcome back to Field Notes. This is where we track signals at the edges of interactive entertainment and emerging tech.

August saw a bunch of news around the topic of so-called “world models”; we’ll dig into a couple items there. We’ll also revisit another favored subject of ours: voice-enabled games.

If you missed our inaugural issue last month, you can check it out here for a download on new smart glasses from Meta, the brewing battle for spatial mapping, and the rapid increase in AI usage among Steam developers.

Last October, I attended the BITKRAFT Summit in San Francisco.1 As you might imagine, AI was a hot topic among the many investors, founders, and operators in attendance.

I asked a prominent VC working at a massive (> $1B AUM) fund what he thought the endgame was for AI and interactive entertainment.

His response? Fully personalized bespoke interactive experiences generated on-demand. A game for every player — no, every whim of every player.

Look, I don’t know if we’ll ever get there. Investors are not exactly known for having nuanced, conservative takes on the potential applications of emerging technologies. There are challenges a-plenty with a vision like that…

…and yet!

The recent developments surrounding so-called “world models” have me feeling a whole lot more optimistic these days.

Briefly, world models “are neural networks that simulate real-world environments as videos and predict accurate outcomes based on text, image, or video input,” according to NVIDIA.

They are often discussed in the context of robotics or autonomous vehicles, but are increasingly being explored in the interactive entertainment space.

There have been a number of startups making grandiose claims in this area, but the two recent instances we’re looking at today come from major players that can feasibly make those sorts of proclamations with a (relatively) straight face.

The first comes from the Google DeepMind team, which unveiled its Genie 3 world model on August 5th.

Genie 3 can simulate physical properties, natural phenomena, and complex environmental interactions in real-time, all while allowing users to freely navigate with simple WASD / arrow controls.

Genie 3 also allows for promptable world events. Here’s how the DeepMind team explains this feature:

Promptable world events make it possible to change the generated world, like altering weather conditions or introducing new objects and characters, enhancing the experience from navigation controls.

This ability also increases the breadth of counterfactual, or “what if” scenarios, that can be used by agents learning from experience to handle unexpected situations.

As you might have inferred from the quote, Genie 3 is composable with other agents, including Google’s own video game-trained SIMA. This opens up a variety of future research and training applications (more on this in a moment).

Not to be outdone, the second major release in August came from Tencent in the form of its Hunyuan-GameCraft model.

Whereas Genie 3 is a generalizable world model with potential applications for interactive entertainment, GameCraft is specifically tailored to games, having been “trained on a large-scale dataset comprising over one million gameplay recordings across over 100 AAA games.”

The GameCraft website leans much more heavily into fantastical, imaginative environments than does the Genie 3 announcement, with the model’s arXiv paper even including this handy comparison table that explicitly calls out “Game Sources”:

Though it lacks Genie’s dynamic interactivity via promptable world events, GameCraft seems to excel at graphical fidelity and movement precision. Tencent’s model unifies both keyboard and mouse inputs, allowing for greater granularity of movement compared to the limited direct action space of Genie 3.

GameCraft also appears to have better 3D consistency and scene coherence. It doesn’t need to account for unexpected variables in the form of additional user prompts and can therefore preserve the original scene information with a greater degree of fidelity. Genie’s environmental consistency, on the other hand, is an emergent capability that may not be as robust or specifically optimized for fast-paced game-like environments.

In either case, both announcements are seriously impressive.

Yes, they still have a long ways to go. First, world models don’t yet maintain consistency very long. What qualifies as “impressive” in this context is little more than a few minutes of environmental coherence and visual memory.

These models are also fairly rudimentary in terms of feedback beyond simple navigation. Collisions, interactivity with objects, and so forth are still a work in progress.

Nevertheless, the usual caveat applies: this is the worst that the technology will ever be.

So, what sort of gaming applications are we actually talking about?

For starters, there seem to be plenty of applications for prototyping and visually concepting game worlds. This goes for both developers and user-facing UGC applications.

When we begin to layer in the aforementioned agentic composability, additional use cases emerge:

• Gameplay simulation (e.g. “what are the common paths players take through this level?”; “where do players tend to congregate in this environment?”).

• New feature testing at scale (e.g. “how would the world change if all players could do X?”).

• Design copilots (e.g. world models embedded in game engines that aid in balancing systems, tweaking mechanics, etc.)

• Automated QA testing of environments.

Essentially, anything that might benefit from live player feedback can be put to the test before ever reaching an actual human.

That’s not to suggest that world models enable the wholesale replacement of those use cases, but rather that those processes — playtests, smoke tests, focus groups, and so forth — could be applied at a lower frequency for greater impact and at potentially reduced costs.

If you’re interested in learning more about world models and Genie 3, specifically, I recommend checking out the recent a16z Podcast episode featuring two of the researchers from the DeepMind team. It’s a good listen that gets reasonably deep in the weeds on this topic.

Speech is a fundamental interaction in our day to day lives, but [its] use in games has been very limited, due to fundamental limitations of technology.

Source: Astrobeam

Voice-enabled games remain a vastly underexplored area of interactive entertainment.

Despite recent improvements in LLM-driven voice recognition capabilities, few developers have attempted to meaningfully integrate human speech as an input into their games.

Of course, this will come as no surprise to avid readers of Always Scheming. We’ve been advocating for a more meaningful exploration of this space for the better part of two years now.2

That’s why we were stoked to learn about Stellar Cafe, a forthcoming title from indie developer Astrobeam.

Stellar Cafe checks a lot of boxes for us:

• Voice-enabled games? ✅

• AI-driven games? ✅

• Spatial gaming? ✅

All the fun of the future in one title. Plus, robots!

Astrobeam was founded by Devin Reimer, the cofounder and former CEO/CTO of Owlchemy Labs. You might know Owlchemy as the creators of VR hit Job Simulator, a title estimated to have made nearly $15M in revenue on Steam alone, according to VG Insights.

Originally founded in 2010, Owlchemy would go on to be acquired by Google in 2017. Reimer stayed on post-acquisition to run the company for another five years before leaving to start Astrobeam.

All told, Reimer has been building VR games for more than 15 years — a figure that few games industry professionals can match. He’s been able to leverage those bona fides alongside his ambitious vision for the future of games into a $3M seed round for Astrobeam back in 2023. The company’s first title (the aforementioned Stellar Cafe) is due out this fall.

We’ll have to wait and see how financially successful Stellar Cafe can be, but from our perspective it has already had a meaningful impact in tackling the challenging technical hurdles at the intersection of AI, voice, and gaming.

Reimer discussed some of the obstacles in a recent GamesBeat piece:

“[We had to] build out pretty complex server-side tech to handle all the steps of converting text into speech running at really quality, high speed, LLMs, and then converting it back in a way that we just couldn’t do on devices themselves.”

“That’s too much for today’s PCs to do. And for mobile, and then for VR, that’s just untenable,” he said.

He goes on to articulate the broader challenges around developing games with AI-enabled voice recognition:

“In the real-time conversation side, we need these responses in hundreds of milliseconds to make sure that we have something that feels like a natural conversation. So I’m confident it will get there at some point,” Reimer said.

“But I also think there’s going to be this push and pull where, as the technology gets better on the device side of things, the quality of LLMs we want to run are going to go up. And so it’s going to be like chasing these things for a while. And so that’s why we became pretty convinced that, like building out a good server-side architecture was going to be the right call for some time.”

Clearly, combining several emerging technologies into a single game is no easy feat, but we’re rooting for Astrobeam. We’ll keep an eye on the game’s performance when it launches and report back.

Fortunately, other games are already trying pave the way for voice-enabled mechanics to make a comeback. The most recent example in this category is a game called Mage Arena: a PvP wizard battler where spells are voice activated.

We first heard about this game from the excellent GameDiscoverCo newsletter which took note of Mage Arena’s impressive debut atop the Steam new releases charts for July 2025.3

Unfortunately, the open nature of the game’s voice comms has exposed it to all manner of hate speech. It’s possible that AI voice moderation could be a potential solution here, though the cost would likely be prohibitive for a tiny development team like this one.

Nevertheless, we wanted to flag Mage Arena as another case study for developers to consider when integrating voice into their games. Here, as with previous titles like Don’t Scream, the mechanic gives the game a viral, streamer-friendly appeal.

In fairness, this sort of implementation has the potential to come off as gimmicky over the long run. We view it as unlikely that Mage Arena will have any meaningful staying power beyond its brief moment in the sun (though we hope to be proven wrong!).

Yet, there is something to be said for finding traction and quick wins. A Mage Arena-sized hit, temporary though it may be, could very well fund additional projects, extend a studio’s runway for further research and technical development (such as that displayed by the Astrobeam team), or even spawn a new franchise that can be built upon.

What’s clear is that there is appetite for voice-based mechanics. Now we just need someone to wrap them into a compelling, sustainable experience.

This month, we’re tackling the final point discussed above. Let us know your thoughts below and in the comments!

That’s it for August! Enjoy your Labor Day weekend and thank you, as always, for reading.

Until next time.