Home News Inworld AI showcases AI cases studies as they move into production

Inworld AI showcases AI cases studies as they move into production

0
Inworld AI showcases AI cases studies as they move into production

The AI ecosystem was not built with game developers’ needs in mind. Kylan Gibbs of Inworld AI told GamesBeat that while AI technologies are impressive in controlled demos they have critical limitations when it comes to transitioning them into production-ready games.

Currently, AI deployment is slowed down because game developers are dependent upon black-box APIs that have unpredictable pricing and changing terms. This leads to a loss in autonomy and stalled innovations, he said. The players are left with “AI-flavored demos” instead of a sustained, evolving experience.

The Game Developers Conference in 2025. Inworld won’t showcase technology just for the sake of technology. Gibbs said that the company is demonstrating to developers how they have overcome these structural obstacles to ship AI-powered video games that millions of gamers are enjoying today. Their experiences show why so many AI-powered projects fail before launch, and how to overcome them.

We’ve seen a shift over the past few years at GDC. Gibbs explained that the transition is from prototypes and demos to production. “When we first started, it was a proof-of-concept. The use case was pretty narrow. It was mostly characters and non-player character (NPCs), with a lot of emphasis on demos.

Now Gibbs said the company is focused more on production, large scale deployments, and actually solving problems.

AI in production: Putting it to work

Inworld AI works with partners such as Nvidia and Streamlabs to develop AI.

In the past, large language models (LLMs), which were used to create games, were too expensive. It was expensive to send a user query to AI across the internet to a datacenter using GPU time. It would send the answer back so slowly that users could notice the delay.

AI processing has been restructured to move tasks from the server side logic to the client side logic. This has helped reduce costs. This can only be achieved if the user owns a machine with a powerful AI processor/GPU. Gibbs explained that inference tasks could be performed on local machines while more complex machine learning problems might have to be handled in the cloud.

I think that we are at a point where we have proven that the stuff works on a large scale in production and that we have the tools to be able do that. Gibbs said that it was a great transition, as we have focused on the root challenges that the AI ecosystem faces. “When you are in the prototyping demo mind-set, a lot works really well, right?” Many of these tools, like OpenAI and Anthropic, are great for demos, but they don’t work when you scale up to multi-million users.

Gibbs stated that Inworld AI will be focusing on solving larger problems at GDC. Inworld AI shares the real challenges that it has faced and shows what can work in production.

There are some very real obstacles to overcome in order to make this work. We can’t do it alone. Gibbs said that we need to solve this as an ecosystem. “We must stop promoting AI, as a panacea or plug-and-play solution. Gibbs looks forward to the proliferation AI PCs.

We have solved problems with a few partner. Gibbs stated that “if you bring all of the processing on to the local machine, a lot AI becomes more affordable.”

This company provides all the backend models, and is working to keep costs down. I noticed that Mighty Bear Games (headed by Simon Davis) is creating games where AI agents play the game, and humans help craft the best agents.

Companions are super cool. You’ll be able to see multi-agent simulations, such as dynamic crowds. Gibbs explained that if you want to focus on a character-based experience, then you can use either primary characters or background ones. “And getting background characters to function efficiently is really difficult because when people look at Stanford paper, they’re trying to simulate 1,000 agents at once. We all know games aren’t built that way. How do you create a sense of scale with millions of characters, while also implementing a level of detail system to maximize the depth of each agent when you get closer?”

AI skeptics?

AI livestreams

I asked Gibbs what he thought about the stat in the GDC 2025 survey, which showed that more game developers are skeptical about AI in this year’s survey compared to a year ago. The numbers showed 30% had a negative sentiment on AI, compared to 18% the year before. That’s going in the wrong direction.

“I think that we’ve got to this point where everybody realizes that the future of their careers will have AI in it. And we are at a point before where everybody was happy just to follow along with OpenAI’s announcements and whatever their friends were doing on LinkedIn,” Gibbs said.

People were likely turned off after they took tools like image generators with text prompts and these didn’t work so well in prodction. Now, as they move into production, they’re finding that it doesn’t work at scale. And so it takes better tools geared to specific users for developers, Gibbs said.

“We should be skeptical, because there are real challenges that no one is solving. And unless we voice that skepticism and start really pressuring the ecosystem, it’s not going to change,” Gibbs said.

The problems include cloud lock-in and unpredictable costs; performance and reliability issues; and a non-evolving AI. Another problem is controlling AI agents effectively so they don’t go off the rails.

When players are playing in a game like Fortnite, getting a response in milliseconds is critical, Gibbs said. AI in games can be a compelling experience, but making it work with cost efficiency at scale requires solving a lot of problems, Gibbs said.

As for the changes AI is bringing, Gibbs said, “There’s going to be a fundamental architecture change in how we build user-facing AI apps.”

Gibbs said, “What happens is studios are building with tools and then they get a few months from production and they’re like, ‘Holy crap! This doesn’t work. We need to completely change our architecture.’”

That’s what Inworld AI is working on and it will be announced in the future. Gibbs predicts that many AI tools will be quickly outdated within a matter of months. That’s going to make planning difficult. He also predicts that the capacity of third-party cloud providers will break under the strain.

“Will that code actually work when you have four million users funneling through it?,” Gibbs said. “What we’re seeing is a lot of people having to go back and rework their entire code base from Python to C++ as they get closer to production.”

Summary of partner demos (19659027)
Streamlabs’ architecture for bringing AI into workflow.

At GDC, Inworld will be showcasing several key partner demos that highlight how studios of all sizes are successfully implementing AI. These include:

  • Streamlabs: Intelligent Streaming Agent ( ) provides real-time commentary, and production assistance.
  • Wishroll (19459091) : Showing Status, a social networking simulation game with AI-driven personalities.
  • Little Umbrella () : The Last Show is a web-based game with witty AI hosts.
  • Nanobit: Winked is a mobile chat with persistent and evolving relationship building.
  • Virtuoso: Gives developers full control of AI character behavior for a more immersive story-telling experience.

Inworld will also feature two Inworld technology showcases.

  • The On-Device Demo: A cooperative game that runs seamlessly on multiple hardware platforms.
  • Realistic Multiple-Agent Simulation: Multi-agent simulation that demonstrates realistic social behavior and interactions.

The critical barriers preventing AI games from being produced and real dev solutions (19659040)
Kylan Gibbs is cofounder of Inworld AI and a speaker at our recent GamesBeat Next event.

Below are seven of the key challenges that consistently prevent AI-powered games from making the leap from promising prototype to shipped product. Here’s how studios of all sizes used Inworld to break through these barriers and deliver experiences enjoyed by millions.

The real-time wall – Streamlabs Intelligent Agent.

Developer problem:Cloud AI that is not production ready introduces response delays which break player immersion. AI response times can be 800 milliseconds or 1,200 milliseconds due to unoptimized cloud dependency. Even the simplest interactions will feel sluggish.

All AI intelligence is server-side. This creates single points of failure, and prevents true ownership. Yet most developers are stuck in perpetual dependency architectures that lock them into this cloud-API only AI workflow.

Inworld’s solution: Logitech’s Streamlabs Intelligence Streaming Agentacts as a co-host, producer and technical sidekick. It observes and comments on game events in real-time, assists with scene transitions and drives audience engagement, allowing creators to focus on their content instead of being bogged down by production tasks. The Streamlabs team said that they tried to build this using standard cloud APIs but the 1-2 seconds delay made the assistant seem disconnected from the action. “Working together with Inworld, our response times were 200 milliseconds, which made the assistant feel present at the moment.”

The Inworld Framework orchestrates behind the scenes the assistant’s multimodal processing, contextual reasoning, adaptive output, and adaptive output. Inworld’s seamless integration with third-party models, the Streamlabs API and voice commands makes it easy for Inworld to interpret gameplay, chat and voice commands and then deliver real-time action, such as switching scenes or clipping highlight. This approach allows developers to avoid writing custom pipelines every time a new AI model or trigger is introduced.

It’s not just faster, it’s the difference between a virtual assistant that feels alive and one that is always a step behind.

The Last Show (19659048): The success tax
The Last Show

The developer problem:Success should be a cause for celebration, not a financial crisis. Yet, for AI-powered games, linear or even increasing unit costs mean expenses can quickly spiral out of control as user numbers grow. Instead of scaling smoothly, developers are forced to make emergency architecture changes, when they should be doubling down on success.

The Inworld solution: Little Umbrella (19459091) – the studio behind Death by AIwas no exception. The studio was almost bankrupted by the success of the game, which reached 20 million players within two months.

Their technical director s hares that “our cloud API costs went up from $5K to $25K in just two weeks.” “We had to throttle user acquisition–literally turning away players– We partnered with Inworld and restructured our AI architecture

They decided to flip the script for their next game by building with cost predictability, scalability, and scalability from day one. The Last Show is a web-based game where an AI host asks funny questions based on topics that players choose or customize. The AI host will roast the players while they submit their answers, vote on their favorites and eliminate the least popular answer.

The Last Show marks the return of their show, designed from the ground up in order to maintain quality and predictability on a large scale. The result? The result? A business model that thrives on success, rather than being threatened by its success.

The quality-cost paradox

How can you be popular? Status knows.

The developer problem:Better AI quality often correlates with higher costs, forcing developers into an impossible decision: deliver a subpar player experience or face unsustainable costs. AI should enhance gameplay, not become an economic roadblock.

The Inworld solution:Wishroll’s Status (ranking as high as No. 4 in the App Store Lifestyle category) immerses players in a fictional world where they can roleplay as anyone they imagine—whether a world-famous pop star, a fictional character, or even a personified ChatGPT. Their goal is to amass followers, develop relationships with other celebrities, and complete unique milestones.

The concept struck a chord with gamers and by the time the limited access beta launched in October 2024, Status had taken off. TikTok buzz drove over 100,000 downloads with many gamers getting turned away, while the game’s Discord community ballooned from a modest 100 users to 60,000 within a few days. Only two weeks after their public beta launch in February 2025, Statussurpassed a million users.

“We were spending $12 to $15 per daily active user with top-tier models,” said CEO Fai Nur, in a statement. “That’s completely unsustainable. But when we tried cheaper alternatives, our users immediately noticed the quality drop and engagement plummeted.”

Working with Inworld’s ML Optimization services, Wishroll was able to cut AI costs by 90% while improving quality metrics. “We saw how Inworld solved similar problems for other AI games and thought, ‘This is exactly what we need,’” explained Fai. “We could tell Inworld had a lot of experience and knowledge on exactly what our problem was – which was optimizing models and reducing costs.”

“If we had launched with our original architecture, we’d be broke in days,” Fai explained. “Even raising tens of millions wouldn’t have sustained us beyond a month. Now we have a path to profitability.”

The agent control issue: Partnership with Virtuos.

Developer problem:Complex narrative games require sophisticated control of AI agents’ memories, behaviors, and personalities, even if they meet sustainable performance benchmarks. The traditional approaches are either unpredictable or require prohibitively complicated scripting, making the creation of believable characters with consistent personality nearly impossible.

Inworld’s solution:Inworld has partnered with Virtuosois a global powerhouse in game development, known for co-developing triple-A titles like Marvel’s Midnight Suns or Metal Gear Solid Delta Snake Eater. Virtuos, with its deep expertise in character and world building, immediately recognized the need to give developers precise control over AI-driven NPCs’ personalities, behaviors and memories. This allows players to influence the story’s outcome and direction by making choices.

Inworld’s suite of generative AI provides the cognitive core which brings these characters to live while providing developers with full customisation capabilities. Teams can fine-tune AI characters to ensure they stay true to their story arcs and evolve logically within the game world. Virtuos’ focus can be on creating rich, immersive experiences with Inworld’s tools.

At Virtuos we see AI as an opportunity to enhance the artistry and bring to life the visions of game developers,” said Piotr Chzanowski, CTO, in a press release. By integrating AI into their creations, we allow developers to add new dimension to their games, enriching them without compromising on quality. Our partnership with Inworld allows us to create gameplay experiences that were previously impossible.

An upcoming prototype will showcase the best of both teams. Media are invited to the Virtuos booth C1515 to see a private demonstration.

The immersive dialogue challenge Winked

Developer problem:Nanobit’s Winked is an interactive narrative mobile experience that allows players to build relationships with characters through dynamic, evolving conversation, including direct messages. To meet player expectations the AI-driven dialogue for players had to be more personal, emotional nuanced and stylistically unique than what was possible with frontier models. The quality required was beyond the capabilities and costs of standard AI solutions, but the scaleability of the solution was a challenge.

Inworld’s solution:Nanobit used Inworld Cloud to train and distill a custom AI tailored specifically for Winked. This model provided a superior dialogue quality, more organic, personal and contextually aware, than off-the shelf solutions, while keeping costs at a fraction of the traditional cloud APIs. The AI seamlessly integrated into Winked’s core gameplay loops, increasing user engagement and maintaining financial viability.

This AI-driven dialog system not only improves player immersion, but also remembers previous conversations and continues the storyline, giving the player relationships that develop as chats progress. This encourages players to have longer conversations with characters and return more often as they become closer.

The multi-agent Orchestration Challenge: Realistic Multi-Agent Simulation

Developer Problem:To create living, believable environments, it is necessary to coordinate multiple AI agents so that they can interact naturally with one another and the player. Developers struggle with creating social dynamics that are organic and not mechanical, especially when they scale up.

Inworld’s solution:This Realistic Multi-Agent Simulation shows how to orchestrate AI agents into cohesive and living worlds by using Inworld. This simulation creates believable, natural social dynamics by implementing sophisticated agent coordination, contextual awareness, shared environmental knowledge and shared environmental information.

These autonomous agents demonstrate how agent orchestration allows for emergent, lifelike behavior at scale. They may form spontaneous crowds around exciting events in-game, react to shared group emotes or engage in multi-character conversation. This technical demonstration highlights the potential for sustained engagement and deep player immersion by bringing social hubs into life. Multiple characters interact with consistent personality, mutual awareness, collective response patterns, and collective responses that create the feel of a truly alive world.

The hardware fragmentation problem: On-device demo

Developer problem:AI functions optimized for high-end hardware fail on mainstream hardware forcing developers to limit their audience or compromise vision. AI vendors also hide critical capabilities needed for on-device analysis (distilled models), deep fine-tuning, distillation and runtime model adaptation to protect recurring revenue and maintain control.

Inworld’s solution:AI hardware for gaming is not a one-size fits all solution. Ensure consistent performance and accessibility on different devices can easily increase complexity and cost. AI solutions need to adapt seamlessly across different hardware configurations in order to achieve scalability.

The on-device demonstration showcases an AI powered cooperative gameplay that runs seamlessly across three hardware configurations.

  • Nvidia GeForce RX 5090
  • AMD Radeon RX RX 7900XTX
  • Tenstorrent Quickbox

It’s not about theoretical compatibility, it’s about achieving consistency across diverse hardware. This allows developers to target a wide range of gaming devices without

The development difference: Moving beyond prototypes

Most AI game projects fail in the gap between prototypes and production. The plugins that are used for prototyping can break in real-world conditions.

  • While they may be useful for prototypes, they often fail in production.
  • Costs explode:Token pricing creates financial cliffs which make scaling unpredictable.
  • AI performance varies significantly between test and production environments.
  • We’ve been watching incredible AI game demos die during the transition to production. “The pattern is always the same: impressive demo, enthusiastic investment, then the slow realization that the economics and technical architecture don’t support real-world deployment.”

    At Inworld, we’ve worked relentlessly to close this prototype-to-production gap, developing solutions that address the real-world challenges of shipping and scaling AI-powered games–not just showcasing impressive demos. Gibbs said that Inworld is eager to share experiences at GDC that not only make it to launch but also thrive at scale. The company’s stand is located at C1615. Gibbs stated that instead of talking about the future gaming with AI, they will show real systems solving real issues, developed by teams that have faced the same problems you are. Gibbs said that the path from AI prototypes to production can be challenging, but it is achievable with the right partners and a strategy that focuses on delivering AI experiences players will love.

    Jim Keller of Tenstorrent will be on a GDC panel with Inworld CEO Kylan Gibbs to examine the broken economic model of AI in gaming and the practical path forward:

    Jim Keller is a legendary hardware engineering who led important processor projects for companies such as Apple AMD and Intel. Keller will be on the GDC panel alongside Inworld CEO Kylan gibbs to examine AI’s broken economy in gaming, and the practical way forward.

    According to Keller in a press release, “current AI infrastructure is not economically sustainable for games at scale.” “We see studios adopting impressive AI features during development, but then stripping them back before launch when they calculate the true costs of cloud computing at scale,” said Gibbs.

    Gibbs expressed his excitement to talk with Keller on stage, about Tenstorrent which aims at serving AI applications at a scale for less than a 100-fold cost. The session will examine concrete solutions to these barriers.

    The discussion will explore concrete solutions for these economic barriers.

    • Significantly cheaper model and hardware options.
    • Strategies to eliminate API dependency.
    • Practical approaches to hybridization that optimize cost, performance, quality, and ROI.
    • Interactive learning systems that improve ROI with time.

    Using Keller’s deep expertise in hardware from Tenstorrent and Inworld, Attendees can expect to gain candid insights on what matters most when bringing AI into practice. They will also learn how to build an AI pipeline that is cost-effective without sacrificing performance or creativity.

    Session details:

    • Thursday, March 20, 9:30 a.m. – 10:30 a.m.
    • West Hall, Room #2000
    • For more details, Visit the GDC Page

    Session with Microsoft – AI innovation for game experience

    Gibbs, Haiyan Zhang, and Katja Hofmann from Microsoft will also be present to discuss how AI can drive a new wave of dynamic gaming experiences. This panel combines research with practical implementation to address the critical challenges that developers face when moving prototypes into production.

    This session demonstrates how our collaborative approach overcomes industry-wide obstacles preventing AI games reaching players. It focuses on proven patterns which overcome reliability, quality and cost challenges that most games never survive. I asked Gibbs how he could convince a developer that AI was a train that they could get on and that it wasn’t a train that was coming straight at them.

    Unfortunately, we couldn’t share the names of all our partners. Many triple-A’s [are quiet]are in the top tier. It’s happening but it takes a lot of effort. We’re engaging with developers to meet their creative requirements. Gibbs explained that if a developer has a game they plan to launch in the next two years and they do not have a clear vision of how they can do it efficiently or at a low cost, they can work with us on that. “There are fundamentally different ways it can be integrated into games.” We’ll be making a lot of announcements this year, as we try to make them more self-serving. Visit the GDC Page

Stay informed!

GB Daily (19659112) Get the latest news delivered to your inbox every day

Read about our privacy policy

Thank you for subscribing. Click here to view more VB Newsletters.

An error occured.

www.aiobserver.co

NO COMMENTS

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Exit mobile version