Bella Chai

AI Researcher, engineer, gamer @EntropyGamesAI | Building AI-native games

[Bella Chai @Jinyu_Chai]

The future of AI is not just about answering what humans say, but understanding what they truly mean. This article explores why games may be the most powerful environment for training that capability.

[Luozhu @LuozhuZhang]

https://t.co/X6NQ8XYazG

[Luozhu @LuozhuZhang]

average weekend: hacking on a robot with an on-device speech model

[Bella Chai @Jinyu_Chai]

@qiqileyuan @karpathy Thanks!!!

[Bella Chai @Jinyu_Chai]

After a year of hard work, we’re finally ready to meet you all.

This is the speech system we believe will redefine gaming. It doesn’t just generate speech , it understands your words, your tone, and your emotion, then responds in real time as part of the conversation.

[Luozhu @LuozhuZhang]

I taught a speech model to understand context in conversation. This is what happened

It adjusts voice and tone to express urgency, comfort, understanding from the dialogue. Just like a real human being

520M model. Runs locally on consumer devices

How this is achieved 🧵

[Bella Chai @Jinyu_Chai]

@LuozhuZhang @karpathy Oh, I see now. I don’t really like the whole machine–human relationship either
we’re still way too early for that :D But Samantha is still kinda cute though

[Bella Chai @Jinyu_Chai]

@LuozhuZhang @karpathy What do you mean by “emotional”

[Bella Chai @Jinyu_Chai]

@LuozhuZhang @karpathy Some info

[Bella Chai @Jinyu_Chai]

I've recently been reading some relevant research and experiments, and I've discovered that AI-driven games can keep players engaged for longer. The reason is simple: NPCs are no longer cold, pre-scripted characters; they can surprise players and offer them unique experiences. This process significantly enhances player immersion and replayability.

However, AI-driven NPCs are not without their controversies. The emergence of LLMs has indeed brought unprecedented playability, but the uncertainty and high API costs have deterred many. I think the key issue lies in how to effectively utilize them.

In my opinion, AI is just the exterior; the story is the soul. A truly enjoyable AI game must have a "director" behind it, mastering key story points and driving the player's emotional climaxes. The AI can freely play with details, but the overall story's rhythm, structure, and tension are controlled and designed by a human.

Beyond that, AI brings more than just fresh dialogue. It empowers NPCs with environmental awareness, allowing them to sense the player's presence and actions. It allows NPCs to express more than just dialogue, expressing genuine reactions complete with emotions, voices, breathing, and even laughter. It enables NPCs to build memories, remembering every detail of their relationship with the player. It truly aligns dialogue and actions, and even radically alters NPC logic through a new behavioral decision tree.

I believe this isn't just a simple functional upgrade; it will usher in a new wave of gaming, rekindling the passion of players who once felt esports wasted. AI narrative games aren't just about "chatting"; they offer players unprecedented and unique experiences.

In future games, storytelling will still be the core of the game, but AI will be the lifeblood that gives that soul warmth and transformation.

[Bella Chai @Jinyu_Chai]

@LuozhuZhang https://t.co/Ny7mKuyudA

[Bella Chai @Jinyu_Chai]

@LuozhuZhang Got me thinking about AlphaStar. DeepMind’s work in strategy game environments still feels ahead of its time.

[Bella Chai @Jinyu_Chai]

@LuozhuZhang Local models are valuable in remote environments, in low-latency interactions, and when privacy must be protected. I remember when Starlink first appeared, many people said it had little use, but over time its importance became obvious.

[Bella Chai @Jinyu_Chai]

I believe SLMs and local models represent the future trend, as they are sufficiently affordable to meet basic daily needs while prioritizing user data privacy and security, aligning more closely with real-world usage requirements. Particularly in the gaming industry, while LLMs can grant NPCs more flexible behaviors, the high costs of API services and concerns about data leaks remain glaring issues. In contrast, SLMs not only offer low-cost and privacy-preserving advantages but also provide players with exclusive memory continuity, allowing NPCs to truly become long-term companions uniquely yours.

[Luozhu @LuozhuZhang]

For AI products, people generally think the intelligence dominates everything, while privacy and cost are seen as secondary. The industry’s path shows this: we’ve spent huge money in leading labs to build the largest models with exceptional intelligence.

But I believe we’ve overlooked another path that’s just as important — the case for local models.

Where local models are slowly emerging:

- Cognitive Kernel of the SLM OS. This is the cognitive core of the OS. It doesn’t need to be very large or know everything. it only needs to understand the user’s intent and call the right apps or tools when needed. Ideally, a few billion parameters will be enough.
It’s built directly into the OS with native support for text/audio/vision. So users never need to download or configure anything, and it will automatically return the result in the right format, whether text/audio or vision.

- Super Personal Assistant App. This is the application layer of the SLM OS. It is built as an execution agent that works offline with access to the local device and application data, coordinating and interpreting your actions.
For example, most AI assistants can only reply to an email. This one can pull from unified application data, summarize your meeting notes, and draft a reply the way you want, while leaving the final decision to send up to you.
It can also learn from user feedback, continually improving how it handles tasks. The killer feature is cross-app automation + local brain search. For instance, when you ask “When did I say XXX?” or “Where is the photo of me and XXX?” it can return the correct result in less than 500 milliseconds.

- Game characters in AI-native games. Traditional games rely on scripts and behavior trees to control game characters. After a few tries, everything feels repetitive and predictable, and players end up quitting. With SLMs combined with natural TTS, that logic is completely changed.
Through deep integration of SLMs with the game engine, every NPC can become a unique companion (with their own personality/background/speaking style). More than that, the storyline can follow the choices made by the player and their companions. This is what we call a “never-ending game.”
And these models live on your device, built right into the game files so you hardly notice them. They can remember the adventures you share, the stories you tell, and the things you care about. Over time, they can feel like your best friend.

Local models win on these factors:

- Low interaction latency: local models can respond in < 500 ms, with some native OS operations in < 50 ms. Game character speaking in < 800 ms close to human conversation speed.
- Private data access: the cognitive kernel of the SLM OS can natively access local data, while LLMs never can. Data quality decides everything for an AI product, so it is reasonable to see local SLMs perform better than LLMs.
- On-device finetuning: we may see better fine-tuning techniques that enable test-time training directly on edge devices. This would allow SLMs to improve personalization by learning from user interactions.
- Everyday tasks: most of the things we do each day are relatively simple. So we’d rather get an 85/100 answer in < 500 ms than wait 10 minutes for an LLM to call multiple tools just to give a 95/100 answer.
- Cost: whether it’s an OS or a game NPC, local SLMs can be used infinitely at zero cost, with no need to worry about inference expenses.
- Ownership: not your weights, not your brain.

Yes, LLMs will continue to get smarter, but most of our daily needs remain simple and unchanged. In some key domains, local SLMs can even perform better than LLMs. I believe we’ll see more impressive SLM use cases in the next 3–6 months, and it shouldn’t be a surprise if some of the best products don’t come from the big labs.

[Bella Chai @Jinyu_Chai]

It's encouraging to see that by 2025, LLM development will no longer be a race to increase parameters, but will instead focus on efficient Model of Evidence (MoE), extended context handling, and building flexible ecosystems.

[Ahmad @TheAhmadOsman]

Comparing &amp; Contrasting Recent LLMs Architecture

&gt; DeepSeek-V3/R1
&gt; OLMo 2
&gt; Gemma 3
&gt; Mistral Small 3.1
&gt; Llama 4
&gt; Qwen3 (dense+MoE)
&gt; SmolLM3
&gt; Kimi 2
&gt; GPT-OSS

Are 2025 LLMs really that different from each other?

MoE, MLA, GQA, sliding window, normalization games &amp; more.

[Bella Chai @Jinyu_Chai]

@tymrtn @jandotai @Alibaba_Qwen @huggingface Fair

[Bella Chai @Jinyu_Chai]

I’ve come to feel that LLMs have already stepped beyond the realm of traditional computer science. They no longer resemble the old “input–compute–output” machines we used to imagine. Instead, they look far more like subjects of biology or neuroscience. When we study the brain, we observe which regions light up under different conditions; when we study an LLM, we watch which modules activate under different prompts. It feels less like operating a machine, and more like investigating a living phenomenon.

Many people compare the workings of an LLM to alchemy: we throw data into the architecture and out comes a result—yet what happens inside remains obscure. This black-box character might signal a shift in scientific paradigm: we are learning to live with incomplete explanations, relying on experiment, observation, and heuristics rather than strict step-by-step logic. In that sense, it mirrors the path of the life sciences.

What’s even more striking is that while, on the surface, an LLM is “just predicting the next word,” in practice it may be building higher-order structures along the way: semantic frames, analogies, even hints of causality. This is reminiscent of human thought. When we speak, it may look like we’re simply predicting the next phrase, but beneath that we’re constantly weaving goals and conceptual networks. For humans, the ultimate meta-goal is survival and reproduction. For LLMs, one could argue they are evolving their own “meta-patterns”，maintaining coherence, serving user intent. These emergent properties beyond prediction might be the real frontier for understanding AI.

Perhaps LLMs are reminding us that the borders between computer science, cognitive science, and philosophy are already dissolving. The challenge ahead may not be “explaining a machine,” but rather learning how to understand a new kind of intelligent phenomenon. And when we lean on the metaphors of neuroscience to describe LLMs, we are already，perhaps unconsciously，shifting them from the category of “tools” toward the category of “life.” That, I think, is the question we urgently need to face.

[Bella Chai @Jinyu_Chai]

@GoogleDeepMind @ScienceMagazine Ai changes world!

[Bella Chai @Jinyu_Chai]

@KMkota0 So cool

[Bella Chai @Jinyu_Chai]

In the past, the value of mods was often tied only to the base game. Many mods were more creative than the original titles themselves, yet because the rights were held by developers, creators could only receive minimal returns, often relying on scattered donations. This stands in stark contrast to the true value they generated. Cases like FiveM, Dota, and Counter-Strike have already shown us that the creativity of player communities is far greater than we might imagine.

I have always believed that a healthy model should never be “mods versus developers,” but rather a three-way win: mods + developers + platforms. Roblox is a forward-looking example: it provides creators with a clear and transparent revenue share, enabling creativity to be transformed into sustainable income.

In my view, the barriers to creation are steadily falling. In the future, a mature platform will not only offer creators fair revenue models, but also a wide range of creative tools that make it easy for anyone to express ideas and channel their creativity legitimately. Outstanding mods will receive official recognition, tiered recommendations, and much like YouTube give rise to a new “creator ecosystem,” where being a “mod author” can become a true profession.

I am convinced that in the near future we will witness a new wave of transformation, one no smaller in scale than the rise of social media or online video. Once creators’ potential is unlocked and rewarded fairly, the game ecosystem itself will grow more open, diverse, and vibrant.

[Luozhu @LuozhuZhang]

I’ve realized there’s a deeply unfair but widely accepted norm in gaming: mods are assumed to be free.

For many successful titles, players don’t just buy the base game. They stay because of the thriving mod community. In extreme cases, the mods are more fun than the original. I’ve lost count of Reddit posts saying: “Without this mod, I wouldn’t even touch the game.”

And history proves it:
- Counter-Strike started as a Half-Life mod, then became a global FPS franchise.
- Dota → Dota 2 / League of Legends grew out of a Warcraft III custom map.
- DayZ began as an ARMA 2 survival mod.
- PUBG evolved directly from DayZ/H1Z1’s “Battle Royale” mod.

These “mod magicians” didn’t just make side content. They invented entire genres that now generate billions in annual revenue.

The most shocking example is FiveM.
A GTA V community-built multiplayer mod with custom servers and roleplay mechanics kept millions of players engaged. Rockstar first cracked down (afraid it would cannibalize GTA Online revenue), then in 2023 had no choice but to acquire the FiveM team (Cfx re).

Imagine this, a community mod was powerful enough to make a multi-billion-dollar studio nervous.

Compare this with software: in open source, you can fork, build, and commercialize freely. It’s celebrated. In gaming, IP is closed and zero-sum. Most studios won’t share revenue with modders, even when their success depends on them.

The alternative is obvious. A platform mindset, not just a content mindset.
Provide creators with robust tools, clear rev-share models, and an ecosystem where the next Dota or PUBG can thrive and reward its creators fairly.

The closest example today is Roblox:
- The @TheSims took over two decades to pass $5B in lifetime revenue.
- @Roblox is set to generate ~$4.4–4.5B revenue in 2025 alone, with ~$5.9–6B bookings, and will pay out over $1B to creators.
- Over 40 million games exist on the platform, and almost none are made by Roblox itself.

In the post-AI era, the more open a platform is, and the more aligned it is with its modders/creators, the more likely it is to win.