Yuanlei Niu

Head of AI, Century Games — creator of Whiteout Survival @WOS_Global. Building stealth mode AI innovation in gaming.

[Yuanlei Niu @truenyl]

I built a 2.5D Splatoon-style game with GPT-6 Astra + Blender.

This one turned out way better than I expected.

I used some assets AI found online, then let GPT-6 Astra work inside Blender to keep improving the layout, materials, lighting, textures and overall composition.

The biggest difference compared with my earlier 3D experiments is that Blender gives the whole thing much more visual polish.

And the gameplay is actually in pretty good shape too.

The shooting feels surprisingly solid, there’s a decent variety of weapons, and the AI enemies already have some numerical difficulty differences instead of all behaving the same.

The game is playable now, and I deployed a build on Vercel so people can try it directly.

https://t.co/h3131SieyB

Now I really want to add multiplayer hhhhh.

[Yuanlei Niu @truenyl]

This would be really helpful for searching through our game assets.

[Shridhar @shridharathi]

Fast image search with Jev

[Yuanlei Niu @truenyl]

@HugoDuprez thanks! want to try and remix sth

[Yuanlei Niu @truenyl]

It looks so cool!

[Hugo Duprez @HugoDuprez]

Jev can generate game levels in real time.

Faster and cheaper structured output could be a big deal for game dev!

[Yuanlei Niu @truenyl]

@HugoDuprez it looks so cool. Can you share the whole workflow?

[Yuanlei Niu @truenyl]

I built a skiing game with GPT-6 + Blender, and this is the first time the 3D quality actually started to feel close to a real game.

I referenced a lot of Mario Kart for the visual quality and UI feel, especially the sense of speed, readability, camera, and the amount of polish hidden in small details.

Compared with my earlier Three.js tests, the jump is pretty noticeable.

The workflow that worked best for me was:

1. Start with concept artLock down the visual style and overall mood first.
2. Find assets close to that styleSearch online for models that already match the concept as closely as possible.
3. Bring in open-source animationsUse existing animation assets instead of trying to generate everything from scratch.
4. Let GPT-6 build and polish the scene in BlenderAssemble the world, place assets, tune lighting/materials, and keep iterating.

The multi-turn iteration was probably the most important part.

I kept using Mario Kart as a reference point and asking for small changes:
adjust the character pose
improve the UI hierarchy
change the animation timing
add more visual feedback
tweak the scene composition
add more pose and pose combo

None of these changes is huge by itself.
But after enough iterations, the game starts to feel much more polished.

[Yuanlei Niu @truenyl]

I accidentally played 30–50 levels of a user-made Piupiu puzzle game in one sitting.

That probably says more than any benchmark :)

Since we started building Piupiu, I basically have an endless supply of puzzle games to play.

As someone who genuinely loves puzzles, this is a very good problem to have.

One of my favorite user creations recently is a really polished Chinese-style sorting game built around porcelain.

The art direction is beautiful, the theme feels unique, and most importantly, it’s actually fun.

What I find interesting is that Piupiu is already pretty good at capturing the creator's gameplay intent.

The core mechanic works. The progression feels coherent. The game has personality.

Of course, this is still very different from building a commercially successful puzzle game.

Great puzzle games usually win on things that are much harder to generate:

carefully tuned difficulty curves, level design that keeps players in flow, monetization bundles, live ops, and acquisition economics that actually work.

Piupiu is not there yet.

But on the pure gameplay side, I think we're getting much closer to something meaningful:

a creator can have a specific puzzle idea, and AI can turn that idea into something genuinely enjoyable to play.

For a puzzle fan, that feels pretty magical.

[Yuanlei Niu @truenyl]

I used GPT-6 Astra to build an Escape from Duckov-style game.

Originally I just wanted to test its 3D scene and map-building ability.

Then I kept adding things for completeness: building, inventory, items, shooting...

And somehow it became basically playable.

The interesting part is that I didn't use Unity or another game engine. Most of it was built with Three.js.

For that setup, I think the result is already pretty impressive.

But the gaps are also obvious.

The scene still struggles with lighting and materials. Three.js can get the structure right, but it still doesn't have the same visual quality as Unity.

Shooting technically works, but the game feel is bad.

And that's probably the expensive part.

Recoil, hit feedback, movement, camera, pacing. Getting all of that to actually feel good would take a lot more time.

I also tried letting AI play the game, record the footage, and make the video by itself.

It works roughly, but scene transitions and continuity are still pretty weak.

My current takeaway:

AI is getting very good at turning a complicated idea into something playable and surprisingly complete.

But going from “it works” to “it feels good” is still a very different problem.

[Yuanlei Niu @truenyl]

As a paying Lumosity user, I definitely believe this.😂

[Brandon Luu, MD @BrandonLuuMD]

Frequent video gamers performed on cognitive tests like non-gamers that were ~13.7 years younger

[Yuanlei Niu @truenyl]

looks so cool!

[Meng To @MengTo]

I built a multiplayer Catan-inspired game entirely with Astra. It's free to play:

https://t.co/rYAUkxN7UD

It's mobile-friendly and has narration, guides, customizations, multiple expansions and a game lobby where anyone can join, chat, and use voice chat. It took 4 days to build and hundreds of dollars in tokens.

Becoming a game dev in 2026 wasn't on my bingo card, but here we are. AI has advanced so much that you can finally make the game you've always dreamed of, even without a team.

[Yuanlei Niu @truenyl]

I spent some time testing GPT-6 Astra's 3D capabilities and built a Whiteout Survival-style scene with it.

My first impression: the improvement is pretty significant.

What seems to matter most is three capabilities working together.

Spatial understanding

This is probably the strongest part.

GPT-6 Astra understands layout, relative position, scale, camera, and how objects should fit together in a 3D space surprisingly well.

For this scene, it understood the overall structure and placement of the buildings much better than I expected.

Coding

The coding capability also feels very strong.

A lot of the structures and buildings were not really “modeled” in the traditional sense.

GPT-6 Astra was writing code to assemble basic geometry and use simple shapes to create the visual result.

For things like buildings, roads, walls, vehicles, and other structured objects, this works surprisingly well.

Computer Use

I've seen some very strong examples of GPT-6 Astra using Blender.

Unfortunately, I don't really know Blender myself :)

So I asked one of our 3D experts to look at the workflow with me.

His take was that GPT-6 Astra is clearly getting much better at actually operating 3D tools, understanding the current state, and making the next move.

That said, for low-poly styles like this one, the difference between doing more work directly in Blender versus building things procedurally through code is not always that large.

So far, I think GPT-6 Astra is especially strong at 3D scene building and layout.

Where it still feels quite different from tools like Tripo or Meshy is precise 3D asset generation.

If I want to accurately recreate a specific object, especially a character, dedicated 3D generation models are still much better.

Putting the two together feels very powerful.

[Yuanlei Niu @truenyl]

@ivantkachukx Exactly 😂 honestly, the fact that people kept using the duct tape was what convinced us it was worth rebuilding properly.

[Yuanlei Niu @truenyl]

Our vibe-coded prototype hit its first real scaling wall at around 100 users.

The first version of Piupiu was built by one PM in about a month and ran on a single cloud server.

That was exactly the right setup when we had no idea whether anyone would use it.

Then people actually did.

At around 100 test users, things started getting painful.

The server couldn't comfortably handle the load anymore, and the codebase wasn't designed for multiple engineers to work on it at the same time.

That's when engineering really had to step in.

I think this is an interesting boundary for vibe coding:

It can get you to real users incredibly fast. But turning that prototype into something scalable, reliable, and maintainable is a very different job.

Vibe coding got us to the first users.

Engineering made it possible to keep growing.

[Yuanlei Niu @truenyl]

A user made this Odyssey-themed game on Piupiu and I loved the art style.

So I spent a little time remixing it into my own version :)

This is the part of AI game creation I'm increasingly excited about.

Not everyone needs to start from a blank page.

A theme, a story, a visual style, any of these can become the starting point for your own interactive experience.

Everyone can create and remix interactive entertainment around whatever they love.

[Yuanlei Niu @truenyl]

The more AI safety regulation I read, the more I think the incentives are much messier than they look.

I think the safety risks are real.

Some of the capabilities we're starting to see are genuinely scary, and I do think we need better safety standards.

But I also keep thinking about how much commercial incentive there is behind some of the proposed regulation, especially around the current IPO timing.

A few things stand out to me.

1. Slowing down the race also saves everyone a lot of money.
Frontier AI is basically a prisoner's dilemma right now.
Every lab would probably prefer not to spend tens of billions every few months just to keep up with the next model. But nobody can slow down alone.

2. Many safety requirements naturally favor closed APIs.
Things like continuous monitoring, identity checks, incident response, or even pulling back access to a model all work much better when the model stays behind an API.
Once open weights are released, the original developer no longer controls how the model is deployed.
So if the safety standard assumes that the developer must always stay in control, closed models have a structural advantage.

3. Compliance can easily become a moat.
Compliance and licensing are huge fixed costs.
A smaller lab, especially an open-source lab from other countries, might have a model that is cheaper or even better, but still have a very hard time entering that market.

A regulation can reduce real risks while also making the industry much more concentrated.

That's the part I think we should be careful about.

[Dario Amodei @DarioAmodei]

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.

Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.

You can read the full post here: https://t.co/OGyPb7yaYt

[Yuanlei Niu @truenyl]

A user recently made this roguelike on Piupiu, and it quickly became one of our most popular games.

Roughly 1–2 hours + ~$10 to get to this level of completeness.

That still feels kind of crazy.

But the interesting part is, building the game is becoming the cheap part.

Making it fun isn't.

The model can generate enemies, upgrades, stats, progression, and a working gameplay loop. But getting the balance right still depends heavily on the creator's taste and judgment.

This one got a lot of love from our users, and I've been playing it quite a bit myself :)

My current takeaway: The model can get you to 60. Maybe the real product opportunity is making it easy for creators to use simple feedback and their own taste to push that to 80+.

And maybe that's true for most creative products where “good” isn't objectively defined.

AI is making game creation cheap. Good game design is still scarce.

P.S. humans still really like competing with each other. A leaderboard is probably a huge difference.

[Yuanlei Niu @truenyl]

We shipped v0.2.5 of our beautiful gaming agent today, just call it Piupiu. A very nice agent to build very nice games.

It feels like the end of a pretty long chapter: turning a vibe-coded prototype into something our small team can actually keep building on.

The very first version was built by one extremely talented PM in about a month.

It worked surprisingly well.

People started using it almost every day, inviting teammates, and eventually the whole thing outgrew the single server it was running on.

That was when engineering really got involved.

At first, we tried the obvious thing: review the existing code, clean it up, and keep building.

That turned out to be much harder than we expected.

The prototype had grown through layer after layer of vibe-coded additions. The code worked, but the module boundaries were messy, the architecture wasn't designed for multiple engineers, and even splitting work across a small team became painful.

So we decided not to do one giant rewrite.

Instead, over the last 3–4 months, we've been rebuilding it module by module, release by release, while keeping the product live and continuing to ship what users needed.

v0.2.5 alone took us about 4 weeks.

A lot of the work has been pretty invisible: architecture, session state, agent reliability, tracing, recovery, tool failures, deployment, the workspace itself.

None of these makes for a flashy AI demo.

But today, most of the core rebuild is finally behind us.

For a team this small, that feels like a pretty big milestone.

More importantly, it finally feels like we have a foundation we can build much more ambitious things on.

More to come on how we think about AI gaming, what we’re learning, and the sometimes painful journey from vibe-coded prototypes to real software. Thanks for following along with our small team.

[Yuanlei Niu @truenyl]

Today was a really special day.

After version after version of iteration, our game agent finally crossed a threshold and it feels genuinely usable now, not just promising.

And today we hit 1,000 registered users and 100 DAU for the first time.

We bought a cake and celebrated together as a team. There’s something deeply satisfying about giving moments like this a little ceremony, pausing to recognize how far we’ve come before getting back to building.

Still a long way to go. But today felt like one of those small milestones worth remembering.

If you just keep building, keep iterating, and keep going long enough, eventually the world answers back. 🎂

[Yuanlei Niu @truenyl]

Just finished carefully going through all the Genie 3 demos. While there’s still some way to go before it fully meets game-level experience and requirements, the progress is stunning—especially compared to Genie 2.

Had an interesting chat with a pre-training researcher yesterday: vision understanding and generation are fundamentally asymmetric. Building robust evals for vision is much more challenging than for text. Human cognition itself is more limited for vision than text, and models can quickly surpass us when guided by strong evals.

Genie 3 really feels like a leap towards a sci-fi future. The pace of progress is unreal.

[Demis Hassabis @demishassabis]

Genie 3 is here - it can generate an entire world simulation that you can interact with in real-time, just from a text prompt! It's pretty mind-blowing really when you stop to think about it, and it's rapidly improving - one day we will be able to build the Holodeck for real!

[Yuanlei Niu @truenyl]

Tried GPT’s study mode—super helpful for someone like me who spends a lot of time learning on the job. It’s a big step up from the usual chat: instead of just Q&A, study mode actively gauges your level, asks insightful follow-ups, and adapts depth based on your feedback.

Perfect for quickly ramping up on a new domain or getting a structured, big-picture understanding. When I asked about reasoning vs base models, the answers were clear and to the point, but the real value was the extra prompts—like asking if I cared about training methods, evaluation metrics, or technical evolution. Sometimes, those follow-up questions are even more valuable than the answers themselves—they highlight perspectives I hadn’t considered and help structure my learning.

Compared to standard chat (which feels more like a passive answer bot), study mode delivers a much more interactive and holistic learning experience.

P.S. No wonder the #1 use case is “Help me with my homework…”—it’s genuinely great for it!

[OpenAI @OpenAI]

As ChatGPT becomes a go-to tool for students, we’re committed to ensuring it fosters deeper understanding and learning.

Introducing study mode in ChatGPT — a learning experience that helps you work through problems step-by-step instead of just getting an answer.