Welcome, humans.
So how good is GPT-6 Astra, anyway?
GPT-6 Astra built a black hole simulator, a Blender scene, a physics game, a sci-fi world, a sound-diagnostic prototype, and Cat Doom from one prompt each.
That is the whole experiment in our new podcast episode: Corey and Grant gave Astra six one-shot build tests, barely touched the steering wheel, and watched what came out the other side. The surprising part was not that every result was perfect. It was how much already worked before iteration.
Here’s our favorite parts:
(6:16) Astra bends light around a black hole: Corey launches photons past the event horizon and the model turns a physics concept into an interactive demo.
(20:55) The Blender reveal: Astra builds “The Last Observatory” from scratch, including the planet, astronaut, lights, telescope, and a six-second camera pullback.
(37:18) Corey’s desk becomes a sci-fi cabin: the blinds rise to reveal a planet outside, while the recreated monitors, fan, desk, and collectibles still resemble the original photo.
(50:30) Cat Doom graduates: one tiny prompt produces Hellcat, multiple levels, weapons, a HUD, mobile controls, sound, keys, and actual Doom-style progression.
(56:07) Chairman Meow (Astra named this) fights back: the boss battle gets hard enough to kill Corey, which also exposes one of the episode’s big lessons about AI game balancing.
There is a pattern across all six demos: Astra can build a lot of the thing now. The remaining work is increasingly about taste, restraint, balancing, deciding what to remove, and knowing what is worth building in the first place.
Why watch this? Because this is a much more useful way to understand a frontier coding model than staring at another benchmark chart. You get to see what one prompt can actually produce, where it falls apart, and where a human still makes the difference.
Watch and/or Listen now: YouTube | Spotify | Apple Podcasts
P.S. The shortest prompt in the whole episode was literally “Make the game Doom end to end, but with cats.” Astra somehow took that as permission to invent Chairman Meow. If you want to try Fable’s version of Cat Doom, check it here.
Keep scrolling for a word from our sponsor, a quick Cat Doom time machine, Thursday’s OpenClaw 2.0 livestream, and all the links to the projects we made (including Fable 5.1 versions for comparison!)
FROM OUR PARTNERS
Plenty of companies can launch an AI pilot. Far fewer know how to turn that pilot into something secure, scalable, and useful in everyday work. Explore “The Enterprise Guide to Scalable AI,” sponsored by Dell Technologies and NVIDIA. The hub looks at what changes when AI moves from pilots into production, including how teams prepare data, choose infrastructure, run agents closer to users, and keep AI systems governed as they scale.
Additional Resources: Cat Doom has receipts
One of the funniest ways to see how fast coding agents have improved is to compare the exact same ridiculous idea over time. The episode does that with Cat Doom, and the older versions make the jump obvious.
The early Cat Doom: one of the oldest versions, from roughly nine months ago.
GPT-5.4 Cat Doom: the next step in the progression.
Fable 5.1 Cat Doom: the September 1 version right before Astra.
The progression matters because the prompt barely changed. The model did.
Want to check out the projects yourself?
We tested the same ideas across Corey’s GPT-6 Astra one-shots, Grant’s versions, and even Fable 5.1 variants. Here are the builds side by side:
🐈 Cat Doom: Corey’s Astra version | Grant version | Fable 5.1 version
🕳️ Black hole: Corey’s Astra version | Grant version | Fable 5.1 version
🔭 The Last Observatory: Corey’s Astra version | Grant version | Fable 5.1 version
🪑 Office Chair Space Program: Corey’s Astra version | Grant version | Fable 5.1 version
🔊 Sound check: Corey’s Astra version | Grant version | Fable 5.1 version
🚀 Orbital Study: Corey’s Astra version | Grant version | Fable 5.1 version
And if you just want to play the polished Cat Doom build: CatDoom is live here.
The useful comparison is how much of the finished experience each model could produce from roughly the same starting idea.
The fun part is comparing them side by side. The prompts were similar. The outputs definitely were not.
🔴 THURSDAY LIVE: OpenClaw 2.0 with Chief Architect Vincent Koc
Thursday, September 17 • Hit “Notify me” on YouTube for the scheduled start time.
Click the image, then hit “Notify me” so you don’t miss it.
OpenClaw 2.0 just landed, and we’re going straight to the person helping build it. Vincent Koc, Chief Architect of OpenClaw, is joining us live for a hands-on tour of the biggest OpenClaw update yet.
Vincent will demo the new release and show how OpenClaw is evolving from an open-source personal AI assistant into a broader agentic computing platform.
We’ll cover:
The biggest OpenClaw 2.0 changes and easier setup.
Running agents across local machines and cloud workers.
Interactive widgets, dashboards, persistent work, automations, loops, and approval controls.
Local models, memory, credentials, permissions, security, and agent interoperability.
Experimental multi-agent features like Swarm and where personal AI agents go next.
Because this is live, we’ll also put your questions directly to Vincent while he walks through the release.
🎙️ In Case You Missed It…
A few recent videos and episodes worth catching up on:
1. New to GitHub? Start here.
TL;DW: Cassidy Williams joined us for a beginner-friendly walkthrough of repos, branches, pushing, pulling, merging, cloning, forking, worktrees, GitHub Actions, and connecting AI coding agents to GitHub.
Why you should watch: If AI can build you an app but “push it to GitHub” still sounds like a threat, this is the missing layer.
Watch: YouTube
2. Want the Cat Doom origin story?
TL;DW: Go back to the older Cat Doom builds and compare what the same basic joke looked like months ago versus Astra now.
Why you should watch: Few benchmarks make model progress as obvious as watching increasingly competent AI try to remake Doom with cats.
Watch: YouTube
One more before you go:
The episode’s biggest takeaway is worth stealing for your own AI work: as capability rises, prompting starts looking less like “tell the model how to build” and more like creative direction. Decide what matters, define the taste, kill unnecessary UI, and iterate on the parts a one-shot build gets wrong.
Subscribe to our YouTube Channel for more!
We’re at 23.4K subscribers. Click above to help us keep climbing.
Subscribe on YouTube to help us bring in more builders, researchers, and guests who can teach you something useful about AI every week.
Stay curious,
The Neuron Team
What do you want to learn about AI?Pick your favorite, then share any others in the "additional feedback" |
|
