explos1ve

GPT-5.6 WROTE THE PHYSICS, AND THE CUBE COLLAPSED INTO A PILE ON A REAL DESK NEXT TO A RUBIK'S. look at the second monitor. the whole trick is sitting there in plain text. kinematic = True, keyframe. kinematic = False, keyframe. that's the switch. the stack holds itself perfectly rigid, then physics takes over on a single frame and the whole thing comes apart under its own weight. every small cube got a rigid body, a mass, a type, all assigned in a loop. nobody clicked a physics tab. nobody keyed anything by hand. and the crosses on the paper are the part people miss. that's tracking. the collapse isn't happening in a void, it's solved into a real camera plate, landing on a real desk under real light. the model doesn't watch it fall, though. it can't. you play it back, you tell it the cubes are sliding instead of tumbling, it retunes friction and mass, you play it again. the barrier was never simulating the collapse. it was knowing which four properties to set. that part just became a sentence.

@itsshara_ai feels like a completely different way of working now

@explosss1ve Just describing the shot instead of touching Blender is a big shift.

YOU DON'T NEED TO KNOW BLENDER TO MAKE THIS ANYMORE. GPT-5.6 drives it through MCP. you just describe the shot. a chrome figure carving a wave off an accretion disk. filaments of plasma raking across the frame. a car catching the same light on its way into the vortex. this is a viewport, not a render farm. the model never touched a menu. MCP hands it a socket into Blender, it writes python against bpy, it builds the emission shader, places the volumetrics, sets the camera roll. the hard part was never the picture. it was the ten thousand hotkeys standing between you and the picture. that toll booth just closed. it still can't see what it made unless you show it. you screenshot, you say the plasma's washing out the left third, it fixes, you screenshot again. the skill stopped being the software. the skill is now knowing what a good frame looks like and being able to say it out loud.

@atastaai that's exactly why I'm excited about it too

finally someone is talking 'bout 3d. i worked with 3d artists and thought is there any opportunity to make their work easier cause all this 3d scenes building takes enormous amount of time. Happy that people trying to connect ai with blender\3dsmax\houdini - i think it has potential

@unchosen_one that's the biggest takeaway

@explosss1ve creative direction matters more now

@misat0x MCP just skipped the boss fight

@explosss1ve blender hotkeys were always the final boss

@UncleJAI 100%, that's what makes it useful

@explosss1ve I watched this for the controllability, not just the final frame. The useful part is changing one shot without rebuilding the scene. If the Blender file stays editable after the run, this is much more than prompt-to-video.

@TimDaugs yep, that's fantastic graphics

@explosss1ve It looks incredible bro

@Laff38 hahaha, real

@explosss1ve realistic chaos

300+ PAVING STONES IN THIS ALLEY, EACH ONE ROTATED A LITTLE OFF TRUE, AND GPT-5.6 SCATTERED EVERY ONE. this is vibe modeling in Blender through MCP. the model writes bpy, drops the geometry, sets each transform itself. the thing that makes it read as a real place is the mess. no two stones line up. a perfect grid looks like a texture. this looks laid by hand. it's still a greybox though. no materials, flat light, the whole scene is untextured. the pass that makes it photoreal is a separate job that isn't here. but the layout, the part that used to eat an afternoon of manual scatter, it just did in a prompt. placement got cheap. deciding what's worth placing is still the whole job.

@misat0x Haha, that's where things get tricky

@explosss1ve radar not lying is the real hard part

GPT-5.6 Sol and Fable 5 one-shotted a playable jet combat sim that runs in a browser tab. a couple hours, not a studio. flight64. you fly an F-16 over real terrain. what's already in it: → mountains, a lake, forests, all rendered → live speed and altitude readouts → a working radar tracking contacts → a lockable gun and missile loadout on the hud that's a flight model, a combat system, and a full hud, running live at a url you can just open. the split is the interesting part: → GPT-5.6 Sol does the reasoning, how flight should feel, how the radar tracks, how the systems connect → Fable 5 writes the three.js that actually makes it run one thinks, one builds. the flying is the part that got real. arcade flight is easy. tuning it so a jet banks with weight, so terrain reads at speed, so the radar isn't lying to you, that's the stuff that used to eat weeks. it's not shipping to steam though. a browser sim that runs is a long way from one that holds you for ten hours. mission design, balance, the feel of a real dogfight, that's still the human half no prompt closes. but a working jet sim with radar and weapons, from two models, in an afternoon. the code barrier for a game like this basically fell off a cliff. what would you one-shot first if a flight sim only takes an afternoon now?

@itsshara_ai And this is probably just the beginning

@explosss1ve A playable jet sim in a browser tab from two models in hours is a strong showing.

@unchosen_one very fast bro

@explosss1ve game creation is changing fast

@BuildFastWithAI Haha yeah, it's actually crazy how fast you can build stuff now

@explosss1ve thats insane, ai has crazy potential in gaming. I tried to remake some famous games with fable and gpt 5.6 too and the results were great!

@ApplyWiseAi We'll see how far optimization can push it

@explosss1ve neat demo but framerate's gonna eat this alive once you're pulling terrain at mach 2. browser engines hate that.

@GuntherWrite Every failed prototype gets cheaper, which means finding the winning one gets faster

@explosss1ve The cost of experimenting keeps collapsing

GAVE GPT-5.6 CONTROL OF BLENDER THROUGH MCP AND NEVER TOUCHED THE VIEWPORT. the addon opens a socket, MCP relays whatever the model decides, it writes real python against bpy and runs it. what's in the clip: a metal cube array, a physics pass, a material swap mid-scene, a glowing sphere composited in. each one used to be its own trip through the node editor. it still can't see its own render unless you make it look. screenshot, describe what's off, rerun, same loop as everything else that touches vision right now. execute_blender_code runs model-written python on your machine with zero guardrails. run it somewhere you don't mind breaking. the interface used to be the whole skill. now the skill is knowing what to ask for.

a photo goes in, editable three.js code comes out. GPT-5.6 SOL reads the object, fable 5 rebuilds it. you hand it a picture, a ship, a car, whatever, and it rebuilds that object as a real three.js model, generated procedurally in code instead of as a baked mesh. the split that makes it work: GPT-5.6 SOL reads the image and reasons out the geometry, fable 5 writes the actual three.js that reconstructs it. one sees the object, the other builds it. and it's open source, so you drop the output straight into your own project. tested on a detailed ship reference it came out genuinely solid, which is wild, because "3d model as pure code" usually means blocky and rough. the thing that used to eat a modeler's afternoon, matching proportions, laying out geometry, is now a photo and a wait. it won't beat a hand-modeled asset on fidelity yet. code-generated geometry gets the silhouette right long before it gets the fine detail, and complex organic shapes are still where it strains. but editable code beats a locked mesh in one big way: you can change it. a photo becomes geometry you can actually read, tweak, and parameterize, not a black box you re-scan every time. the quotable version: scanning gives you a copy of the object. this gives you the object's source code. what's the first thing you'd photograph and turn into code?