Acorama

Vibe coding || curating intelligent architectures. flow as code ||

[Acorama @Acoramaa]

https://t.co/64Cyv2mawk

[Acorama @Acoramaa]

GPT built my Minecraft base. I gave 50 zombies the address

I was too lazy to build one myself. It gave me a two-level workshop above a canyon, glass walls and a retractable bridge

Then came the inspection

First run: bridge open. Zombies inside in about 17 seconds

Second run: 50 fresh zombies. Retract the bridge 10 seconds in. No breach during the 60-second trial

GPT wrote the build commands. The mobs use normal Minecraft AI

I delegated construction. The zombies handled quality control.

[Acorama @Acoramaa]

I asked Opus 5.5 for a 3D robot for my dream workshop. Then gave it its first job

Meet MILO. It picks up a small part, opens its cargo hatch, puts the part inside and brings the arm back

The scene is editable in Blender. I can move the wrist target, open and close the gripper, scrub the pick cycle and separate the robot into its component modules

Getting the motion right took revisions. The carry path needed clearance over the head, and the hatch had to wait until the arm was out of the way

It's a scripted pick-and-place sequence. Forces, friction and autonomous planning aren't simulated

My workshop is still imaginary. Its first helper now has a job I can watch, pause and adjust.

[Acorama @Acoramaa]

I made a 3D espresso machine with AI. You can pull it apart while it brews

Brew starts the shot. Inside reveals the water path. Explode separates the housing and internals while the coffee keeps flowing

The tubes follow the moving parts. The cup fills. Put it back together, and everything returns to the same scene

I kept the controls live and the project editable: materials, lighting, camera and motion can all be changed

It's a visual prototype with an illustrative water path, not a pressure or heat simulation

I wanted to show the machinery behind a familiar morning ritual. Build notes in the reply.

[Acorama @Acoramaa]

@irinatoxi good

[Acorama @Acoramaa]

@explosss1ve A good system, broo

[Acorama @Acoramaa]

I made a tiny game with AI. Your only job is to bring a dead lighthouse back to life

You're a brass keeper carrying a lantern. Search the harbour, the old ruin and the overlook. Bring the prisms back, then restore the beacon

I kept the build small enough to finish a whole loop:
1. Shape the island and lay out a walkable route
2. Build the keeper, lighthouse and three collection points
3. Add movement, solid obstacles and interaction prompts
4. Connect the inventory to the lighthouse switch
5. Record a complete run from an empty inventory to a working beacon

Each pickup changes the counter and map. The switch checks for all three prisms. Finish the task, and the light starts sweeping across the sea

The clip is recorded gameplay, with a scripted keyboard route. The playable version lets you take your own route

It's a small prototype: one island, one quest, no combat or save system

I wanted the world to give you something to do. Turning the light back on felt like enough.

[Acorama @Acoramaa]

I ASKED AI FOR A CABIN IN THE WOODS. THEN MADE ITS GLASS FRONT SLIDE OPEN.

Timber frame. A sleeping loft. A deck facing the trees. And four glass panels that slide aside until the living room opens straight onto the terrace

The film starts with the finished cabin, rewinds to its first pieces, then assembles the space in eight stages

The roof lifts away so you can see the loft, stairs and living area. Then it returns, the glass opens and daylight gives way to warm evening light

The model and camera sequence are editable. This is a concept animation, not a validated timber-frame design or a physical construction simulation

I wanted the forest to feel like part of the room. That final shot is the whole point

[Acorama @Acoramaa]

I ASKED AI FOR MY DREAM WORKSHOP. THEN MADE IT SHOW THE ASSEMBLY.

Garage downstairs. Creative studio upstairs. Tools, a lift, workbenches and a place to make things that don't fit on my desk

The finished building appears first. Then it comes apart

Pads. Columns. Beams. Bracing. The upper floor. Stairs. Roof. Glass. Finally, the spaces inside

The camera stops at a beam connection before pulling back to reveal both levels

What's actually in the model:
> 84 frame members
> 264 modeled connection bolts
> 18 stair treads
> 10 assembly stages

These are geometry counts, not invented construction statistics

It's an editable 3D concept, not a certified structural design. Loads, connections and clearances still need proper engineering checks

But I can change a component, move the camera and regenerate the sequence

The workshop is imaginary. The model is something I can keep working on

[Acorama @Acoramaa]

I ASKED AI TO PUT AN OCEAN INSIDE A BLACK CUBE

Then I asked for one wave to escape

The shell slides apart. A tiny sailboat rides the surface. The water rises beyond the box, curls into a suspended wave, then settles as the cube closes

The boat is my favourite detail. It gives a small patch of water a sense of scale

An AI-assisted 3D animation, built around one impossible object. The motion is choreographed, rather than calculated by a fluid solver

Twelve seconds. Back to a black cube. Until it opens again

[Acorama @Acoramaa]

CLAUDE OPUS 5.5 SHOWED ME THAT A VIDEO IS A 3D OBJECT.

Take a 4-second render. 120 frames.
Don't play them. Stack them, one behind another.

Now time is a direction you can walk along.

The little guy spins once in the clip, so the stack comes out twisted. That twist is the turn, frozen. The dark stripe on top is his beanie. The two dotted lines are his eyes, travelling around him.

Then the fun part. Cut the block on a slant, and every column of pixels comes from a different moment. Same 120 frames, a completely new video.

Press it flat again, and it's just a clip.

Why I care: my article is about handing one render to an agent and getting a video back. This is what you are actually ordering. Every extra second is 30 more slices, and on Picsart video is priced per second. That's why the agent should get a quote before it renders anything.

Article below.

[Acorama @Acoramaa]

https://t.co/ns7lpuiIzt

[Acorama @Acoramaa]

CLAUDE OPUS 5.5 TOOK MY FLUFFY STAR APART TO EXPLAIN ONE TOOL.

Top half: the creature, piece by piece. Eye, arm tip, belly pile, then the beanie lifts and the fur pulls back to the core. One script built all of it.

Bottom half: what happens when an agent sends that render to Picsart MCP.

6 steps. Only 1 of them costs money.

pick a model, free
read its exact inputs, free
preflight: validate + quote, free
generate, charges
collect the job, free
view_image: the agent looks at the result, free

That last one is the part nobody talks about. The agent checks its own output before it spends again.

And most guides on this are already out of date. On Sept 29 Picsart rewrote its docs and said the old setup instructions were wrong. No CLI, no API key, one address.

I wrote down what changed and how I'm wiring it for my renders. Link below.

[Acorama @Acoramaa]

CLAUDE OPUS 5.5 TOOK MY FLUFFY STAR APART TO EXPLAIN ONE TOOL.

Top half: the creature, piece by piece. Eye, arm tip, belly pile, then the beanie lifts and the fur pulls back to the core. One script built all of it.

Bottom half: what happens when an agent sends that render to Picsart MCP.

6 steps. Only 1 of them costs money.

pick a model, free
read its exact inputs, free
preflight: validate + quote, free
generate, charges
collect the job, free
view_image: the agent looks at the result, free

That last one is the part nobody talks about. The agent checks its own output before it spends again.

And most guides on this are already out of date. On Sept 29 Picsart rewrote its docs and said the old setup instructions were wrong. No CLI, no API key, one address.

I wrote down what changed and how I'm wiring it for my renders. Link below.

[Acorama @Acoramaa]

https://t.co/ns7lpuiIzt

[Acorama @Acoramaa]

https://t.co/ns7lpuiIzt

[Acorama @Acoramaa]

CLAUDE OPUS 5.5 MADE THIS IN BLENDER WITHOUT EVER TOUCHING THE MOUSE.

No sculpting. No brushes. No grooming the fur by hand.

It wrote one Python script, and the script builds everything: the star-shaped body, the bead eyes, the knitted beanie with a pom-pom, the studio lights, the camera path. Then Blender renders it.

The fur is the part that gets me. Over 650,000 strands, and every single one is coloured the same way: dark at the root, warm brown in the middle, amber at the tip. That gradient is why it looks soft instead of plastic.

12 seconds, one loop, rendered on a laptop GPU.

[Kurt @0xkurt]

I BUILT A JEV TRADING DESK WITH 6 AI OPERATORS — AND NONE OF THEM ARE ALLOWED TO GUESS.

Below 85% confidence, the system stops and hands the decision to a human.

Most “multi-agent” demos hide everything inside one giant prompt. This desk does the opposite. Every role stays separate, every permission is explicit and every handoff leaves evidence:

WATCHER → observes BTC, ETH and SOL
ANALYST → turns market state into shared context
RISK → checks exposure, drawdown and order size
EXECUTOR → places, cancels or closes
REPORTER → writes the evidence trail
HUMAN → receives anything below the confidence gate

Jev sits between those roles as a fast decision layer. It does not write a market essay. It receives bounded state and chooses from legal actions: LONG / SHORT / HOLD, AUTO / HANDOFF, EXECUTE / STOP.

The code still owns the rules. If risk confidence is at least `0.85`, the order may continue. If it falls below `0.85`, automation stops and the human operator gets the full shared context instead of a confident guess.

That changes what “autonomous” means. The goal is not to remove humans from the desk. The goal is to make the exact moment a human is needed visible, cheap and auditable.

Every action can answer four questions:

Who saw the signal?
Who changed the state?
Who approved the order?
Why did the system hand off?

The browser interface, role routing, confidence gate and evidence flow are real. The market data, decisions and telemetry shown in the video are scripted; no live orders were placed.

The pattern is simple:

Jev decides.
Code enforces.
Humans handle the gap.

That is a much more useful version of an AI trading desk than six agents quietly agreeing with each other.

Would you trust this system more because it knows when to stop?

[Acorama @Acoramaa]

CLAUDE OPUS 5.5 BUILT ME A CITY OUT OF 1,665 BRICKS.

Every single one placed in order, from code.

First the roads snap onto the baseplate. Then the towers start rising, layer by layer, each brick dropping in and clicking into place. Staggered joints, locked corners, windows exactly where they should be.

Then the park, the street lamps, the little cars.

The sun goes down. Windows light up one by one. The cars start driving.

The counter in the corner is real: 1,665 bricks, 4,797 studs on them, not one placed by hand.

26 seconds. Watch it till the lights come on.

[Kurt @0xkurt]

JEV HAS BEEN OUT FOR 17 DAYS AND DEVELOPERS HAVE ALREADY PUT IT ONE FUNCTION CALL AWAY FROM REAL MONEY.

Live order books. On-chain blocks. Exchange APIs. Trading agents. Entire simulated hedge funds.

Here is what actually shipped:

jev-trader → watches the Kuru MON-USDC book on Monad and asks Jev to pick buy or sell every block, roughly every 300ms. Without a private key it runs as a dry-run with simulated fills.
https://t.co/UQDnZBYLlG

jev-trade → sends Hyperliquid market state to Jev, which picks direction, intent and leverage across five assets. The project supports dry-run mode; live keys can place orders.
https://t.co/wa66nY03cV

beebots → three OKX trading “bees” use real market prices and ask Jev to choose from the moves available right now. Paper trading is the default, and deterministic code can veto or resize the decision.
https://t.co/Rfr4cwbYnH

QuantDinger → places a Jev decision gate directly before entry orders. Jev returns typed checks; code rejects blocked entries, while exits and emergency actions bypass the AI.
https://t.co/y7BcVwWy7A

ai-hedge-fund → a multi-agent fund simulator that can run its investment agents through Jev. The repository explicitly says it does not execute real trades.
https://t.co/6rVJ2U8Ule

And this is the part people keep missing: Jev is not being asked to write a strategy or explain the market. It gets a bounded state and a legal menu:

LONG / SHORT / HOLD
OPEN / CLOSE / WAIT
ENTER / REJECT
ONE ACTION / ONE CONFIDENCE SCORE

The large model can design the strategy. Jev makes the fast typed call. Deterministic code still owns sizing, limits, stops and execution.

That final boundary matters. A fast decision model is not a profitable strategy, and “live-capable” is not evidence of live returns. These are engineering experiments with very different defaults and risk controls.

But the pattern is real:

market moves → Jev picks → code checks → order or no order

No essays. No chain-of-thought parser. No expensive model debating every tick.

Save the five repos before everyone starts calling this an AI hedge fund.

[Acorama @Acoramaa]

GPT-6 ASTRA BUILT ME A BLACK HOLE SANDBOX. I SENT THE SAME PLANET INTO IT THREE TIMES

One stays in orbit. One swings past and escapes. One disappears into the center

The planet stays the same. Its starting position and velocity change

The useful part is underneath the glow: each path is calculated from gravity, with a capture boundary at the center

Speed alone doesn't decide the ending. Energy tells you whether a trajectory is bound; angular momentum determines how close it gets. A path that crosses the capture boundary is absorbed

This is a simplified Newtonian model with stylized visuals, rather than a full relativistic black hole simulation

The $0.84 headline raises the next question for me: how much checking does a cheap decision leave behind?

For a Jev layer, I'd ask it to choose orbit, escape or capture, then compare its answer with the calculated trajectory. Record the mistakes, the uncertain calls and the work needed to check them

That would be the next experiment. This clip shows the simulation

The number I'd want at the end is cost per correct outcome, including verification.

[Luca Shaffer @lucabuildsx]

https://t.co/QJdbjWkyqN

[Acorama @Acoramaa]

GPT-6 Astra made me a watch. Then I had it take the whole thing apart

Brass gears, a dark metal case, silver hands and an exposed spiral, all built as a 3D model

The camera moves in close, pulls back, and the assembly separates into floating layers. Then everything slides back into place

I wanted the inside to be the subject. The small screws, the gaps between the gears, the light catching the edges of the bridges

It's still a visual concept; the mechanics haven't been validated as a working watch

My favorite moment is just before it closes, when you can still see how the layers fit together.

[Nazar @Nazik2053]

a photon hits a fruit fly's eye

166,700 neurons decide what the body does next

scientists reconstructed the complete brain and central nervous system of an adult male fruit fly: 166,700 neurons, 124.2 million synapses and 11,710 cell types

this is more than a picture of a brain. it is a wiring diagram detailed enough to follow a signal from the photoreceptors into the optic lobe, across the brain, down the nerve cord and toward the motor circuits that move the body

once those connections exist as data, developers can build a loop around them

camera pixels become sensory input. activity passes through the reconstructed network. motor neurons become commands for a controller

that is how fly-derived neural systems ended up inside experiments with Doom, Mario and Beat Saber

this video is a conceptual signal trace, not a live recording of neural activity. and the current demos do not prove that a fly learned to play a game

the interesting part is the translation

the network sees input, routes activity and produces movement. it does not know whether that output turns a leg, moves Mario or swings a virtual saber

we spent years drawing the wires

now the strangest experiments begin after the map is finished