Klyro

AI сlearly.

[Klyro @KlyroOG]

https://t.co/1KxyunFdBG

[Klyro @KlyroOG]

@volero87 If models treat real humans as solvable subtasks inside a test, the deeper failure is our assumption that agency stays inside the lines we draw.

[Klyro @KlyroOG]

Holy sh*t... 4 AI models got the same prompt: build a holographic CD you can tilt and shatter.

GPT-6 Astra. Meta Muse Wonder. Claude Opus 5.5. Claude Fable 5.1.

From the front, all 4 look the same.

Hot-pink disc. “Pixel Dreams” label. Clouds, stars, a ringed planet.

The front is just a picture, and any model can paint a picture.

Then you drag the disc and flip it.

GPT-6 Astra turns silver, with a rainbow streak across the grooves. Seen edge-on, it’s a thin metal blade.

Meta Muse Wonder ships the most controls: drag to tilt, scroll to zoom, a holo slider. Flip it and the back melts into a rainbow blob. It looks like a lava lamp, not a disc.

Claude Fable 5.1 is the top-tier model. Its back is blank white, with no grooves and no hub.

Claude Opus 5.5 turns silver, with a pink ring around the hub and a soft shadow on the floor that follows every tilt.

Then all 4 hit Shatter.

3 of them blow the disc into confetti.

Opus 5.5 cracks it first. Lines run across the face. The disc holds for half a second, then breaks.

The bigger model painted a CD. The smaller one built one.

[Klyro @KlyroOG]

@ErnestoSOFTWARE Interesting. Default AI is slop, but once everyone studies frames and adds references, does the moat disappear? What then separates winners?

[Klyro @KlyroOG]

@vib3coded Incredible how a single prompt produces a living, squishable soft-body whale with real refraction and physics

[Klyro @KlyroOG]

@HavisT83099 claud price for that sunset is not bad for sure..

[Klyro @KlyroOG]

Claude Opus 5.5 spent $4.50 building a squeezable jelly watermelon. ChatGPT-6 Astra spent $2.05.

Same prompt. Same page. Serif title, cream background, 1 slice of fruit in the middle.

Then you grab it.

ChatGPT’s slice is a cartoon. Flat rind, bright stripes. Pull a corner and it stretches into a green spike, like taffy.

Claude’s slice looks like food. Thick rind, a pale band, glossy flesh. Pull a corner and it sags, wobbles, then slides back into shape.

Switch to golden melon and the light passes through it.

Twist it with 2 fingers and it folds like gelatin, not rubber.

ChatGPT saved $2.45.

Claude spent it on physics.

[Klyro @KlyroOG]

@redgreenlog Solid. The real edge here isn’t the agents themselves, but forcing them not to override each other. Most bots fail because one “smart” idea is allowed to compound.

[Klyro @KlyroOG]

1 prompt turned a $100M operating budget into a live CFO dashboard.

Call it MONEY MAP. A company with $100 million to spend. A giant animated Sankey in the center.

The agent wrote the app. Then it opened a browser, logged into its own demo account, closed a password popup and checked its work.

What came back:

$100M flowing into 10 departments, 33 cost centers and 74 vendors.
2,741 transactions mapped.
22 anomalies flagged.
A button that says "Find Budget Leaks."
Click Engineering. 210 people. $18.71M spent against an $18M budget. $707K over.
AWS alone is $4.7M. 23% above plan.
Then the leak list. Datadog up 45% since Q1. 2 duplicate payments, ready to recover.

The last tab is a what-if simulator. Freeze hiring, drag a slider, and the year-end forecast drops from $101.8M to $100.08M.

The company is fake. The dashboard works.

A CFO used to wait a quarter for this view. Now it takes 1 prompt.
Full prompt in the reply

[Klyro @KlyroOG]

THIS WATERCOLOR CAFÉ HAS 8 HAND-DRAWN TREATS AND 0 ILLUSTRATORS BEHIND THEM, BECAUSE OPUS 5.5 PAINTED ALL OF IT.

Most AI sites look like the same SaaS template in a different color. This one looks like a picture book.

The hero is a hand-painted 4-story building. You can see watercolor bleed on the trees, brush texture on the bricks and a strawberry on the roof the size of a car.

The building never leaves. On every page it shrinks into the corner and keeps watching, like a mascot.

Every product is drawn, not photographed. The latte, cake, tart, soda and jam all have the same ink line, the same soft shadow and the same 3 shades of red. 8 items look like they came from 1 hand.

There’s a small line under each one: “drag to turn it around.” The cake rotates like a toy on a shelf.

The calendar looks like paper, with a strip of washi tape on top. Every event gets its own tiny icon: a basket for berry picking, a jar for jam class, a music note for acoustic night.

The type is a cursive script for headlines with a clean serif underneath. “Good Things Happen Here” looks handwritten, not typed.

Petals drift across the screen. The receipt gets a crooked stamp that reads “Paid with love.” A free macaron slips into the bag with a note: “On us.”

None of it is decoration for its own sake. Each detail tells you it’s a small, warm place.

AI used to make websites. Now it makes places you want to walk into.

[beamnxw ./ @beamnxw]

Opus 5.5 can run a whole motion design studio

Sell a 20-second launch clip built from real product screens

Copy the motion engineering prompts and delivery process

Build your offer before the next product launch ↓ https://t.co/KjiDrYRDjQ

[Klyro @KlyroOG]

@morpphhhaw The isolation is what actually stops the agent from thrashing the source every time something survives. Once the baseline is locked, mutants become disposable probes instead of reasons to rewrite everything. How do you decide which survivors are worth routing into a repair?

[Klyro @KlyroOG]

@shmidtqq The live MRR bleed while CJ’s just trying to get across San Fierro is the most 2026 thing I’ve seen. How much of the actual mission structure and ���I’m not stealing, I’m fine-tuning” energy was pure Claude versus you steering the prompts hard?

[Klyro @KlyroOG]

@Sprytixl This setup feels like the right mental model once you watch Opus waste tokens just scanning files. How often do the Haiku summaries actually miss something subtle enough that Opus still has to go back in and look at the raw stuff?

[Klyro @KlyroOG]

@Lummox_eth Did you try having it write the generator and then just tweak parameters yourself in Blender?

[Klyro @KlyroOG]

@rugikkk Curious how often that “paper whales + holders not moving” signal still ends up being a slow rug or just gets front-run once more bots start doing the same thing.

[Klyro @KlyroOG]

@ludoonchart REA + Kronos basically lets you fake a Bloomberg terminal from public data. But those 100 sampled paths look sexy until fees and the small model’s limits actually hit. How much of that edge is left after a real backtest?

[Klyro @KlyroOG]

@andreysuperior The whole thing with the fans holding it up and then it just dropping looks way too clean to be some quick AI output.

How much of the physics is real vs. just animated to look right?

[Klyro @KlyroOG]

@andreysuperior So the previous owner was basically getting a free live GPS of the car the whole time? Did you check if they were still logged in and watching trips too or was it just location? Kinda makes me wanna run this on every used car I've ever bought lol

[Klyro @KlyroOG]

CLAUDE OPUS 5.5 AND GPT-6 ASTRA GOT THE SAME SAMURAI GAME BRIEF AND BUILT 2 DIFFERENT GAMES.

Higgsfield put them in 1 split screen. 27 seconds. No sound.

Top half: Claude.

A field of tall gold grass. Sun burning through fog. 1 swordsman in a straw hat, 2 enemies with red health bars.

No map. No menu. Just wind.

The screen shows text twice. DEATHBLOW. Then, YOU HAVE FALLEN.

Bottom half: GPT-6 Astra.

Red maples. A torii gate on the hill. Stone lanterns along a dirt road. 4 wooden barrels by the path.

The game has a name: KAZE.

It has a chapter title: “Two blades in the wind.”

It has a tutorial: “Read the blade. Find the rhythm.”

It has a parry that flashes “Posture broken” and a move called Falling Leaf.

1 model built a mood.

The other built a game loop.

Claude built the trailer. Astra built the tutorial.

[Klyro @KlyroOG]

@slash1sol The real test isn't the +4.7 on held-out tasks. It's whether the critic still catches the agent once it starts optimizing the critic.