Acorama (@Acoramaa) header
Acorama (@Acoramaa) profile picture

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

Miami
Joined July 2026
Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)Photo by Acorama (@Acoramaa)
An AI optimizer makes your chip better at exactly one thing: whatever you wrote down. Everything else is luck Yesterday Claude's optimizer beat a greedy layout on wiring. Same toy chip, 13% less wire. Today I ran the heat Look at the hot spot. 93.3°C. The optimizer never knew heat existed. It beat the greedy layout's 96.8°C almost by accident, just by spreading blocks out to make room for wires So I added one line to its goal: a penalty for every spot above 82°C. Same optimizer, same random start, 600,000 moves The peak dropped to 85.5°C. Almost 8 degrees cooler on the same 125.6 W And it paid for it. 35% more wire than yesterday's winner That's the line from the article I keep coming back to: an optimizer can only account for the constraints you give it. It found a cooler chip the moment I asked, and a longer one without asking AI can try 600,000 layouts in two and a half minutes. Picking the trade you can live with is still on us
2
1
2
51
2,678
The scariest thing about AI in chip design is that it doesn't trust a layout just because it looks right. We do Look at this. One toy AI chip, 83 buses, placed two ways Layout A puts each block in the best spot for itself, one at a time. Neat, and the wires look short on paper. The last block had no legal spot left and got squeezed in Then the router tries to connect everything. 10 spots run out of room Layout B is Claude's optimizer moving all 29 blocks at once. 600,000 moves in 103 seconds, and early on it takes worse moves on purpose so it can't get stuck. It looks messier and its wires are 10% longer on paper Routed, it uses 13% less wire. 1 spot over capacity instead of 10 That's the trap the article describes: placement decisions interact, and a choice that looks good locally makes the later steps harder To be fair, this is a toy. No timing, no billions of cells, and simulated annealing instead of the reinforcement learning Google uses for real TPUs But the lesson scales. What AI takes over first is the moment we stop searching because the layout already looks finished
3
7
1
94
5,842
AI is coming for engineers. And the terrifying part is that engineers are helping it happen. Look at this. You're watching the insanely complex layout of a computer chip, with layers of microscopic connections that normally require years of expertise to design and optimize. Now AI is learning to handle parts of that process. Chip placement, routing, design optimization. Work that once demanded highly specialized engineers is becoming increasingly automated. And it gets worse. Engineers are building AI-powered tools to design the very chips that will run the next generation of AI. I'm not saying every engineer is getting fired tomorrow. But when one engineer can potentially do the work that previously required an entire team, things get uncomfortable very quickly. We spent years worrying about AI replacing artists and writers. Meanwhile, it's working its way into the jobs we thought were too technically complex to automate.
5
15
2
155
9,136
I had Claude build a GPU server in Blender, scan it, then kill one of its six fans The GPU two rows behind it went from 81°C to 110°C Here's the prompt, swap in any [DEVICE]: Write one Blender Python script that builds [DEVICE] as a 9:16 thermal scan Look: smoky transparent glass parts with thin dark outlines on white, soft studio light 1. Model every part as its own object, with its real power draw 2. Get every temperature from a simple model, never by hand: air enters at 25°C and warms as it passes each part, each part = local air + power × heatsink resistance 3. Each part starts as glass, then switches to a thermal colour (blue 20°C to red 110°C) as a sweep passes over it 4. Break one thing (a fan, a vent, a heatsink) and keyframe the hottest part climbing to its new temperature 5. Export the camera and each part's 2D box per frame so a tracking HUD can go on top 1080×1920, 30 fps. Render one still first and check it before the full run
1
11
526
I had Claude build the fish in code and solve both before filming: > Aim: light bends at the surface, so the bug isn't where the fish sees it > Jet: one squirt would spread out and arrive weak > Rule: every number on screen comes from the solve, nothing typed in • What the aim solve gave > The bug appears at 54° > It really sits at 40° > The fish fires 14° lower than it sees • What the jet solve gave > 40 droplets leave over 35 ms > The first at 3.0 m/s, the last at 5.0 m/s > The tail catches the head and all of it lands at once, 80 ms later Then the replay, 30× slower: gill covers snap shut, the piston drives the water down the barrel, the jet bunches up on the way After that it comes apart. Eyes, jaw, barrel, gill covers, pump, fins, tail, 8 shell rings. At 0:19 nothing is attached The prompts are the same 7 from the article ↓
10
583
I had Claude build a GPU server in code, then assemble it the way the parts list orders it. It got stuck 3 times before the last screw A parts list is sorted by part, not by reach. Nobody checks whether each part can still get to its place after the ones before it So Claude ran the build as a dry run: > Order: 12 steps, parts-list order first, big parts in early > Reach: every part is lowered straight down and checked against what's already installed > Screws: all 53 get the same test for the screwdriver path • What run 1 found > 8 of 16 memory sticks blocked, the CPU heatsinks cover the inner slots > 3 of 7 baseboard screws sitting under the cross brace > A power plug that can't go in, the rear GPU heatsinks overhang the connector • What run 2 changed > Memory goes in before the heatsinks > Screws get driven before the brace goes on > The cable gets plugged before the GPUs drop in Three steps moved. 0 blocks, 53 of 53 screws The first thing the check caught was my own model. The heatsink fins sat lower than the memory, so no order could ever have worked Same parts, new order. The fix was the sequence, not the design How the check works is in the reply ↓
1
8
369
A turbofan hides two shafts, one inside the other, turning at different speeds I had Claude build the engine in three.js - 22 fan blades, 11 compressor stages, 6 turbine stages - then pull it apart along its axis The gauges read the same shafts that spin the blades
17
516
To test Claude Fable 5.2, I asked it to take my hexapod off-world Same 14 kg robot, same code. It weighs 52 N on Mars, 23 N on the Moon and 18 N on Europa, and every foot finds its own ground Drive it yourself, link below glittery-stardust-885b97.net…
5
11
264
Owls lock their grip with a ratchet in the tendon, so they can sleep on a branch without holding on I asked Claude to build that as a machine. 0.17 s of grip at 20× slower - 6 teeth click and it holds with zero power Then it takes the whole owl apart. By 0:19 nothing is attached
1
11
164
I gave Claude one picture of a robot chameleon. It rebuilt every plate in code, checked its renders against the picture, then took the whole thing apart on camera A new part comes loose every 2.5 seconds. By 0:19 nothing is attached to anything The article has the 7 prompts I used, what each check round caught, and the job file
1
11
445
I asked AI to build a mechanical mantis shrimp that could punch a brass target and take itself apart The first hit lands before the two-second mark. Then the glass shell lifts and the camera moves in > the coil compresses while the club is held back > the latch pin lifts, the club swings and the brass plate snaps back > seven armor plates lift off the spine, and each eye splits into a lens, a ten-blade iris and a socket The same timeline runs the parts, the readouts and the sound. I checked the close-up frames and moved a side panel down so you can see the spring through the whole replay The real animal inspired the spring-and-latch idea. This is my animated mechanical version, built with AI.
1
9
247
this is going to put a LOT of 3D animators on notice I showed Claude one reference clip and asked for a mechanical jellyfish. it built the whole thing in code, then took it apart on camera > no 3D software. every rib, hose, lens and link is three.js code > the glass bell sits over 16 ribs and 144 plates, and it actually contracts on every pulse > 16 tentacles, 26 links each. one wave runs down every arm and the tips curl into spirals > 8 lenses on the rim, each opens into socket, light ring, 10-blade iris and dome > one script runs it all: the cursor hitting each tab, the jet pressure card, the callouts and the sound > it can't watch video, so it renders stills at the busiest moments and fixes what looks off before the full render > rendered at 1.25x, downscaled to 1080p, score sits at -14 LUFS one reference clip in, this jellyfish out
2
17
538
I gave AI one picture of a robot chameleon. It rebuilt the whole thing in code, then took it apart. > the picture was only a target. every plate, gear and lens is three.js code > the eyes are two loose turrets with a 10-blade iris, each one tracks on its own > the tongue is a 16-link chain with a spring and a suction cup. it fires and pulls back > a glass window opens and the core slides out with 3 gears still turning > 34 shell plates, 12 spine plates, 26 tail segments in a spiral > a new part comes loose every 2.5 seconds until 0:19, when nothing is attached to anything 235 lines of code. the picture and the job file I gave AI
8
385
Claude built me a mechanical scorpion. At full speed the sting is gone in 0.12 seconds So I replayed it 20x slower, faded the shell and left the mechanism running A motor turns a cam. The cam winds a spring. A latch holds it until it doesn't, and the whole tail whips over the body 8 legs walk in an alternating tetrapod gait, always 4 feet on the ground. Every gauge reads the same sim that moves the parts 38 joints, 205 lines of code, then it takes itself apart on a full turn and walks off
2
9
605
Claude built me a mechanical hummingbird. At full speed the wings are just a blur So I slowed it down 40x, faded the shell and kept the drive running A small motor turns a 10:30 gear pair, a crank and two telescoping rods. Each rod rocks one wing. You can watch the blade flip at the end of every stroke At 0:12 it backs off the flower and comes apart into ten assemblies. Then it puts itself back together, speeds up and leaves Every wing angle, the crank, the readouts and the sound follow one shared phase Fictional mechanism, original geometry and sound. The flight is choreographed, there's no aero solver in it
12
344
Claude built my mechanical chameleon. The fly gets less than two seconds It grips the perch, turns its eyes and fires the tongue. Then I made it open its head and repeat the catch in slow motion At 0:07 you can follow the carriage sliding on its rails, the spool turning and the tongue pulling the little robotic fly back in through the faded shell Then the head panels separate while the catch runs again. The parts return to their places and it gets one more fly The useful part of the build: tongue, carriage, spool and captured target all follow one animation state. Change the stroke and they stay together Made with Claude, including the geometry, motion and original soundtrack. It's a fictional mechanism with a choreographed catch I've put the build brief and the animation recipe in the first reply.
2
177
Claude built me a mechanical dragonfly. Then I made it expose the wing mechanism while it kept moving Four transparent wings, six articulated legs and a segmented body. The little bronze linkages are the part I wanted to get close to The sequence slows the wings down, fades the shell and opens a detail view of the drive. Then the head, battery, wing assemblies and body panels separate, come back together and fly off One phase drives the cams, wing angles and the moving diagram. The assembly labels follow the parts as they separate The geometry, motion and soundtrack were made for this clip. The flight path is choreographed; this is a fictional mechanism study My favourite moment is the shell fading while the drive keeps turning.
1
1
12
546
This is insane. It's not a video model. It's a web page. Claude Opus 5.5 wrote every part of it in code: the six-legged walker, the gait solver, every gauge and button. The legs really walk. Every foot is solved with inverse kinematics, and the support polygon, leg loads and footfall chart all come from that same solver. Switch the gait from tripod to wave and it slows down and keeps 5 feet on the ground. How it's built is in the quoted post
2
11
414
Claude Opus 5.5 built a mechanical jellyfish I can take apart, with its own control panel. I gave it one line: "mechanical jellyfish, field simulator". This came back. the build: > white shell on 16 hinged petals, a glowing core with a 3-ring gyroscope inside > every petal, roller and gauge is three.js + canvas code. no models, no stock UI > 10 arms, 200 segments, 400 rollers. each arm is solved joint by joint every frame > the dashboard is live: pressure, thrust, arm curl and depth all read from the sim > it rendered test frames and fixed them for 3 rounds before filming > every 3 seconds something opens: the lens unscrews, the bell blooms, the pump drops out spinning > at 0:16 it blows apart on a 330° turn and reassembles before the turn ends 24 seconds, 280 lines of code, and it swims out of frame on its own.
8
356