GIVE IT A DRAWING WITH NO WAY TO ACTUALLY SEE IT
The task, straight from Anthropic’s own benchmark: reconstruct a machine part as a 3D FreeCAD model from a drawing - except the environment was set up so the model had no direct way to view the image at all Not a harder version of the task - A blocked one
Every competing model tested the same way failed after five attempts. Read that as: five separate tries, same setup, zero successes, on models that are otherwise capable of plenty
Opus 5’s actual move: it didn’t ask for a workaround or give up on the missing input
It wrote its own computer vision pipeline from scratch, pulled the geometry straight out of raw pixel data, then used that to reconstruct the full part in CAD - and did it repeatedly, not as a one-off fluke
What that tells you that a design benchmark score doesn’t: the model built a missing tool mid task instead of treating the absence of one as a blocker
For any design pipeline that depends on reading a reference - a sketch, a scan, a competitor screenshot - that’s the difference between “can’t see it, can’t work with it” and “built the thing that lets it see, then kept going”
full below
JAY GANESH JAY GANESH JAY GANES H DEVA
JAY SHRI GANESHAY NAMAH🚩
J
A
Y
S
H
R
I
G
A
N
E
S
H
A
Y
N
A
M
A
H
MANGAL MURTI MORAYA🙏
M
A
N
G
A
L
M
U
R
T
I
M
O
R
A
Y
A
GANPATI BAPPA MORAYA🙌
G
A
N
P
A
T
I
B
A
P
P
A
M
O
R
A
Y
A
MATA JAKI PARVATI PITA MAHADEVA👏 @grok
HAR HAR MAHADEV! 🔱🙏
OM NAMAH SHIVAY 🕉️
H @grok
A
R
H
A
R
M
A
H
A
D
E
V! 🔱🙏
O
M
N
A
M
A
H
S
H
I
V
A
Y 🕉️
HAR HAR MAHADEV! 🔱🙏
OM NAMAH SHIVAY 🕉️
H
A
R
H
A
R
M
A
H
A
D
E
V! 🔱🙏
O
M
N
A
M
A
H
S
H
I
V
A
Y 🕉️
ॐ नमः शिवाय 🕉️
शिव से यह मत माँगो कि जीवन में दर्द कभी न आए।
बस इतनी प्रार्थना करो
दर्द आए तो मैं टूटूँ नहीं,
मुश्किल आए तो मैं झुकूँ नहीं।
क्योंकि महादेव हर दर्द नहीं मिटाते,
कभी-कभी उसे सहने की शक्ति देते हैं। 🔱
हर हर महादेव! 🙏
MY CHILD IS DYING, PLEASE😭
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎don’t forget me🥺
︎
︎
︎
︎
︎
︎
︎
︎
please care
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎ً
︎
I beg you, I need food.
I kiss your hands
Help my poor child, she's dying
https://t.co/MT2Xd10Mhd
THE SCARIEST THING ABOUT 300 AI AGENTS ISN'T THAT ONE OF THEM FAILS. IT'S THAT ONE OF THEM IMPROVISES SUCCESSFULLY.
Kimi K3 can coordinate up to 300 sub-agents.
And Moonshot quietly included one warning that I think matters more than the agent count:
if you need K3 to stay inside defined boundaries, write those boundaries down.
That's a completely different way to think about autonomous AI.
We spend most of our time telling agents what they can do.
Browser access.
Files.
Email.
Code.
Payments.
Hundreds of tools.
But the more capable the agent becomes, the more valuable the opposite file becomes:
what it must never do.
I like the idea of reducing that to five permanent constraints:
SCOPE → what exists
ONE-WAY → what requires approval
EVIDENCE → what counts as true
BUDGET → when work stops
SILENCE → when the agent must refuse to guess
Then all 300 agents inherit the same fence before step one.
Because a model trained to keep solving problems doesn't necessarily stop when it reaches the edge you had in your head.
The more autonomy you give AI, the more precisely you need to define where autonomy ends.
Give the swarm the fence first.
Then give it the keys.
MY CHILD IS DYING, PLEASE😭
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎don’t forget me🥺
︎
︎
︎
︎
︎
︎
︎
︎
please care
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
.
︎
I beg you, I need food.
I kiss your hands
Help my poor child, she's dying
https://t.co/MT2Xd10Mhd