Spent years designing cement jobs on a screen. On the screen they always looked great. Then you pump it and the well does whatever it wants. It’s like baking a cake inside a pipe you can’t open.
The chicken story is the same thing. Every weight on the way in was fine, and the water was still in the bird.
Designing is getting cheap. Knowing what really happened down there is where the money is.
AI may make hardware rediscover a very old bottleneck: reality has a refresh rate. You can generate a million designs before lunch, but the machinist, thermal chamber and chicken still refuse to operate at token speed. At some point intelligence gets cheap and contact with reality becomes the scarce resource.
CAD Arena can rise while the lossy step this piece is about stays unmeasured. A model that one-shots a bracket still hands a machinist a file that has to become stock, workholding, and a toolpath, which is the translation step where the errors enter. I would score the same design revision twice, once on the file and once on a measured part, before reading a benchmark gain as the model getting better at hardware rather than at drawings.
Christian, the idea of a lossy interface between the digital and physical worlds really resonates.
We're experiencing a version of this at Flox, where we're building Nature Intelligence, an operational intelligence layer for machines interacting with the natural world.
The difficult part isn't just sensing or deciding what to do... it's understanding what actually happened after a system acted, whether that action made a difference, and what can be learned from it.
In the natural world, those outcome signals are noisy, incomplete, and often ambiguous.
I suspect the systems that successfully close this real-world feedback loop will have a significant advantage as the underlying models continue to improve.
Excellent post! I spent 35 years in manufacturing and supply chain management, and you nailed the high level requirements across the design chain. Great stuff!
Spent years designing cement jobs on a screen. On the screen they always looked great. Then you pump it and the well does whatever it wants. It’s like baking a cake inside a pipe you can’t open.
The chicken story is the same thing. Every weight on the way in was fine, and the water was still in the bird.
Designing is getting cheap. Knowing what really happened down there is where the money is.
AI may make hardware rediscover a very old bottleneck: reality has a refresh rate. You can generate a million designs before lunch, but the machinist, thermal chamber and chicken still refuse to operate at token speed. At some point intelligence gets cheap and contact with reality becomes the scarce resource.
CAD Arena can rise while the lossy step this piece is about stays unmeasured. A model that one-shots a bracket still hands a machinist a file that has to become stock, workholding, and a toolpath, which is the translation step where the errors enter. I would score the same design revision twice, once on the file and once on a measured part, before reading a benchmark gain as the model getting better at hardware rather than at drawings.
Superhuman Hardware Engineer..no. Manufacturing Engineer yes. The "How" is called "kaipability" dyor 👍🏽
Christian, the idea of a lossy interface between the digital and physical worlds really resonates.
We're experiencing a version of this at Flox, where we're building Nature Intelligence, an operational intelligence layer for machines interacting with the natural world.
The difficult part isn't just sensing or deciding what to do... it's understanding what actually happened after a system acted, whether that action made a difference, and what can be learned from it.
In the natural world, those outcome signals are noisy, incomplete, and often ambiguous.
I suspect the systems that successfully close this real-world feedback loop will have a significant advantage as the underlying models continue to improve.
Excellent post! I spent 35 years in manufacturing and supply chain management, and you nailed the high level requirements across the design chain. Great stuff!