4 ms·

I don't see that day coming in my lifetime. There are simply too many design variables. You can't engineer a totally new product with AI. It's one thing to make

I don't see that day coming in my lifetime. There are simply too many design variables. You can't engineer a totally new product with AI. It's one thing to make improvements to an existing design.

brudgers 11y ago
Most products aren't totally new. Some things that are currently designed, such as buildings, are relatively simple and largely built the same way they were decades and centuries ago. Same is true for furniture, hand tools, cookware, cutlery, etc. The thing is, the idea that there are many design variables is contrary to recent AI development. Instead of creating a bunch of rules and putting them in a specification, the machine just looks at what does and doesn't work and decides what might work.
- ddingus 11y ago
  BTW, existing, rule based CAD nails those cases. Hell, I was showing people how to make parametric buildings in the mid 90s... Those standard cases have been ripe for the picking, and it's been done for a long time now, usually with a combination CAD and rule system, and verify ways for humans to input various things. There will continue to be nice gains and a simplification of CAD as well as the factoring it out on those. We have never needed CAD to do that stuff. CAD just helped do it better, more efficiently. For the growing body of stuff we do need CAD to do, these approaches are nowhere near the maturity needed. Novel things do require people, and they must iterate, collaborate, analyse, etc... CAD is the best we've got. When we get a real AI, that has real analytic and creative skill, maybe. Will be exciting too. Long way off though.
  - brudgers 11y ago
    It's a question of who gets to "input various things". CAD systems whether or not they incorporate heuristics or parametrics are based on: human owner -> human designer -> production [possibly human] -> artifact However, the important high level abstraction is: human owner -> artifact where the "->" simply accounts for the time gap that we call "production". Getting closer to the high level abstraction is why people pay for design. For many projects having humans in the "->" is seen as a cost not a benefit...they're forced to pay a designer for time spent rounding corners and gradienting drop shadows because those things are more fun and easier than iterating. It's also more fun than writing up vast arrays of heuristics and parametric constraints. Computers are really really good at iterating. With machine learning, they don't require long lists of heuristics and parametric constraints. Just a lot of examples and an unimaginably large number in human terms of CPU cycles. That's not a place AutoDesk as a publicly traded company can pivot. So in the 1990's when you were developing computational approaches, they pivoted toward a product line architecture that allows Autodesk to roll out features at regular intervals. If it had wanted to go computational, it would have doubled down on AutoLisp with macros and promoted developers. There's a business reason that full blown AutoCad developer tools require approval and that there's no AutoCad app store.

I don't see that day coming in my lifetime. There are simply too many design variables. You can't engineer a totally new product with AI. It's one thing to make improvements to an existing design.

brudgers 11y ago

Most products aren't totally new. Some things that are currently designed, such as buildings, are relatively simple and largely built the same way they were decades and centuries ago. Same is true for furniture, hand tools, cookware, cutlery, etc. The thing is, the idea that there are many design variables is contrary to recent AI development. Instead of creating a bunch of rules and putting them in a specification, the machine just looks at what does and doesn't work and decides what might work.

ddingus 11y ago
BTW, existing, rule based CAD nails those cases. Hell, I was showing people how to make parametric buildings in the mid 90s... Those standard cases have been ripe for the picking, and it's been done for a long time now, usually with a combination CAD and rule system, and verify ways for humans to input various things. There will continue to be nice gains and a simplification of CAD as well as the factoring it out on those. We have never needed CAD to do that stuff. CAD just helped do it better, more efficiently. For the growing body of stuff we do need CAD to do, these approaches are nowhere near the maturity needed. Novel things do require people, and they must iterate, collaborate, analyse, etc... CAD is the best we've got. When we get a real AI, that has real analytic and creative skill, maybe. Will be exciting too. Long way off though.
- brudgers 11y ago
  It's a question of who gets to "input various things". CAD systems whether or not they incorporate heuristics or parametrics are based on: human owner -> human designer -> production [possibly human] -> artifact However, the important high level abstraction is: human owner -> artifact where the "->" simply accounts for the time gap that we call "production". Getting closer to the high level abstraction is why people pay for design. For many projects having humans in the "->" is seen as a cost not a benefit...they're forced to pay a designer for time spent rounding corners and gradienting drop shadows because those things are more fun and easier than iterating. It's also more fun than writing up vast arrays of heuristics and parametric constraints. Computers are really really good at iterating. With machine learning, they don't require long lists of heuristics and parametric constraints. Just a lot of examples and an unimaginably large number in human terms of CPU cycles. That's not a place AutoDesk as a publicly traded company can pivot. So in the 1990's when you were developing computational approaches, they pivoted toward a product line architecture that allows Autodesk to roll out features at regular intervals. If it had wanted to go computational, it would have doubled down on AutoLisp with macros and promoted developers. There's a business reason that full blown AutoCad developer tools require approval and that there's no AutoCad app store.