Thanks @theinformation for highlighting Build-A-Chip-Bench.
In 2024, we started exploring how AI could help design chips, from digital to analog.
That work kept bringing us back to one question: can AI carry design intent all the way through implementation while meeting both functional and physical requirements?
In collaboration with @nvidia, we created Build-A-Chip-Bench to put that question to the test.
In our initial evaluation, the highest frontier model score was only 16.6%.
There’s a lot of work ahead.
But these failures are also encouraging: they help us clearly see the next research questions.
We’d welcome AI research and engineering teams to explore these questions together and bring chip superintelligence to reality.
In 2024, we started exploring how AI could help design chips, from digital to analog.
That work kept bringing us back to one question: can AI carry design intent all the way through implementation while meeting both functional and physical requirements?
In collaboration with @nvidia, we created Build-A-Chip-Bench to put that question to the test.
In our initial evaluation, the highest frontier model score was only 16.6%.
There’s a lot of work ahead.
But these failures are also encouraging: they help us clearly see the next research questions.
We’d welcome AI research and engineering teams to explore these questions together and bring chip superintelligence to reality.
Anthropic is telling IPO investors that its 15-person chip team is already using Claude to speed up chip design, potentially giving it an advantage over established chipmakers like Nvidia and AMD.View quoted post on X
Read more: thein.fo/4zBsdSc
