Chip design. I briefly worked in a VLSI lab, and making custom chips for tasks can improve performance by 10x to 100x, but it’s hard and expensive. If you can design chips fast and cheaply, then you can flash FPGAs with custom circuits at runtime, dynamically reconfiguring your CPU/GPU as needed to be ultra performant at any task.
I’ve dreamt of this for years, and now RL has shown some promise in similar domains (AlphaTensor) so this dream may some day be a reality.
This is on my Ai wishlist now as well. Seen the use of Ai on jets and that already was quite interesting, can only imagine he raw performance were going to pump out by doing this
Chip design. I briefly worked in a VLSI lab, and making custom chips for tasks can improve performance by 10x to 100x, but it’s hard and expensive. If you can design chips fast and cheaply, then you can flash FPGAs with custom circuits at runtime, dynamically reconfiguring your CPU/GPU as needed to be ultra performant at any task.
I’ve dreamt of this for years, and now RL has shown some promise in similar domains (AlphaTensor) so this dream may some day be a reality.
This is on my Ai wishlist now as well. Seen the use of Ai on jets and that already was quite interesting, can only imagine he raw performance were going to pump out by doing this