I don’t really agree. I had a Xerox 1108 Lisp Machine in the 1980s and loved it, but special purpose Lisp hardware seems like a waste of effort. I set up an emulator for the 1108 last weekend, and yes, I really did enjoy the memories, and things ran an order of magnitude faster than on the 1108 in the 1980s.
Then, I appreciated my M1 MacBook Pro running SBCL, LispWorks, Haskell, Clojure, and various Scheme languages - all with nice Emacs based dev setups. Life is really good on modern hardware.
The 1108 wasn't really special purpose Lisp hardware. One could run other operating systems on it. What made it special purpose was the loaded microcode for the CPU.
> Life is really good on modern hardware.
Agreed: On modern CPUs.
More support for the additional hardware features like GPUs, media processing engines and the neural network engines (see the M1 Pro/Max/Ultra) would be welcome.
The best bet for getting GPU deep learning support, I use Anaconda/conda, using the Apple M1 channel. That said, I usually use my Linux GPU rig or Colab for deep learning.
I feel like a lot of posts like this are pining for the complete lisp machine -user environment- and overestimating how necessary/important the hardware architecture would be to getting back to that today.
I can manage to context switch between different lisps fine but I do sometimes wonder in e.g. a slime+SBCL setup how much that context switching is costing me.
Then, I appreciated my M1 MacBook Pro running SBCL, LispWorks, Haskell, Clojure, and various Scheme languages - all with nice Emacs based dev setups. Life is really good on modern hardware.