vans163

There is a big problem here Nx.max is wrong, it should be Nx.clamp_min (does not exist), I am not sure how to clamp a tensor with Nx.

I can rebase this once the code review is squared away. BTW I am using this like this. Maybe there should be an option to allow the subtraction, or would...

I am debugging this and the LinAlg version is much slower (or inaccurate) compared to the squareroot version on both the CPU and GPU, a huge factor slower. The workload...

This is missing unit tests, if its good to merge as you say let me add them today.

@seanmor5 Have a look over I think this is final.

> Hey @HurricanKai! I don't have significant experience with reinforcement learning, though I think it should be possible with Axon's loop API and/or some custom implementation. Here is a detailed...

> blake3 is really fast + can parallize. Noticed a 60x speedup on 8core vs sha256 (i guess divide by 8 to remove the multithreaded advantage if that somehow should...

> @vans163 we can use anything that is part of Erlang/OTP. I assume that's not the case for blake3, correct? It seems it has to be included in openssl first,...

Can this get tagged on too, https://bugs.erlang.org/browse/ERL-1159. Basically to not push up the REPL prompt on Logger outputs. ``` iex(2)> :erlang.spawn fn()-> Process.sleep(1000); IO.puts "hi" end #PID hi iex(3)> iex(3)>...

@hauleth I think this PR https://github.com/erlang/otp/pull/2523#issuecomment-581790305, would allow one to easily set standard_io as the output device in logger? Probably remedying this? Then OTP error reports (like a crash report...