Ariel Kwiatkowski

Re: changelong - do you mean in the docstring? That's already done. The docs site page (I think) gets automatically generated from that. Re: experiments - probably too much effort...

Sorry about the delay. I was a bit concerned about potential performance implications of this, since jax jit can be a bit moody sometimes with its optimizations, but from a...

> I don't think it's an MDP if you change it to be a termination I don't really buy this. The observation is a pixel rendering in a radius around...

Is there anything forcing us to break compatibility with 3.8? My general intuition is that we can keep it around for now, and once we need any features that were...

This isn't some dogma that we must always follow. My point is a bit more from first principles - there's a difference between actively and passively supporting some version. With...

Was this tutorial discussed with anyone, on Discord or otherwise?

I might address some other points later (discussing with Vincent "offline" in parallel), but one thing I want a sanity check on regarding the last message: > So for all...

I'd really, really like to ask people to turn down the temperature and stop overexaggerating. > This is a major problem because this API is essentially eliminating all the benefits...

> seed the action space Done > sync/async Done > termination vs truncation This test specifically checks this one failure case on the simplest scenario, i.e. a fixed-length episode which...

Hi, I don't think there's any comprehensive benchmarking effort, but I have a bunch of curves from a heavily modified* version of the GRPO recipe. Note that all of this...