https://rlplays.com/game

This is NOT a quick-one-off vibe/LLM-coded project.

I started this project out of curiosity: I wanted to build an RL-based game as there were very few out there (e.g. Sony GT Sophy). And I wanted to learn the core RL foundation in a practical/useful manner.

I built on top of Puffer - but the training speed was not up to my needs so I rewrote the core with a ground-up native eval/training loop with multithreaded GPU batching (gonna be a part of the next Puffer release). [ Unaffiliated plug: Puffer is an excellent OSS library - check out https://puffer.ai ]

I trained the RL agent using curriculum learning + self-play. The demo showcases this self-play as well - which you can play against yourself, like an RL agent would!

Technical details in my blog in the link above.

Comments URL: https://news.ycombinator.com/item?id=47436311

Points: 1

# Comments: 0

I am just saying it's not as flexible/cost-free as you would on a 'normal' von Neumann-style CPU.

I would love to see Rust-based code that obviates the need to write CUDA kernels (including compiling to different architectures). It feels icky to use/introduce things like async/await in the context of a GPU programming model which is very different from a traditional Rust programming model.

You still have to worry about different architectures and the streaming nature at the end of the day.

I am very interested in this topic, so I am curious to learn how the latest GPUs help manage this divergence problem.

I understand with newer GPUs, you have clever partitioning / pipelining in such a way block A takes branch A vs block B that takes branch B with sync/barrier essentially relying on some smart 'oracle' to schedule these in a way that still fits in the SIMT model.

It still doesn't feel Turing complete to me. Is there an nvidia doc you can refer me to?

GPUs are still not practically-Turing-complete in the sense that there are strict restrictions on loops/goto/IO/waiting (there are a bunch of band-aids to make it pretend it's not a functional programming model).

So I am not sure retrofitting a Ferrari to cosplay an Amazon delivery van is useful other than for tech showcase?

Good tech showcase though :)

Paperbacks required authors to spend the same amount of time/effort to create content with a vastly expanded market and distribution mechanism.

TikTok and Insta created N creators to M consumers where N is nearly the same as M. Making the distribution channels bigger but effortless to create content doesn’t magically equate quality paperbacks with short form hummingbird-attention videos.

Being an engineer and coding at this stage/level is just remarkable- sadly this trade craft is missing in most (big?) companies as you get promoted away into oblivion.

Not sure how apocryphal a tale this is but it does speak volumes to how customer obsessed these companies were.

**insert inception meme here**

Joking aside: this points to MSFT moving away from the whole Mono/Maui investments and into Aspire or whatever they call it. Without MSFT backing this I am not sure if there is much more future left for MAUI (or dotnet on mobile in general).

Avalonia is great though.

Don't be silly and assume if I assign it multiple times in an if condition it's mutable - it's constructing the object as we speak, so it's still const!!!

C# gets this right among many other things (readonly vs const, init properties, records to allow immutability by default).

And the funny thing is the X thread has lots of genuine comments like 'yeah, just wrap a lambda to ensure const correctness' like that's the okay option here? The language is bad to a point it forces good sane people into seeing weird "clever" patterns all the time in some sort of an highest ELO rating for the cleverest evilest C++ "solution".

I was hoping Carbon was the hail mary for saving C++ from itself. But alas, looks like it might be googlified and reorged to oblivion?

Having said that, I still like C++ as a "constrained C++" language (avoid clever stuff) as it's still pretty good and close to metal.

Software (iOS26), services (Music/Tv/Cloud/Apple Intelligence) and marketing (just keep screaming Apple Intelligence for 3 months and then scream Liquid Glass) ---- on the other hand seem like they are losing steam or very reactive.

No wonder John Ternus is the widely anticipated to replace Tim Cook (and not Craig).

If you could yank the screen out, it probably evens out :)

I have seen quite a few such announcements from competitors that tend to be so close that I wonder if they have some competitor analysis to precede the Goliath by a few days (like Google vs rest, Apple vs rest etc).

And they still are alongside us right now. Which to me is fascinating.

Perhaps `dot`, `cross` etc might be useful additions.

Seriously though this is inevitable as Fuchsia wound down and Android essentially proved its worth running on Windows (which was a nice experiment I guess).

It may be hard to find an audience (revenue) but these kind of gems should be encouraged similar to traditional art form.

It is somewhat obscene that there is art in galleries out there with far less effort and talent that go for millions in both funding and revenue.