michalsustr

Correct. Erlang also uses green threads?

Yes indeed, unless you use ropes or other specialised structures

I’m not familiar with Haskell concurrency. The combination of green threads and large memory allocations due to immutable data structures sounds like it would be hard to implement a web server handling 10k+ concurrent requests on commodity hardware?

Btw. too bad author talks about microsecond guarantees usage but does not provide a link, that would be interesting reading.

I recently attended a STAC conference where they claimed the GPUs themselves are not deterministic. The hand-wavy speculation is they need to temperature control the cores and the flop ops may be reordered during that process. (By temperature I mean physical temperature, not some nn sampling parameter). On such large scale of computation these small differences can show up in the actually different tokens.

The blurring in the sheets and the topo loss reminded me of https://arxiv.org/abs/2408.05446

This is why we use squash like here https://docs.gitlab.com/ee/user/project/merge_requests/squas...

My hunch is that this effectively creates a differentiable minimax “search” “tree” that can be backpropagated through. Not a tree — a dag really — and not search, but learning. :)

> the fastest, heavily-optimised Java solution runs in just under a second on my machine

I don’t understand how this is possible. The file in question has 13GB, while the fastest commonly available SSDs are 12400 MB/s. Am I missing something?

Thank you!

Can you please give a link to the soviet paper? Thank you very much :)

Energy and mass are equivalent and their relation is E=mc^2

Why would one prefer to add sudoedit X to sudoers rather than updating file access privileges of X directly?

Just curious about arguments for this use case.

Great question! I would say the main reason why this work is significant is that StarCraft agent was bootstrapped from human replays, and Dota agent did not have game-theoretic guarantees on minimizing exploitability (i.e. how far the strategy is from Nash equilibrium). The R-NaD algorithm with neural nets (behind Stratego) starts from scratch, and has game-theoretic guarantees.

In principle, the AlphaStar's league approach (from StarCraft) could be done also in Stratego, and it would be very interesting to compare the two approaches. Note that AlphaStar is more expensive: it required to train N competing agents with pair-wise evaluation costing N^2, while Stratego's NeuRD trains a single agent.

Interesting! We have very recently wrote a paper on the various tree representations [1]. Maybe it is the same distinction you found on your own? I judge this based on the screenshots on your website. Public state trees are amenable to nice factorization over the hands, and indeed most implementations use it, but formally they write they use the history tree notation.

[1] https://arxiv.org/abs/2112.10890

Congrats on putting this together, it's not easy! I'm curious, what papers have you used for your implementation? I know there are a couple of reimplementations of DeepStack out there.

Disclaimer: I'm doing phd in this area, generalizing to harder games than poker.