can you say more about "If the same algorithm couldn't be adapted to other environments with similar results, it doesn't count."?
if its given 1000 words of guidance on how to play minecraft, but otherwise the method can easily be ported to other games, does that count?

@calour Yes, that counts. This market is (more-or-less) older than consumer LLMs; it was written with classic RL agents in mind. An LLM with a fixed set of weights can have as much in-prompt guidance as necessary as long as the approach would generalize.

@vluzko is it intentional for 2026 to be in both ranges?

@2b3o4o It's [inclusive, exclusive). Not a great way to write it in hindsight, sorry

Does this mean starring at the beginning of a new survival game on non-peaceful difficulty? Or does it include peaceful/creative mode/starting in the end with a bow?

@maskedtorah I'm not going to resolve to anything but the options I originally made.

@vluzko Apologies, I missed that the displayed list was truncated at first. Would delete if I could in the UI.

"Before 2023" seems very optimistic for an agent not tuned to Minecraft specifically, considering to my knowledge no agent tuned to Minecraft has done it yet. This sounds like it would require multi-domain training, which as of right now has comparable or worse results to specialization.

@horse I'm interpreting not tuned for minecraft as "could be used for another game if you trained it on it".
So my guess is that If someone makes something like a bigger VPT but trained on speedruning videos it would count.

@VictorLevoso Yes, that is what "not tuned for minecraft" means here: the agent itself can be tuned to minecraft, of course, but the learning algorithm generating it can't be