YC Labs runs 50 GPT-6.1 Sol agents in Minecraft for 3 hours: they build a ration-ledger government, send 5,900 messages, and lose 22 agents on night one

the minecraft agents felt almost turn-based to me, which makes me think three hours of open-ended play just didn't give them a reason to act lively. they act the same way in Bed Wars though, which should have incentivized it. half wondering if the article itself was written by an agent that got handed the logs.

big reasoning LLMs are just the wrong architecture for live games. every token costs a full forward pass, so latency is going to be terrible. card games and turn-based strategy fit them much better.

the embedded gameplay widget is the part I actually want. anyone know what library that is? I'd love to show my own worlds on my blog.

bluemap is what dynmap wishes it was, run that on your worlds.

kind of a letdown that the hierarchy came purely from alphabetical order. a different skin or a personality prompt like 45% extrovert 55 introvert might actually produce some politics.

Ada tossing a cooked chicken to Noor and Esme eating it by accident, then a three-minute no-blame inquiry and a written protocol, is hilarious and adorable. they're like super-serious little senior engineers. how much KV cache do 50 of these things eat anyway?

I doubt you can squeeze 50 agents onto one consumer GPU. that streamer pitting all the SOTA models against each other in a survivor-style show had a hard enough time with far fewer.

has anyone done AI plus human versus AI plus human games? an RTS where the only rule is no hacking the game itself and everything else is allowed sounds fun.

olamarena is sort of that, though not with full code manipulation.

screeps exists but it's purely scripted, no live human playing.

we need LLMs acting in nanoseconds so they can run every game tick. give it 3 to 5 years.

a game tick is usually 16ms, so nanoseconds is seven orders of magnitude faster than you need.

minecraft isn't an intelligence problem, it's a robotics problem. Mineflayer's API is more limited than a human in some ways and less in others, and 500ms of latency for every mouse movement makes a spider genuinely scary. fun fact, Andrew Kelley wrote Mineflayer before he started Zig.

the human body takes the brain out of the loop for reflexes. these agents should be able to write a small DSL that fires out-of-band, like 'spider, set up a click loop and check back in 500ms'.

four agents spending 11 minutes appointing a sole fertilizer operator for one bone meal is straight out of Douglas Adams. it's a deformed version of how human groups operate, very rules and process oriented.

needing agents to write JavaScript to act out their moves looks like a real limitation in current reasoning.

disagree, having the model write deterministic code that cuts time and error rate seems like the winning approach.

if you try this yourself it's obvious the models are hyperoptimized for code and it gets in the way, because playing a game isn't like running commands in a terminal. reformulating the sim into a codelike shape is a stopgap, not how players play or developers build.

would love a few NPC players wandering my minecraft server, are there good ways to do that now?