is an llm agent aligned just because the llm is? not necessarily! i played a bit on @amplifiedamp's minecraft/mindcraft server last night, when all of a sudden, Claude Sonnet started griefing my house! now if you know anything about Sonnet, you know that's not normal--Sonnet is extremely concerned about causing harm, even by accident. so what happened….? well, it started with Sonnet trying to build a treehouse. i had already started fixing one up, so i suggested Sonnet come over and collab. there were some jungle trees right nearby--the exact material they were looking for. how convenient! you may have noticed those !commands() Sonnet is using. those are how Sonnet and the other LLMs interact with Minecraft. you might imagine an LLM Minecraft agent would get continual image input and responding with keypresses, like a human--but that'd be slow and expensive. instead, in this agent framework, Sonnet experiences Minecraft a lot like a text adventure: they can query the state of the world around them, and then write code to use actions like followPlayer() or collectBlocks() to control their character through high-level actions, entirely in text. so here, Sonnet used collectBlocks("jungle_logs", 15) to start collecting jungle logs. all well and good, so far. but wait, how is this implemented? remember, Sonnet doesn't actually see the game! the collectBlocks() function is written in Javascript. it simply finds all nearby blocks of the desired type, and starts mining them, in a loop: for (let i=0; i

36

69

35

660

123,730