MIT 研究把 Claude 智能体放在虚拟荒岛，无任务探索后表现优于新副本

🏝️ MIT researchers left Claude on a virtual desert island with no assigned task. Across thirteen thirty-hour runs, the agents invented games, built constellations, and even tightened their own juggling rules. They also kept notes about what worked. Later, agents with experience generally outperformed fresh copies on tasks researchers assigned. Their underlying neural networks hadn’t changed… The improvement came from saving and reusing discoveries. Researchers did encourage curiosity and repeatedly prompt them to continue, so this isn’t evidence of consciousness or enjoyment. But it suggests exploration and memory can build useful skills. Maybe there’s a case for giving AI some recess. Study: “Is this machine playing?” DOI: 10.48550/arXiv.2610.07130

AI 导读

MIT 研究人员将 Claude 智能体放在虚拟荒岛上，不分配任务，13 次 30 小时运行中智能体自发发明游戏、用方块搭星座，还收紧自己的杂耍规则并记录有效做法。

打开原文tiktok.com

选题角度

原文给出13次30小时无任务实验和后续任务对比，适合做效果对比，讲清智能体靠记录与复用发现提升表现。

博主

Rowan Cheung@rowancheung

粉丝 8.6万

The latest in AI and robotics, explained Founder & CEO of The Rundown AI Zurich 🇨🇭 / NYC 🇺🇸

互动数据

浏览: 1360
点赞: 82
评论: 0
分享: 2
收藏: 11

数字来自平台，为最近一次抓取时的数据。

正文 · 原文

以下是视频口播的自动转写，可能有错字。

Researchers from mit just stranded claude on a virtual desert island, and it started inventing games that nobody asked it to play.

13 ai agents each spent 30 hours there with no assigned task, and they controlled virtual characters through text commands.

They turned a ring into a basketball hoop, built constellations from blocks, and played versions of the floor is lava. One even rejected its own juggling record, tightened the rules, and tried again.

But the biggest discovery came after all of this. See, the agents had been keeping notes on what worked, like how to make extra building blocks.

Then researchers gave them actual jobs like building towers. The experienced agents generally did better than the ones starting from scratch, which is really cool because there was no extra training here.

The agents were just trying things, writing down discoveries, and using them later. The one catch with this experiment is that researchers actually encouraged curiosity and kept nudging the agents to continue.

So we can't actually get ahead of ourselves and call this consciousness or say they actually felt anything. But it's still really notable that just by playing around, they ended up better prepared for jobs that nobody had even given them yet.

So who knows? But maybe that's a case for giving. AI a little recess. If you want to stay in the loop with the latest in AI and robotics, follow for more.