论文88°

Anthropic发现Claude内部推理空间J-Space,可读可操控

Must-read research by Anthropic. Here is the simple explanation and why this is a big deal. We s...

精选理由

Anthropic找到了Claude思考时的内部工作区J-Space,能看能改,比以往猜输出靠谱多了。

AI 摘要

Anthropic新研究发现Claude内部存在一个名为J-Space的全局工作空间,与链式推理或草稿板不同,它是训练中自发出现的推理机制。通过J-Space,研究者可以直接读取和修改模型内部的信息流动。实验表明,该方法首次实现了对模型推理过程的直接观测和操控,而非仅从输出文本推测。这为可解释性提供了新工具,可能用于验证模型行为、实施更好护栏和预测危险场景。

原文 · elvis

Must-read research by Anthropic. Here is the simple explanation and why this is a big deal. We s...

Must-read research by Anthropic. Here is the simple explanation and why this is a big deal. We suspect LLMs perform "internal reasoning". But little is known or do good methods exist to understand it. Anthropic claims that J-Space (which differs from chain-of-thought or scratchpad), emerged on its own through training and provides a window into how Claude "reasons" internally. In other words, this shows that Claude has a sort of internal workspace where information gets held, combined, and passed between different parts of the model. They can read from it, and they can steer the model by changing it. As it is the case with these reports, the consciousness angle will get all the attention. However, the bigger story is that for the first time you can point to a specific place inside the model where reasoning is staged, rather than guessing at it from the text that comes out. This, of course, changes what interpretability can be. We spent years inferring what a model was doing from what it said. Now there's a mechanism to observe directly, and a direct lever to move. This could enable even more advanced levels of "reasoning" in LLMs and bridges gaps in frontier intelligence and world models. If you can see where a model holds an idea, you can also verify it, audit it, and catch it working toward a goal you never gave it. You can implement better guardrails and predict dangerous/unwanted scenarios better. Anthropic @AnthropicAI New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude. Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 2 🔄 1 ❤️ 12 👀 1396 📊 3 ⚡