In a recent experiment on Claude Sonnet 4.5, a large language model (LLM) from Anthropic, researchers asked it to count to five and, at the same time, “introspect(内省，自我反省) deeply”. The artificial-intelligence model complied, returning: “One . Two . Three . Four . Five .” So far, so normal(到目前为止，一切正常).
在最近一项针对Anthropic公司的大型语言模型（LLM）Claude Sonnet 4.5的实验中，研究人员要求它数到五，同时“深入内省”。这个人工智能模型照做了，回复道：“一。二。三。四。五。”到目前为止，一切正常。
During the task, the researchers were watching what was going on inside the LLM’s many layers of artificial neural networks. When a user asks a question, the words are turned into chunks of text (tokens(令牌，文本单元)) that are then converted to numbers. These are then passed through the layers of artificial neurons until they reach the final layer, which produces a token. Repeat that process a bunch of times each second and you get a series of tokens that turn into a sentence or some other more complex response.(每秒重复这个过程多次，你就能得到一系列token，它们会变成句子或其他更复杂的回应。) For a long time these layers have remained a black box(黑匣子，不透明系统), largely impenetrable(不可穿透的，难以理解的) to anyone wanting to know how or why LLMs do things in the way they do.
在任务期间，研究人员观察着LLM多层人工神经网络内部发生的情况。当用户提问时，单词被转换成文本块（token），然后转换为数字。这些数字再经过人工神经元层，直到到达产生token的最终层。每秒重复这个过程多次，你就能得到一系列token，它们会变成句子或其他更复杂的回应。长期以来，这些层一直是一个黑匣子，对于任何想了解LLM为何以及如何以这种方式运作的人来说，基本上是无法窥探的。
Anthropic’s researchers were able to look inside those layers, using a mathematical tool they had developed to do so. What they found surprised them.(他们的发现令他们惊讶。) As the model counted the sequence of numbers, different words popped in and out of existence(忽隐忽现，出现又消失) in the layers underneath. One was “countdown”. About halfway through the output, “half way” appeared. Then the words “consciousness”, “AI” and “Claude” all appeared. After the model spat out the number five, it said nothing else to the researchers before issuing a full stop, but the word “done” appeared in its neural layers.(在模型吐出数字五之后，它在发出句号之前没有再对研究人员说什么，但“完成”这个词出现在它的神经层中。)
Anthropic的研究人员利用他们开发的一种数学工具，得以窥视这些层的内部。他们的发现令他们惊讶。当模型数数字序列时，不同的单词在其底层中忽隐忽现。一个是“倒计时”。在输出进行到大约一半时，“中途”出现了。然后，“意识”、“AI”和“Claude”这些词都出现了。在模型吐出数字五之后，它在发出句号之前没有再对研究人员说什么，但“完成”这个词出现在它的神经层中。
Read the rest of our cover package
阅读我们封面报道的其余部分
Intel inside
英特尔内置
Think different
不同思考
Don’t be evil
不要作恶