A new interpretability paper makes a technical discovery. Then it frames the finding in terms of Global Workspace Theory. This framing is completely irrelevant to the findings. To understand this argument, let’s take a quick look at how large language models work. When a large language model processes a sentence, each word or piece of a word – a ‘token’ -- gets converted into a vector, a list of thousands of numbers. This is what allows an LLM to process text based information. That vector doesn’t get replaced as it moves through the model’s layers. Instead, each layer, through two mechanisms called attention and a multilayer perceptron, adds information onto the vector. Attention allows each token’s vector to pull in relevant information from other tokens in the sentence. The perceptron transforms each vector independently using fixed weights, adding conceptually relevant information. Both of these processes happen at each layer, and the whole cycle repeats roughly one hundred times The accumulating vector, carried all the way from input to output, is called the ‘residual stream’. Only at the final layer does the model produce a probability distribution over all possible next tokens, from which the next most likely token is selected (e.g., ‘Paris’ for ‘Is the capital of France’. This Anthropic paper then asks: which specific directions (if any) in the residual stream, in the middle layers of processing, are causally determining that final output distribution? Anthropic developed a measurement technique called the Jacobian lens, or J-lens to investigate this question. It works by taking a specific direction in the residual stream and measuring how adjusting that direction shifts the output probability distribution — not in one sentence, but across many different sentences. If a direction consistently pushes the output toward France-related tokens, for example, regardless of surrounding context — whether the sentence is about geography, food, weather, or anything...