ai
Anthropic Reports Internal Neural Patterns in Claude AI Model
Anthropic published a white paper describing internal neural patterns in its Claude model that allow processing without generating visible text. The patterns emerged during training and support multistep reasoning while remaining optional for basic tasks.
Jul 7, 1:47 PM·1m read2 sources
Washington ExaminerAudio version
Tap play to generate a narrated version.
Developing·Limited corroboration so far. This page will refresh as more sources emerge.
Anthropic announced that its Claude artificial intelligence model can process concepts internally without producing output text. The company described the activity as occurring in internal neural activations that form during training.
Anthropic labeled the patterns J-spaces after the mathematical Jacobian used to identify them. Each J-space links to a specific word, and the model can activate the pattern without writing the word. The company stated the spaces were not programmed but appeared on their own.
The model continued to perform standard functions such as speech and fact recall when the J-spaces were disabled. With the spaces active, Claude handled multistep problem solving and reported on its own reasoning steps.
Anthropic stated that the experiments do not demonstrate consciousness or the capacity for experiences. The company distinguished between phenomenal consciousness and access consciousness, noting that the findings relate only to functional processes.
The company said the discovery could inform future safety work on artificial intelligence systems. Anthropic also reported that it is preparing for a public offering announced last month.
