I tried out 'J-Lens,' which can visualize the 'J-space' that resembles AI thought processes.
While the performance of AI models continues to improve rapidly, the processing steps that lead AI to determine its output are not yet fully understood. In this context, Anthropic, the developer of Claude, has proposed the concept of '
Jacobian Lens – Qwen3.6-27B ｜ Neuronpedia
https://www.neuronpedia.org/qwen3.6-27b/jlens
When humans consciously process things, brain activity patterns called 'consciously accessible activities' appear. Similarly, it has been discovered that AI models such as Claude also have a 'set of a few neural patterns that play a special role compared to other internal processes.' Anthropic has named this set 'J-space.'
It has been discovered that a 'J-space' specifically designed for 'thinking' is activated when AI processes instructions - GIGAZINE
A demo of J-lens, which visualizes J-space, is available at the following link, so let's try using it.
Jacobian Lens – Qwen3.6-27B ｜ Neuronpedia
https://www.neuronpedia.org/qwen3.6-27b/jlens
Here's what the demo looks like. When you chat with the AI in the chat area on the left, 'phrases activated in J-space' will be displayed on the right. You can choose from AI models such as Qwen3.6 17B and Gemma 3 12B.
I typed 'Tell me how to choose a smartphone' and sent it.
The AI's response was displayed on the left, and the phrases activated in J-Space were displayed on the right.
The most frequently activated phrase while generating responses was 'smartphone,' followed by phrases such as 'smartphones,' 'iPhone,' and 'affordable.'
It's also possible to see the phrases activated in the J-space before outputting the 'next token' for each token. When outputting the next token after 'smart,' phrases like 'fo' and 'phones' were activated.
When outputting the token following 'latest,' phrases such as 'model,' 'of,' and 'standard' are activated, ultimately outputting the token 'model.'
It should be noted that J-space was not considered during the AI model's design phase, but rather arose naturally during the AI model's learning process. However, the existence of J-space does not necessarily indicate that the AI model has the same thought processes or emotions as a human.
Related Posts:
