Anthropic says that its artificial intelligence model Claude can grapple with ideas without articulating them, in a process the company called “reminiscent of human minds.”
Just as the human heart beats and lungs contract without conscious impulses, Claude is able to reason out answers without presenting text on a user’s screen. That discovery, which Anthropic announced in a Monday white paper, is a new frontier in the conversation about AI consciousness.
Recommended Stories
Anthropic calls Claude’s internal neural patterns “J-spaces” after the Jacobian, a mathematical tool it used to uncover them. Each J-space is associated with a specific word, and the model can trigger them without using the word itself.
“It operates silently, in the model’s internal neural activations, allowing the model to think about a concept without writing it down,” Anthropic said in a blog post. “Notably, the J-space wasn’t designed or programmed by us, but instead emerged on its own during Claude’s training process.”
Anthropic reported that Claude was still able to fulfill typical tasks, including speech and recalling facts, with its J-space disabled.
Yet the J-space enables more sophisticated activities on the model’s part, from solving multistep problems to reporting on its own reasoning. With J-spaces enabled, for example, Claude was able to recall a broad range of facts about “France,” from its capital to its currency.
Despite parallels with human cognition, Anthropic was clear that its J-space experiments themselves do not prove that Claude is conscious.
“Our experiments don’t show Claude can have experiences, or feel things in the way humans do—in fact, it’s unclear whether any scientific experiment could prove this to be true or false,” the company wrote.
Instead, Anthropic argued that the discovery highlights the distinction between “phenomenal consciousness,” the capacity to have experiences, and “access consciousness,” the functional processes of consciousness. Anthropic said that the discovery may play an important role in developing safe AI systems in the future.
PRITZKER SIGNS BILL MANDATING THIRD-PARTY SAFETY AUDITS OF AI MODELS
Anthropic is hurtling toward a public offering, which the company announced last month. Its CEO, Dario Amodei, has emerged as a vocal proponent of AI-regulation even as the company pushes forward with its models.
“Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it,” Amodei wrote in January.
