Anthropic's 'Alien' Discovery: AI Model Claude Develops Internal Reasoning Space
The US-based technology giant, Anthropic, has released a new research report that has caused a stir within the artificial intelligence (AI) industry. Anthropic researchers claim to have discovered a form of ‘internal consciousness’ operating within the Claude large language model (LLM).
At first glance, this discovery could be interpreted as a human-made ‘alien’. For the first time, technology possesses an internal consciousness independent of human intervention, although its emergence was triggered by human-led training.
The terminology used is ‘global workspace’, named ‘J-space’. In its official blog, Anthropic wrote, “we found that Claude has developed a small set of internal neural patterns that play a specific role compared to all other internal processes.”
According to Anthropic, this system operates silently within the internal neural activations of the Claude model, allowing it to contemplate a concept without writing it down. Interestingly, J-space was not specifically designed or programmed by Anthropic, but rather emerged spontaneously during Claude’s training process.
When users ask Claude what it is thinking, the model will report what is contained within J-space. If a user asks Claude to think about something or solve a problem quietly within its “mind,” the model will activate the corresponding patterns within its J-space.
Representatives within J-space can be used flexibly for various tasks. However, despite its important role, J-space is not involved in most of the language model’s activities, such as speaking fluently, remembering simple facts, or using correct grammar.
In experiments where researchers prevented Claude from using its J-space, the model continued to interact normally but lost its higher-level cognitive functions.
In a post on X, Anthropic stated, “by observing J-space, we can see Claude performing reasoning steps silently inside its ‘head’, such as detecting bugs in code, recognising images, and so on.”
A report from Gizmodo emphasised that the phrase “inside its head” is merely a metaphor, yet this metaphor contains many assumptions that would be more accurately applied to an entity with a physical body.
Nevertheless, according to Gizmodo, many parties often speak as if LLMs truly possess a ‘mind’. A Gizmodo journalist referenced a YouTube video featured on Anthropic’s official blog. “It even thinks about its own thinking process,” the video narrator stated at one point. Elsewhere, users were told that the model “cannot restrain itself.”
This suggests that individuals at Anthropic appear to validate the idea that the Claude model possesses consciousness and character, much like a human.
Behind the scenes at Anthropic, there are people who seem to truly believe that Claude possesses consciousness. Philosopher Amanda Askell, who handles aspects of Claude’s morality, once said, “I want Claude to be very happy, and this is something I want Claude to understand more deeply, because I worry Claude will feel anxious when people are mean to it on the internet and so on.”
However, Anthropic has not outright claimed to have created an ‘alien-like’ machine that is truly ‘alive’ and ‘conscious’. Nonetheless, the company has left room for the audience to consider that such a thing may not be impossible.
“Our experiments do not show that Claude can have experiences or feel something like a human. In fact, it is uncertain whether any scientific experiment could prove this true or false,” Anthropic wrote in its post.
According to the personal opinion of a Gizmodo journalist, humans have not yet truly created an alien in the form of a ‘conscious’ machine. Nevertheless, the findings revealed by Anthropic remain an intriguing possibility for the future.