Is AI becoming conscious? Anthropic's latest research opens model "brain"

Is AI becoming conscious? Anthropic's latest research opens model "brain".

In the past two years, large model competitions have mainly focused on two directions: one is to continuously improve model capabilities, and the other is to try to understand why models can perform more and more smartly. How does AI collaborate with different capabilities to complete complex tasks?

On July 7, Beijing time, Anthropic released a latest study, "A Global Workspace in Language Models", which made an enlightening discovery on the interpretability of the model: there is an information organization model similar to the human brain inside the Claude model, based on which "conscious" reasoning can be carried out.

Analogous to the human brain, there are some unconscious information processing processes, and there are also some perceptible and controllable conscious processing processes: such as carefully planning a shopping location.

Neuroscientists and philosophers call the latter brain activity "consciously accessible", which is different from all unconscious processing processes. Researchers discovered a "consciousness access" architecture similar to the human brain inside the Claude model and named it "J-space".

It is much like the "thinking workspace" inside the model, which is responsible for thoughtful reasoning and reportable ideas, independent of automatic language processing.

This opened up the "black box" of large models, allowing researchers to realize that the internal operations of large models are not chaotic, but organized in a way similar to human thinking.

It is worth mentioning that Anthropic emphasized that J-space is not a structure designed or programmed by humans, but was formed by itself during Claude’s training, probably because it is an effective way of organizing calculations.

This shows that the workspace that supports conscious access is not exclusive to the human brain, but appears to be a general solution adopted by intelligent systems to solve certain specific problems. Although J-space plays an important role, it is irrelevant to most features of language models.

In experiments, Anthropic researchers tried to completely delete J-space to verify its effect. The study found that without J-space, Claude was still able to speak fluently, classify emotions, answer multiple-choice questions, and extract facts from articles in much the same way as before.

However, the model was unable to perform tasks that require higher-order thinking: multi-step reasoning dropped to almost zero, and generalization and creative capabilities were lower than a full model on a much smaller scale. The researchers' experiment was inspired by an important theory in neuroscience - "Global Workspace Theory".

This theory depicts the brain as a collection of professional systems, with independent "expert systems" responsible for functions such as vision, hearing, language, and memory.

When people need to complete complex thinking, each system will send information to a shared space, where it will be integrated and then broadcast to other areas to form a unified decision.

Anthropic found that Claude's J-space has particularly tight connections to the rest of its neural network, allowing it to function as the human brain's "broadcast." For example, when a model answers a complex question, different abilities such as language understanding, mathematical reasoning, code knowledge, and world knowledge are not necessarily completed independently of each other.

Instead, information fusion is completed at a few key locations and then drives subsequent calculations. This means that there may be a coordination mechanism similar to a "central workspace" within the model. Does this information mean that AI has consciousness? This is the most likely place for misinterpretation after the release of this study.

Finally, Anthropic believes that large models have a consciously accessible architecture, which is a similarity in computational structure and does not mean that the model has feelings and consciousness in the psychological sense. If this research direction continues to advance, its future impact may extend beyond academia.

This research provides the possibility to monitor the "inner thoughts" of the model. A deeper understanding of the calculation process inside the model is expected to help researchers more accurately locate the causes of hallucinations, erroneous reasoning or unsafe behaviors in the model, thus improving the reliability of AI systems.

In addition, interpretability capabilities may also become a new dimension of future model competition. At present, the large model industry has entered a stage of performance convergence, and companies are increasingly required to prove that models are not only more capable, but also more transparent, controllable and trustworthy.

This research may also influence future model architecture design. If J-space is indeed an important part of complex reasoning, then large models in the future may have new optimizations around information integration mechanisms, rather than just relying on expanding parameter sizes or increasing training data. (

#Stocks #AI