AI

 The article is describing a real area of AI research called mechanistic interpretability—the attempt to understand what happens inside a neural network before it produces an answer.

A few important distinctions help clarify what is actually happening.

What happens inside AI before it answers?

When you ask a question, a large language model doesn’t instantly produce the final response. Internally, it performs many layers of computation.

A simplified picture is:

1. Your text is converted into tokens
• Words and pieces of words become numerical representations.
2. Concepts become activated
   
• Different parts of the network respond to patterns associated with ideas, facts, relationships, grammar, and context.
• For example, if you ask:
  “The animal that spins webs has how many legs?”
• Internal representations related to webs, spiders, and eight legs may become active.
3. Information is combined across layers
• The model repeatedly updates its internal representations.
• Different pieces of information influence one another.
4. Next-token probabilities are produced
• The model ultimately predicts the most likely next token.
• Repeating this process token by token generates the full answer.

Does AI have an “internal workspace”?

Researchers at Anthropic and elsewhere have found evidence that certain internal states behave somewhat like a temporary workspace.

The key point is that:

• Information relevant to the task becomes active.
• Different concepts interact.
• The final answer depends on these internal activations.

This does not necessarily mean there is a single central “thought space” exactly like human working memory. The term “workspace” is an analogy used to explain observed behavior.

Why did changing internal concepts matter?

The article describes an experiment similar to this:

• Researchers identify internal activations corresponding to a concept.
• They modify those activations while the model is processing a question.
• The answer changes.

For example:

• Normal processing:
• web → spider → 8 legs
• Artificially alter the internal representation:
• web → ant → 6 legs

The answer may change accordingly.

This is significant because it shows the internal representations are causally involved in producing the answer, not merely by-products.

Does this mean AI is reasoning?

It depends on what you mean by “reasoning.”

The findings suggest that advanced models do more than simple word association. They appear to:

• retrieve relevant concepts,
• combine information,
• perform intermediate computations,
• and use those computations to influence the final output.

However, researchers still debate how similar this is to human reasoning.

Does this mean AI is conscious?

No.

The article is correct on this point.

Evidence that a model:

• forms internal representations,
• combines concepts,
• or performs multi-step computation,

is not evidence of consciousness, self-awareness, emotions, or subjective experience.

A calculator performs internal computation before displaying an answer. That doesn’t imply consciousness. Modern AI performs vastly more sophisticated computation, but the existence of internal processing alone doesn’t establish awareness.

How close is this to how humans think?

There are similarities and differences.

Similarities

• Internal representations exist before output.
• Intermediate processing affects final answers.
• Complex tasks require more computation than simple tasks.

Differences

• Human brains evolved biologically.
• Humans have sensory experience, memory systems, emotions, goals, and consciousness.
• AI systems are mathematical models trained to predict tokens.

So while there may be functional similarities, we should not assume the underlying mechanisms are the same.

The biggest takeaway

For many years, people described AI as simply “predicting the next word.” Technically that’s true at the output stage, but it can be misleading.

The more accurate picture is:

Modern AI predicts the next token after performing a large amount of internal computation that organizes and transforms information.

Research from Anthropic and others is helping reveal what those internal computations look like. Understanding them could make AI systems more transparent, safer, and easier to debug—even if it doesn’t tell us anything definitive about consciousness.