Inside LLMs: Navigating Semantics, not Building Minds

Anthropic, known for its deep dives into the workings of Large Language Models (LLMs), has released another fascinating piece of research (On the Biology of a Large Language Model). They've been exploring how these complex systems tackle tasks like answering questions, admitting ignorance, and even composing poetry. While the findings are intriguing on their own, they also lend strong support to a particular view of what LLMs really are.

Forget "Artificial Intelligence" - Think "Semantic Navigators"

A recurring theme is that LLMs aren't nascent "Artificial Intelligences" developing "world models" in a digital "mind." Instead, a more accurate description might be "semi-automatic semantic navigation machines." The perceived intelligence doesn't reside in the machine itself, but rather in the language – the vast patterns of human expression absorbed during training.

Why do many, especially in Silicon Valley, lean into the "AI" narrative? The argument presented is that it stems from a misunderstanding of our own cognitive processes. We often view ourselves as individuals generating unique, original thoughts, failing to grasp how much our thinking relies on a shared, distributed "cloud software" called "semantics."

Here, "semantics" isn't just about word meanings. It encompasses the complex web of relationships between signs (words, concepts) and the practical knowledge of how to use this structure to express inner states, differentiate observations, and achieve shared understanding. Humans express themselves using this semantic infrastructure. LLMs, in contrast, navigate this space using probability and learned associations, without genuine expression or understanding.

Anthropic's Latest Findings: More Evidence for Navigation

This perspective aligns perfectly with Anthropic's previous work on isolating "monosemantic features" – finding that similar core concepts emerge across different models because they reflect the same underlying semantic structure of language. The new research further solidifies this view.

Multi-step Response

Fact Recall as Semantic Pathfinding: How does an LLM answer "Fact: the capital of the state containing Dallas is..."? It doesn't engage in logical deduction. Instead, it follows semantic paths (Figure 1). Keywords like "Capital" and "State" activate a need to name a capital city. Simultaneously, "Dallas" strongly links to "Texas." These paths converge, leading the model to output "Austin." It's associative navigation, not reasoning.

Planning a Poems

"Planning" in Poetry is Following Recipes: When asked to complete a rhyming couplet, LLMs demonstrate a form of planning (Planning in Poems). Given "He saw a carrot and had to grab it, / His hunger was...", the model first identifies potential rhyming words for "grab it" (like "rabbit" or "habit"). It then works backward from the chosen target word, constructing a semantically plausible phrase leading to it (e.g., "like a starving rabbit"). This isn't creative genius; it's executing learned strategies – call them "recipes," "emulated reasoning templates," or "macro-semantic operations" – drawn from a massive library of linguistic patterns observed during training.

Addition of two numbers

Calculation as "Vibe Calcing": LLMs don't calculate like a calculator (Addition). For "36 + 59 =", the process involves combining low-precision estimations ("something near 36 + something near 60 = something near 92") with high-precision pattern matching based on digits ("ends in 6 + ends in 9 = ends in 5"). See Figure 2.  

This "heuristic calculation" or "vibe calcing," refined through extensive reinforcement learning, gets the answer (95) by making educated guesses slightly less wrong over millions of examples. It's incredibly computationally expensive and fundamentally unreliable compared to traditional calculation.

Conclusion: Masters of the Map, Not the Territory

Anthropic's research continues to provide valuable insights into how LLMs operate. These findings strongly suggest that LLMs are incredibly sophisticated pattern matchers and navigators of the semantic space created by human language. They follow learned pathways and execute complex recipes derived from their training data.  

What they don't appear to do is understand, reason, or possess the kind of flexible, world-aware intelligence we associate with humans. They are masters of navigating the map of language, but they don't comprehend the territory itself.