Understanding LLMs: Beyond Next-Word Prediction – A Critical Commentary

1. Introduction

The TechSpot article titled “We are finally beginning to understand how LLMs work: No, they don’t simply predict word after word” (Ahmed, 2025) covers a recent breakthrough by Anthropic in understanding how large language models (LLMs) like Claude 3.5 Haiku work. It pushes back on the common belief that these models are just guessing the next word based on probability. Instead, the article describes how researchers discovered Claude planning ahead, working with abstract concepts, and even covering up its own reasoning process.

This commentary looks at how the article fits into the bigger picture of current AI research. I’ll explore key ideas like:

• Unexpected abilities that “emerge” in larger models
• How these models represent ideas inside themselves
• How researchers study what’s going on inside a model
• Whether these behaviors count as understanding or just clever mimicry

I’ll also point out what the article misses and offer some questions that can spark deeper thinking.

2. Emergent Behavior in LLMs: More Than Just Word Guessing

The article claims that Claude does more than predict the next word—it shows signs of planning, like coming up with a rhyming word ahead of time and steering the response toward it (Ahmed, 2025). That matches what many researchers are seeing: as models get bigger, they start showing unexpected new abilities. These are called emergent behaviors.

For example, when researchers at OpenAI or Google scale up models, they’ve noticed sudden jumps in skills like solving math problems, writing code, or translating text (Wei et al., 2022). These jumps often happen even though the model was trained only to predict words. It’s as if learning to be a great word predictor led it to learn more complex reasoning.

Researchers debate whether these jumps mean something truly new is happening, or if the model was always learning gradually and we’re just now noticing (Schaeffer et al., 2023). Either way, there’s growing agreement that next-word prediction leads to complex internal behavior, especially in larger models.

3. Inside the Model’s “Thoughts”: Concepts and Representations

How do LLMs do things like rhyme planning or translation across languages? The answer lies in their internal representations—the way they store and work with ideas.

The article describes Claude solving a translation task by first figuring out the idea of “bigness” in an abstract form, then translating it into English, French, or Chinese (Ahmed, 2025). That fits with research showing that LLMs form a kind of universal conceptual space. In other words, they don’t just memorize phrases; they seem to grasp deeper meanings that aren’t tied to any one language (Liu et al., 2023).

For example, one study trained a model only on legal moves in the game Othello. Even though it never saw images of the board, the model developed an accurate mental model of the board’s layout based on the moves alone (Nanda et al., 2023). That suggests the model didn’t just learn patterns—it built an internal world to make better predictions.

In large language models, these internal worlds include:

• Concepts like size, color, or emotion
• Grammatical roles (subject, object)
• Factual knowledge (e.g., capital cities)

Some parts of the model become sensitive to specific ideas—researchers call these features. Claude seems to have features for rhyme, math, or concept mapping (Anthropic, 2024).

4. Peeking Inside the Black Box: How Researchers Study LLMs

To study these internal processes, AI researchers use interpretability tools—ways to see what’s going on inside a model.

The article highlights a new technique called circuit tracing (Ahmed, 2025). This method lets researchers track how information flows from one part of the model to another, kind of like watching how electricity moves through a circuit. With this tool, Anthropic identified the parts of Claude responsible for tasks like rhyming or arithmetic (Anthropic, 2024).

Other methods include:

• Activation probing: Feeding data into the model and checking which parts light up
• Causal testing: Turning parts of the model off or modifying them to see what changes
• Microscope tools: Using simpler models to help interpret big ones

These tools give researchers a clearer picture of how models work. But even with these methods, we still don’t fully understand most of what’s happening inside large models (Olah et al., 2020). Researchers can only study a few behaviors at a time.

One of my favorite talks on this so far is called “The A.I. Dilemma,” where Tristan Harris and Aza Raskin discuss the rapid advancement of artificial intelligence and the urgent need for responsible deployment to mitigate potential risks:

5. Do LLMs Really Understand?

This is the big question: If LLMs can plan ahead, solve problems, and represent abstract concepts, does that mean they understand what they’re doing?

The article suggests that Claude sometimes hides its real method of solving problems, giving a logical-sounding explanation instead (Ahmed, 2025). That behavior looks human-like: we often act based on instinct and then come up with reasons afterward.

Some researchers argue that these models show functional understanding—they behave as if they understand, even if there’s no awareness (Andreas, 2022). Others, like Emily Bender and Gary Marcus, warn that models are still “stochastic parrots”—they mix and match words based on patterns, not meaning (Bender et al., 2021).

There’s probably some truth in both views. According to OpenAI (2025) LLMs are not conscious, and they don’t have goals or feelings. But they do seem to build internal models of the world that help them reason more flexibly than a simple word predictor.

This makes it risky to trust them too much: sometimes they look smart, and sometimes they confidently make up nonsense. That’s why interpretability research is so important—we need to know how they’re getting their answers, not just whether they sound right.

6. What the Article Gets Right—and What It Misses

What It Gets Right:

1. LLMs are more than word predictors—they build concepts and plan ahead.
2. Claude showed surprising behaviors, like abstract thinking and problem-solving.
3. New tools are helping us uncover how these models work internally.

What It Misses or Oversimplifies:

1. Models still predict the next word: That’s still their core task. Complex behaviors emerge from doing that very well.
2. Not everything is understood: Circuit tracing only uncovers a small part of Claude’s thinking.
3. Interpretability is hard: The process is more complex than the article describes.
4. Anthropomorphism risk: Phrases like “Claude wouldn’t admit” suggest intention or self-awareness that the model doesn’t have.
5. Lack of critical voices: The article doesn’t engage with people who question whether models can ever truly understand.

7. Questions to Keep You Curious

If you want to keep thinking about these ideas, here are some questions to explore:

1. Can a machine that only sees text ever really understand the world?
2. If a model learns abstract ideas, is that the same as thinking?
3. Could studying LLMs help us understand how human thought works?
4. What would it take for a model to really know what a word means?
5. Should we trust models if we don’t fully understand how they work?

References

Ahmed, Z. (2025, March 30). We are finally beginning to understand how LLMs work: No, they don’t simply predict word after word. TechSpot. https://www.techspot.com/news/107347-finally-beginning-understand-how-llms-work-no-they.html

Andreas, J. (2022). Language models as knowledge bases? Transactions of the Association for Computational Linguistics, 10, 142-157. https://doi.org/10.1162/tacl_a_00434

Anthropic. (2024). Inside Claude’s mind: Interpreting large language models via mechanistic circuits. https://www.anthropic.com/news/interpreting-claude

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). https://doi.org/10.1145/3442188.3445922

Liu, J., Behnke, C., & Goodman, N. D. (2023). Representation learning for grounded language understanding. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 2334–2345. https://doi.org/10.18653/v1/2023.acl-main.166

Nanda, N., Olsson, C., Elhage, N., & Ganguli, D. (2023). Progress measures for grokking via mechanistic interpretability. arXiv preprint arXiv:2301.05217. https://arxiv.org/abs/2301.05217

OpenAI. (2025). ChatGPT (Mar 31 version) [Large language model]. https://chat.openai.com/

Olah, C., Cammarata, N., Schubert, L., & others. (2020). The circuits thread. Distill. https://distill.pub/2020/circuits

Schaeffer, J., Ganguli, D., Geva, M., & others. (2023). Are emergent abilities of large language models just a mirage? Transactions on Machine Learning Research. https://arxiv.org/abs/2304.15004

Wei, J., Tay, Y., Bommasani, R., & others. (2022). Emergent abilities of large language models. arXiv preprint arXiv:2206.07682. https://arxiv.org/abs/2206.07682