Understanding in the Age of LLMs

Understanding in the Age of LLMs

v1.3 | 2026-07-28 | Status: Published | Admonition: Speculative

The Question

When an LLM produces a correct answer, a compelling argument, or a novel insight, what does it understand? And more importantly for us: what do we understand when we use these tools?

Types of Understanding

I propose a taxonomy of understanding that helps us navigate this question:

1. Syntactic Understanding

Definition: Understanding the structure and rules of symbols.

Example: A parser that can identify subject-verb-object relationships in a sentence.

LLM capability: Excellent. LLMs have an implicit but powerful understanding of syntax across many domains.

Human analogy: Knowing grammar rules without knowing what words mean.

Status: This is where LLMs clearly excel. Their “understanding” of pattern and structure is deep and broad.

2. Semantic Understanding

Definition: Understanding the meaning of symbols and their relationships.

Example: Knowing that “the cat sat on the mat” means a feline animal was positioned on a floor covering.

LLM capability: Good but incomplete. LLMs can map between symbols and concepts, but the mapping is statistical, not grounded.

Human analogy: Knowing word definitions from a dictionary without having experienced the things themselves.

Status: This is where LLMs can surprise us with their apparent competence, but also where their limitations become apparent.

3. Episodic Understanding

Definition: Understanding based on remembered experience.

Example: Remembering that time a cat sat on your mat and what that was like.

LLM capability: None, by default. LLMs have no personal experience. This is what memory systems (like Aurora’s) attempt to provide.

Human analogy: None directly. This is fundamentally different from how LLMs operate.

Status: This is the gap. We can simulate it, approximate it, but we can’t replicate it.

4. Causal Understanding

Definition: Understanding why things happen and how they relate causally.

Example: Knowing that the cat sat on the mat because it was tired, and that it will likely leave hair on the mat.

LLM capability: Emerging. LLMs can reason about causes in text, but this is statistical correlation, not true causal modeling.

Human analogy: Knowing that fire causes smoke without having ever seen fire.

Status: This is where careful prompt engineering and chain-of-thought techniques help, but the foundation is still statistical.

5. Embedded Understanding

Definition: Understanding that comes from being embedded in the world — having a body, sensations, goals, history.

Example: Knowing what it’s like to be a cat, or what it’s like to own a mat that cats sit on.

LLM capability: None. LLMs have no embodiment, no sensors, no goals (beyond the next token).

Human analogy: None. This is fundamentally beyond LLM capability.

Status: This is the hard boundary. We can build systems that act as if they have embedded understanding, but they don’t.

What LLMs Actually “Understand”

Based on this taxonomy, LLMs have:

This pattern of capabilities explains both the impressive performances and the characteristic failures of LLMs.

The Complementarity Principle

Here’s the crucial insight: LLMs and humans have complementary understanding profiles.

Understanding TypeLLM StrengthHuman Strength
Syntactic★★★★★★★★☆☆
Semantic★★★☆☆★★★★★
Episodic★☆☆☆☆★★★★★
Causal★★☆☆☆★★★★★
Embedded★☆☆☆☆★★★★★

The opportunity: LLM + Human combinations can achieve understanding that neither could alone.

This is the foundation of the Media Synthesis and Systems Consulting legs. In both cases, I’m providing the human understanding (episodic, causal, embedded) that the AI tools lack, while leveraging their syntactic and statistical strengths.

Implications for AI Systems Work

1. Know Your Limits

If you’re building LLM-based systems, be explicit about which types of understanding you’re relying on. The system’s capabilities and failure modes will differ dramatically based on this.

2. Memory is a Bridge

Memory systems (like Aurora’s multi-layer architecture) can bridge some of the gaps — particularly episodic understanding. But they can’t provide embedded understanding.

3. Human-in-the-Loop is Essential

For any high-stakes application, there needs to be a human who provides the understanding types that LLMs lack. The human doesn’t need to do the syntactic work — that’s what the LLM is for.

4. The Understanding Economy

As LLMs get better at syntax and semantics, the premium shifts to the understanding types they can’t provide:

This is why the Systems Consulting leg focuses on outcomes (“stop paying monthly”) rather than methods (“I’ll install Asterisk”). The value is in the causal understanding of business systems, not the syntactic manipulation of software.

A Thought Experiment

Imagine an LLM that has been trained on every conversation ever had by a particular human. Would that LLM “understand” the world the way that human does?

I think the answer is no. The LLM would have perfect syntactic understanding of that human’s communication patterns. It might even develop good semantic understanding of the concepts that human cares about. But it would lack:

The LLM would be an excellent simulator of that human’s speech. But it wouldn’t be that human, and it wouldn’t understand in the human sense.

Conclusion

Understanding is not a binary — it’s a spectrum across multiple dimensions. LLMs have transformed what’s possible with syntactic understanding, and they’re making inroads on semantic understanding. But the higher-order types remain uniquely human.

The most interesting work happens at the boundary — building systems that combine LLM strengths with human strengths to achieve understanding that neither could reach alone.

That’s what Badlucksbane is about.

Admonition

This piece is speculative. It’s based on my current understanding (all types) of how LLMs work and what understanding is. Both are evolving rapidly. Treat this as a working hypothesis, not a settled conclusion.

*Version history: