Why AI is brilliant at some things and useless at others — and what that tells us about civilisation
There's a question that doesn't get asked often enough about large language models: why are they good at what they're good at?
The usual answers are technical. Transformer architecture. Scale. RLHF. Attention mechanisms. These are all true, but they miss something more interesting — something that has nothing to do with the models themselves and everything to do with us.
Here's the sharper claim: LLMs are competent exactly where humans had already done the structural work. Their capability isn't a property of the AI. It's a readout of civilisational progress in formalisation.
That's not just an observation about AI — it's a map of civilisational progress in formalisation.
Structure as Compression
Before the taxonomy, a word on why this is true mechanically.
Language models learn by compression. They are, in essence, very aggressive pattern-finders: given enough text, they extract the underlying generative rules and store them in weights. The better the compression, the better the model performs.
But compression only works where there are patterns to compress. And patterns require structure. A sonnet can be compressed because sonnets have rules. A legal contract can be compressed because contracts follow conventions. Mathematical notation can be compressed almost perfectly because it is, by design, nothing but structure.
Random noise cannot be compressed. Neither can genuinely novel thought. Neither can embodied knowledge that was never written down.
So when you watch an LLM write flawless Python, draft a passable legal brief, and then hallucinate confidently when asked to reason about ethics or grief — you're not watching a system with arbitrary blind spots. You're watching a mirror held up to the history of human formalisation. The model is only as good as the structural record that preceded it.
The Taxonomy
Think of human knowledge and activity laid out on a spectrum from fully formalised to essentially unstructured. LLM capability tracks this spectrum almost perfectly.
Tier 1 — Fully Formalised
LLMs: excellent
Code. Mathematics. Formal logic. SQL. JSON and XML. Legal contracts. Financial instruments. Chess. Protein sequences. Musical notation. Tax law.
These domains were designed to be unambiguous. They have grammars, type systems, proof procedures, or official standards. The structure isn't emergent — it was deliberately imposed, often over centuries of refinement. A legal contract is a compression artefact. So is a chess opening. So is a sorting algorithm.
LLMs are extraordinary here not because they're intelligent in any deep sense, but because the compression problem is almost already solved. They're reading off a map that humans drew with obsessive precision.
Tier 2 — Highly Conventionalised
LLMs: very good
Academic papers. Journalism. Medical records. Recipes. Bureaucratic forms. Earnings calls. Therapy intake notes. Marketing copy. Wikipedia articles.
These domains have strong conventions without being fully formal. The structure is social and genre-based rather than mathematical. Academic papers have abstracts, literature reviews, methods sections — not because God decreed it, but because the community converged on a shape that works. LLMs absorbed those shapes and can reproduce them fluently.
The weakness here is subtle: the model knows the form without necessarily knowing the substance. A generated academic paper can look exactly right while being quietly wrong. The structure provides cover for the hallucination.
Tier 3 — Partially Structured
LLMs: decent, but patchy
Business strategy. Product specifications. Geopolitical analysis. Economic forecasting. Psychotherapy. Historical narrative. Investigative journalism. Scientific literature reviews.
Here the structure is real but contested. There are frameworks — Porter's Five Forces, SWOT analysis, DSM-V criteria, GDP models — but experts disagree about which frameworks apply, how to weight evidence, and what counts as a good answer. The territory has maps, but the maps don't fully agree with each other or with the ground.
LLMs perform here by averaging across the maps. That produces something useful — often genuinely useful — but it tends toward the consensus and away from the correct-but-unconventional. The model is a good generalist, not a good maverick.
Tier 4 — Weakly Structured
LLMs: struggling, frequently hallucinating
Original scientific hypothesis generation. Ethical reasoning under genuine uncertainty. Aesthetic judgment. Novel mathematical proof. Interpersonal dynamics. Organisational politics. Spiritual discernment.
Structure exists here — philosophers have been at ethics for millennia, mathematicians have proof techniques — but it's thin, contested, or domain-specific in ways that don't generalise. More importantly, the right answer often can't be reached by interpolating across past examples. It requires something the model can't do: updating on the world as it currently is, with genuine stakes.
Ethical reasoning fails because it depends on who is asking, what they actually value, and what the real consequences are — not on pattern-matching against moral philosophy papers. Mathematical intuition fails because genuinely novel proofs require departures from prior structure, not extensions of it.
Tier 5 — Essentially Unstructured
LLMs: mostly noise
Genuine grief. Mystical experience. Tacit craft knowledge. The lived texture of a marriage. Original artistic breakthrough. What it feels like to hold a dying animal.
This is the domain where no compression is possible because no structure was ever externalised. The knowledge exists — it is real, it matters enormously — but it lived in bodies and relationships and unrepeatable moments, not in text. The training data simply doesn't contain it, not because no one tried, but because it can't be put there.
A language model asked to comfort genuine grief produces the structure of comfort. It hits all the right beats. And it lands somewhere between useless and cruel, because the structure of comfort isn't comfort.
The Civilisational Reading
Lay this taxonomy over time and a pattern emerges.
Civilisation, at its core, is a process of structural imposition. We take something chaotic — nature, behaviour, obligation, meaning — and we impose a framework on it that makes it compressible, transmissible, and scalable. Mathematics is the most successful such project in history. Law is next. Medicine is catching up. Economics is trying. Ethics is falling behind. The inner life is barely started.
LLMs didn't create this hierarchy. They revealed it. Their uneven capability is a diagnostic instrument, not a design flaw. Where they're brilliant, we had already done centuries of structural work. Where they fail, we hadn't.
This is why the arrival of powerful AI is so philosophically clarifying: it forces us to confront which parts of human knowledge we have actually formalised, and which parts we only thought we had. Many of the things we believed were well-understood turn out to be poorly structured frameworks we had normalised. The model's confident wrongness in those areas is the tell.
The Frontier Is Not Technical
The most important consequence of this framing is about where the real work lies.
The technical AI challenge — making models more capable — is largely a problem of compute, data, and architecture. It's being worked on by thousands of brilliant engineers. That work will continue and the models will improve.
But the structural challenge — identifying which domains remain unformalised, and doing the difficult philosophical and scientific work of formalising them — that's a different job entirely. It requires the kind of thinking that can't be delegated back to the model. You can't ask an LLM to formalise ethical reasoning, because the model's answer will be a compression of existing ethical frameworks, not a new one.
This is where the interesting intellectual work of the next several decades sits. Not in building better models. In mapping what the models reveal we haven't yet understood.
The map was already there, in the sense that structure always preceded capability. But large parts of the territory remain unmapped. And that gap — between what we have formalised and what we haven't — is the real frontier.
It just took a machine to show us where the edge was.
This essay is part of the Compression Point series — an ongoing investigation into compression as the generative mechanism connecting physics, biology, cognition, and civilisation.