AI’s Missing Layer: Why the Future Might Belong to Symbolic Knowledge Engines Connected by LLMs

Perhaps LLMs are not the brains of AI. They are the universal semantic interface to a new cognitive infrastructure.

For a while I’ve had the feeling that the public debate about artificial intelligence might be aimed at the wrong target.

People talk about large language models (LLMs) as if they are embryonic digital minds. Regulators then shape policy around fears of a model “waking up.” That framing might be wrong. I suspect it misreads the technology and confuses the part of the stack that matters least with the part that will matter most.

LLMs might not be the ultimate brains of AI.

LLMs might be the universal semantic interface. They are the first systems that let any person express intent in natural language and have machines interpret, execute, and explain that intent. They translate between human thought and machine logic. They are not the engine. They are the protocol.

The real power will sit behind them:

millions of auditable, high-stakes, domain-specific symbolic knowledge engines containing the rules, constraints, definitions, and causal structures society depends on.

LLMs will build these engines, update them, verify them, and mediate how humans interact with them. Neural models will handle language, perception, and flexible reasoning. Symbolic systems will hold the stable memory and enforce correctness.

I suspect this hybrid architecture is the next layer of civilisation.

1. LLMs are not the processor. They are the interface.

LLMs are extraordinary at:

• interpreting intent
• summarising and contextualising
• orchestrating tools
• translating between abstractions
• turning ambiguous prompts into structured plans

But they are not reliable reasoners. They are probabilistic. They drift. They cannot guarantee consistency across tasks or across time.

In computing terms, LLMs behave as semantic interfaces: a translating layer that converts human intention into machine operations, and machine logic back into natural language. They connect systems that would otherwise be incompatible.

They enable cognition, but they do not anchor knowledge.

For that, we need something structurally different.

2. The return of symbolic systems, with a new role

Symbolic AI once tried to model the whole world through logic. It failed. The world was too fluid; the rules were too brittle.

This time symbolic systems return in a narrower, more powerful role:

• representing the knowledge that must be right
• encoding rules, constraints, exceptions, and thresholds
• providing versioned, transparent, machine-checkable memory
• anchoring high-stakes decisions where “vibes” are unacceptable

Tax. Medicine. Aviation. Law. Accounting. Engineering. Compliance. Safety. Scientific protocols. Value investing systems like the one I talk about on the QAV podcast. These are rule-governed domains where correctness and explanation matter.

LLMs change the economics of symbolic systems. They can:

• extract candidate rules from documents
• propose and refine ontologies
• detect contradictions
• generate test cases
• mediate updates
• interface symbolic engines with each other

But the hand-off from neural to symbolic is the danger zone.

The formalisation gap: the core unsolved problem

Real-world expert knowledge contains ambiguity, context, implicit assumptions, and exceptions. Today’s best LLM-assisted formalisation pipelines still miss nuances and produce rule sets with significant error rates.

If the LLM adds or corrupts a rule during formalisation, the resulting mistake becomes a perfectly executed error. This is the “logic bomb” problem.

Hybrid architectures therefore require:

• automated verification
• contradiction detection
• provenance and version control
• simulation environments
• human oversight for the highest-risk layers

And there is a deeper frontier: auto-formalisation with formal verification.

In this model, the LLM writes rules in a formal language like Lean or Coq, and an independent theorem prover checks correctness. The LLM proposes; the prover disposes. AlphaGeometry is the first major demonstration of this loop.

This shifts correctness from something humans inspect to something compilers prove.

3. Where symbolic systems work, where they break, and the messy middle

Symbolic systems excel in domains with stable definitions and low tolerance for error:

• tax and accounting
• clinical guidelines
• safety specifications
• value investing rulesets
• engineering constraints
• regulatory frameworks

They fail in open, fluid domains like social interaction or creative work.

But most real domains live in between. Even tax law uses terms like “reasonable.” Even driving contains hard constraints defined by physics.

The real frontier is partitioning:

• what must be symbolic
• what can remain neural
• how the two communicate
• how errors propagate between them

The emerging pattern across 2024–2025 systems is consistent:
neural models propose possibilities in ambiguous zones; symbolic constraints define what is allowed, safe, or legal.
It’s not “neural or symbolic.” It’s a negotiation.

Examples already in production:

• AlphaGeometry mixes neural proposals with formal deduction
• DeepSeek-Math uses neural search with rule-based checking
• OpenAI’s o3 drives tools under strict symbolic constraints
• Lean/Llemma blends learned tactics with proof assistants

The pattern is clear: a hardened symbolic core surrounded by a flexible neural layer.

4. QAV as a practical example of symbolic knowledge becoming executable

QAV — a rules-based value investing methodology — shows exactly why symbolic systems matter.

Today, QAV’s rules live as prose in a long document. An LLM can interpret the rules, but it can also:

• misapply a threshold
• forget an exception
• blend two rules
• invent a detail

A symbolic QAV engine transforms the method into something deterministic:

• every rule becomes explicit and checkable
• intrinsic value calculations become code, not text
• liquidity and sentiment filters execute exactly
• every recommendation produces a rule trace

The QAV document remains the human-readable specification.

The symbolic engine becomes the executable truth.

The LLM becomes the interface and the maintenance assistant.

This pattern generalises across thousands of domains.

5. Personal vs centralised symbolic knowledge engines

Some symbolic engines will be personal — reflecting individual expertise, company-specific processes, research frameworks, or bespoke methodologies.

Many more will be institutional:

• tax ontologies
• medical guidelines
• safety standards
• engineering codes
• regulatory rulebooks

These will be maintained by governments, standards bodies, and major firms.

Whichever institutions control the canonical symbolic rulebases will control the reasoning layer of civilisation.

This is the real battleground of AI governance in the next decade.

6. The competing path: end-to-end latent reasoning at scale

There is a rival vision: purely neural systems with huge amounts of “thinking time” at inference.

Recent reasoning models like OpenAI’s o3/o3-pro, the GPT-5-series reasoning variants, Google’s Gemini 3 Pro, and DeepSeek’s R1 / DeepSeekMath-V2 show that with enough test-time compute, a neural network can simulate many symbolic behaviours internally.

They already demonstrate:

• chain-of-thought style reasoning
• scratchpad planning and tool use
• multi-step code synthesis and self-correction
• theorem-like reasoning in math and formal domains

In many areas they now beat classical neuro-symbolic hybrids on:

• raw performance
• speed
• engineering simplicity
• cost

On a whole class of semi-high-stakes tasks — contract review, accounting workflows, compliance scanning, due-diligence checklists — these opaque neural systems are already competitive with senior human specialists.

Economically, the pressure will be brutal:
many organisations will choose “good enough but opaque” unless failure is catastrophic.

The problem is opacity.

When a latent reasoning model makes a serious error, there is:

• no explicit rule trace
• no clear ontology
• no guarantee the same reasoning holds tomorrow

For medicine, aviation, critical infrastructure, regulated finance, and national security, that is not acceptable.

In those domains, society will demand:

• explicit rules and constraints
• provenance of changes
• audit trails
• independent verification

The argument here is not that symbolic systems win everywhere.

It’s that for the decisions we most care about, opacity is itself a failure mode.

Both paradigms will coexist.

7. The ontology problem: how symbolic engines talk to each other

Symbolic systems traditionally fail to interoperate. If two systems disagree on the definition of “profit,” “risk,” or “diagnosis,” they cannot exchange information without manual reconciliation.

This is the Tower of Babel problem.

The missing piece is that LLMs are not only interfaces for humans; they are ontology translators. They reconcile definitions, map concepts between rulebases, mediate contradictions, and generate interoperability layers.

Institutions will maintain canonical ontologies.

Individuals and companies will maintain local ones.

LLMs will translate between them.

This is how a decentralised symbolic world becomes coherent instead of fragmented.

8. Why policymakers need to shift focus

Policy discussions are still obsessed with model size, training data, and content filters. Of course these issues matter, but perhaps they are not where the true systemic risk lies.

In hybrid systems, the catastrophic failures will come from:

• misencoded symbolic rules
• missing exceptions
• contradictory clauses
• outdated guidelines
• drift in regulatory ontologies

The LLM will faithfully execute and explain a flawed rulebase.

The failure will be in the knowledge layer, not the language layer.

That means regulated industries will need:

• provenance tracking
• version control
• automated verification
• audit logs
• simulation-based testing
• rulebase certification
• knowledge engineers as a core profession

This is where safety, liability, and governance will converge.

9. A realistic, ambitious vision for the next cognitive infrastructure

Here’s the world we are heading toward:

• LLMs act as the universal interface for human and machine knowledge.
• Symbolic systems hold the long-term structure and enforce correctness.
• Millions, eventually billions, of knowledge engines arise.
• Institutions maintain canonical ones; individuals maintain local ones.
• LLMs mediate, translate, reconcile, and explain.
• Neural reasoning and symbolic verification converge into hybrid loops.

Fragments of this world already exist:
neural-guided theorem provers, self-verifying code synthesis, medical pilots that mix LLMs with symbolic guidelines, and regulated-industry prototypes where LLMs act as the front-end to certified rulebases.

This is not speculative. It is emerging.

Human knowledge no longer dies in human heads or hides in PDFs. It becomes:

• structured
• executable
• interpretable
• updateable
• preserved across generations

Language becomes the bridge.
Neural networks become the translators.
Symbolic systems become the anchor.
Humanity becomes the author of its own extensible minds.

This is the next layer of civilisation.