The 2% Con: AI Solved the Flashy Bit and Left the Hard Architecture Behind

AI has done something extraordinary.

It has industrialised generation.

Give it a prompt and it can produce code, designs, analysis, reports, images, software and apparently intelligent reasoning at ridiculous speed.

That part is real.

The uncomfortable question is whether we have mistaken the most scalable part of intelligence for the whole thing.

Let's call generation the 2%.

Not because anyone has measured intelligence and discovered precisely 2% of it lives inside a Large Language Model. The number is deliberately provocative.

The point is simpler:

We have industrialised the generator. We have not industrialised most of the machinery required to make generated intelligence cumulative, controlled and durable.

And the awkward thing is that computer science knew about much of this difficult machinery decades ago. We did not suddenly discover the problem after ChatGPT.

AI Used to Worry About Keeping the Books

Go back to 1979. Jon Doyle published his Truth Maintenance System.

Its purpose was remarkably relevant to today's problem: reasoning programs needed to record why they believed things, maintain dependencies between beliefs and revise those beliefs when later discoveries contradicted their assumptions.

Before transformers, GPUs or anyone asking AI to generate a horse wearing sunglasses, researchers were already asking:

In other words, they were trying to build the accounting department of intelligence.

Then Came Belief Revision

By the mid-1980s, the AGM framework for belief revision was asking an even more direct question:

How should an existing body of beliefs change when new information arrives? And how do you make that change while disturbing the existing belief state as little as reasonably possible?

That is remarkably close to what we are now calling the AI Cumulative Hiatus Delta: the gap between generating a plausible next answer and reliably maintaining the next trustworthy state.

The problem is nearly half a century old. We still do not have a general answer.

Expert Systems Hit the Hard Bit Head-On

The expert-system boom tried another route. Build an explicit knowledge base. Represent facts. Represent rules. Separate knowledge from the inference engine operating upon it.

That gave systems something current generative models generally do not possess internally: explicit, inspectable knowledge structures that humans could modify.

And then reality arrived. The industry hit the knowledge acquisition bottleneck. Extracting real expertise from humans, structuring it correctly, representing exceptions and keeping the resulting knowledge base current was painfully difficult.

The hard part was not making the inference engine run. It was building and maintaining the knowledge state. There is a lesson in there.

Then Scaling Won

By 1990, knowledge representation was still being described as the heart of artificial intelligence. Then statistical machine learning got very, very good.

Instead of manually representing the world's concepts, relationships, exceptions and rules, we learned to train large models that could infer distributed representations from data. The results were spectacular — speech, vision, translation, language generation, transformers, GPT. The field had found something that scaled.

But symbolic AI was not replaced because every difficult question about knowledge, belief, revision and state had been solved. A different branch simply became vastly more scalable. That distinction matters.

The Creative Department Staged a Coup

Modern generative AI is magnificent at the creative part: recognising patterns, predicting continuations, generating candidates, recombining concepts, approximating reasoning, producing possibilities.

That is the glamorous department. It walks onto the stage. It writes Shakespeare. It writes Python. It passes exams. It generates video. Everyone applauds.

Meanwhile, the accounting department is downstairs shouting:

Nobody invited accounting to the keynote.

The Difficult Bits Didn't Disappear

We still need:

These are not small finishing touches. They are the machinery that prevents intelligent output from becoming uncontrolled state mutation.

And current research is still visibly working on fragments of the same problem. Continual learning exists because sequential updating can degrade previously acquired knowledge. Formal verification can prove only what has been explicitly represented. Testing can expose observed failures but cannot cover an unrepresented world. Knowledge graphs can traverse only the relationships they contain. Provenance can establish where something came from without making it true.

The old problems did not disappear. We simply became extraordinarily good at another part of the machine.

The Immaculate Conception Strategy

This is where the current AI trajectory becomes interesting. The dominant direction increasingly looks like:

more parameters + more compute + more context + more agents + more self-checking = eventually, the rest of intelligence

Perhaps. But there is no demonstrated general mechanism showing that scale alone produces persistent authoritative state, controlled cumulative change, complete dependency management, independent settlement, or protection against things the system does not know it failed to represent.

Better generation clearly produces better generation. What has not been established is that a sufficiently good proposal engine automatically creates the rest of the architecture. That is the Immaculate Conception Strategy: make the generator bigger until the missing machinery appears by itself. It may work. But at the moment, it remains an assumption.

We Copied the Creative Cortex and Forgot the Organism

AI became extremely good at modelling parts of cognition. Then the success became so overwhelming that we began treating cognition as though it were the whole organism. But biology does not work like that. Creativity does not maintain the entire organism. Other systems regulate, preserve, correct, reject, repair, remember, and respond to disturbance.

Human civilisation works the same way. Individual humans are unreliable reasoning machines. So we invented accounting. Science. Engineering configuration control. Legal records. Audit. Testing. Version control. Peer review. Certification. Institutions.

We did not solve unreliable human intelligence by waiting for a human with a sufficiently large brain. We built the missing architecture around it.

The 2% Con

So the "2% con" is not that AI is fake. It is not. The technology is astonishingly real.

The con is subtler. We may have collectively mistaken spectacular progress in the most scalable part of intelligence for evidence that the remaining architecture would follow naturally.

It has not. Not yet.

We know how to industrialise the generator. We do not yet know how to industrialise the full machinery required to preserve, revise, verify and extend established knowledge over thousands of consequential changes.

That remaining architecture may come from old AI ideas. Or biology. Or formal methods. Or causal discovery. Or systems engineering. Or something nobody has invented yet.

But another trillion parameters do not make the unanswered question disappear. They simply make the creative department even better.

And at some point someone still has to ask accounting: what exactly changed, why did it change, what does it affect, and do the books still balance?

That may turn out to be the other 98%.