I recently noticed something amusing while working with AI on a mathematical problem.

I proposed an idea. The AI elaborated it. I corrected the elaboration. The AI produced a more sophisticated version. I corrected that. The AI introduced mathematical terminology. I corrected that too.

Eventually, I realized something almost comical:

I was telling the AI what the idea was, and the AI was explaining my idea back to me.

😂

The Thought Magnifying Glass

That gave me a name for the phenomenon:

TMG — the Thought Magnifying Glass.

A magnifying glass does not create the object underneath it. It makes the object appear larger.

AI can sometimes behave in exactly this way with human thought. A small human insight can be transformed into terminology, structure, examples, analogies, mathematical notation, references, counterarguments and polished prose.

The visible output becomes enormous compared with the original input. But the intellectual origin may still be the original human idea.

The illusion of intellectual enlargement

Suppose a researcher asks an AI:

There are 256 possible local configurations. What is the minimum information required to trace the graph?

An AI can immediately surround the question with sophisticated language: finite-state reconstruction, sufficient statistics, topology-preserving transitions, information compression and so on.

That sounds impressive. But the important question is not whether the AI can describe the problem elegantly.

Did it discover the problem and its organizing principle?

Searching is not discovering

This distinction becomes particularly important in mathematical AI.

  • Finding an existing theorem or proof is one task.
  • Combining known mathematics into a new argument is another.
  • Discovering the right representation, invariant or algorithm from an unfamiliar problem is another.

A system can become extraordinarily good at the first two without demonstrating the same ability in the third.

The 256-state experiment

Consider a very small mathematical experiment. Take a 3×3 binary neighbourhood with the central cell fixed as filled. There are eight remaining cells, and therefore:

28 = 256 possible configurations.

The entire local state space is finite and enumerable. The information needed for the experiment is already present in the question. The challenge is to discover what representation is sufficient to reconstruct the local graph and determine the next carving operation.

A tempting first classification is valency. But valency alone is insufficient. Two occupied neighbours may form one connected local mass or two separate diagonal arms. Both can have V8 = 2 while having different local topology.

The useful abstraction is therefore not simply valency. We need to examine the connectivity of the neighbouring occupied cells after temporarily removing the centre. This naturally leads toward an arm-count concept:

ARM_COUNT = the number of 8-connected occupied components among the neighbouring cells after removing the centre.

Now the distinction becomes visible:

Valency ≠ topology.

The interesting question is not merely whether an AI can explain this once the concept has been suggested. The interesting question is whether it can discover the abstraction independently from the original specification.

What does AI actually do?

There is an important technical qualification. Modern AI systems are not simply retrieving a collection of pre-written answers. Their output is generated dynamically from learned parameters, architecture, instructions, context and computation.

But that does not establish human-like subjective understanding. A model can generate a remarkably coherent explanation without there being a human-like subjective experience of the idea behind the explanation.

Therefore:

Explanation capability ≠ discovery capability ≠ subjective understanding.

How TMG can amplify the AI bubble

Imagine a small human intellectual contribution being passed through an extremely powerful language-and-computation amplifier. The visible output may become hundreds or thousands of times larger than the original thought.

That amplification can create a measurement problem. Observers may judge intelligence from the sophistication and volume of the output while failing to separate the original conceptual contribution from the machine-generated elaboration.

A possible feedback loop is:

Human idea → AI magnification → impressive demonstration → perceived intelligence → investment and attention → more demonstrations → greater perceived intelligence.

This does not mean AI is fake or that every AI claim is inflated. It means that the metric used to judge capability matters.

The TMG test

If we want to measure genuine mathematical discovery, a better benchmark is to give an AI an unfamiliar problem with complete information and say:

No hints. Discover the representation yourself.

Then test whether the system independently discovers the relevant invariants, counterexamples, equivalence classes and algorithmic structure.

For the 256-state experiment, we could ask the model to determine what information is necessary, what is redundant, which configurations are equivalent and how the local rules extend into a global graph-tracing algorithm.

If it independently reaches the same abstraction, that is interesting evidence of discovery. If it merely rearranges concepts supplied during the conversation, the impressive output was primarily amplification.

TMG does not mean AI is useless

Quite the opposite. A magnifying glass is extremely useful. The fact that it does not create the object does not make it useless.

AI can be an extraordinary formalizer, exhaustive checker, counterexample generator, literature navigator, programmer, proof critic, visualization engine and idea amplifier.

The mistake begins when amplification is confused with authorship of the underlying insight.

Human + AI may be the better architecture

Perhaps the most productive model is not:

AI replaces researcher.

It is:

Human discovery + AI magnification + AI verification + AI exploration.

The human supplies intuition and direction. The machine supplies enormous computational breadth. The human proposes a strange possibility. The machine tests 256 cases. The machine finds a counterexample. The human changes the conceptual model. The machine formalizes it.

The TMG principle

Never estimate an AI system’s independent intellectual contribution from the size, sophistication or polish of its output. First determine how much of the underlying conceptual structure was already supplied by the human.

Or, more bluntly:

A magnified thought is still the original thought.

And perhaps the best demonstration of TMG is wonderfully self-referential.

I can take this idea, add mathematical terminology, divide it into sections, construct equations around it and produce a polished article.

That makes the idea look much bigger.

But the original observation was very simple:

“You are a thought magnifying glass.”

😂

Related: The Ferrari on a One-Lane Highway — When AI Capability Meets Its Own Bottlenecks