Beyond What Works: Ideas, Intelligence, and the Courage to be Creative

When we talk about intelligence, we often treat it as the ability to solve problems.

A problem appears. A system searches. A solution is found.

But this picture hides something deeper.

Not all solutions are of the same kind, because not all acts of intelligence operate at the same level. Some systems generate solutions within an established space. Others generate the spaces themselves.

To understand this distinction, we need to be precise about our terms.

An idea is a solution within a space. Intelligence is the ability to generate ideas within a given space. Creativity, in a functional sense, is the ability to generate new spaces by operating over generators themselves.

This is not a poetic distinction. It is a structural one.

Ideas are generated, not primitive

An idea is not a fundamental unit.

A proof, a hypothesis, a design—these are all outputs of a system operating under constraints. They emerge from representations, transformation rules, and evaluation processes.

Every idea presupposes:

• a way of representing the problem
• a way of generating candidates
• a way of evaluating outcomes

So rather than treating ideas as primitive, we treat them as outputs of a generator operating in a constrained space.

Evolution does not search every form

Biology provides a clear example.

Evolution is often described as a search over all possible organisms, guided by natural selection. But this is incomplete. Organisms do not vary arbitrarily; they vary according to the structure of the developmental systems that produce them [1].

The extinct Tasmanian wolf (thylacine) and modern wolves diverged over 100 million years ago, yet evolved strikingly similar forms. This is not just because the environment favored those forms. It is because both lineages inherited similar generative resources—developmental systems that bias what can be produced [2].

Selection chooses among possibilities.

But the generator defines the possibilities.

And crucially, those generators are themselves products of prior evolution.

Science operates within generators

Scientific reasoning follows the same structure.

Scientists do not explore all possible explanations. They operate within frameworks—mathematical, conceptual, and representational—that determine what kinds of explanations can exist.

Kuhn called these paradigms [4]. Within a paradigm, intelligence produces solutions. But when the paradigm itself changes, the space of possible solutions changes with it.

For centuries, geometry meant Euclidean geometry. The introduction of non-Euclidean geometry redefined what “space” could be, enabling entirely new lines of reasoning and ultimately new physical theories [5].

In quantum mechanics, the shift from probabilities as primitives to amplitudes as primitives fundamentally altered the structure of reasoning about uncertainty and interaction [6].

These were not just better answers.

They were new generators.

A functional hierarchy

We can now define a clean hierarchy:

Level 0 — Ideas

Solutions generated within a space.

Level 1 — Intelligence

A generator that produces ideas within a fixed space.

Level 2 — Creativity

A generator that operates over generators, producing new spaces of ideas.

Creativity, in this sense, is not about novelty or artistic expression. It is a mechanism for modifying the structure of possibility itself.

It is a generator of generators.

Formal notation

More precisely:

• Level 0: An idea is an element $x \in \mathcal{S}$, where $\mathcal{S}$ is the solution space defined by a generator.
• Level 1: Intelligence is a function $G: \mathcal{P} \to \mathcal{S}$ — a generator that maps problems to solutions within a fixed space.
• Level 2: Creativity is a meta-function $\mathcal{M}: \mathcal{G} \to \mathcal{G}'$ — a transformation that takes generators and produces new generators, thereby redefining the space of reachable solutions.

These levels are not merely conceptual. In the supporting experiments, we construct systems at each level and observe their qualitatively different behaviors.

Creativity as a system

This definition aligns with—but sharpens—existing work.

Margaret Boden distinguishes between exploratory, combinational, and transformational creativity [7]. What we are calling creativity corresponds specifically to her transformational category: changing the space itself.

But here we go further.

We treat creativity not as a descriptive label, but as a functional system:

Creativity(G₁) → produces G₂
G₂ → produces ideas

Creativity is not just producing unusual outputs. It is modifying the underlying generator.

Pseudo-code: intelligence vs creativity

Intelligence (fixed generator)

function solve(problem, generator):
    ideas = generator.generate(problem)
    return select_best(ideas)

This is standard intelligence: search within a fixed space.

Creativity (generator over generators)

function create_new_generator(problem, generators, meta_generator):
    candidate_generators = meta_generator.generate(generators)

    evaluated = []
    for g in candidate_generators:
        ideas = g.generate(problem)
        score = evaluate(ideas)
        evaluated.append((g, score))

    return select_best(evaluated)

Creativity operates at a different level. It does not generate solutions directly. It generates new ways of generating solutions.

The observable difference

The distinction between intelligence and creativity is not just theoretical — it produces measurably different dynamics. To demonstrate this, we simulate two systems searching a rugged fitness landscape with many local optima:

1. Intelligence (fixed strategy): A hill-climbing optimizer with a fixed Gaussian step size. It can only find solutions reachable by small perturbations of its current position.

2. Creativity (meta-strategy): The same base mechanism, but a meta-system periodically evaluates and mutates the search strategy itself — changing step sizes, switching between exploration methods, and generating entirely new strategies.

The result is striking:

The intelligence system (blue) converges rapidly to a local optimum and plateaus. The creativity system (red) shows the same initial convergence, but then makes a discontinuous jump — escaping the local basin when a strategy mutation opens access to a previously unreachable region of the landscape.

On the fitness landscape itself, the intelligence trajectory stays trapped in one basin. The creativity trajectory leaps across the landscape to find the global optimum (gold X) — a peak that no amount of refinement within the original strategy could reach.

This is exactly the pattern we should expect: the fixed system plateaus; the creative system exhibits discontinuous jumps. The jumps do not come from better execution within the existing strategy. They come from changing the strategy itself.

python run_experiment_1.py --n-steps 1000 --n-peaks 20

Generators have histories

Generators do not appear out of nowhere.

Developmental systems evolve. Mathematical frameworks are constructed over time. Scientific paradigms emerge through accumulated constraints and successes.

Most of us only ever see the solutions. We see relativity, but not how Einstein arrived at it. There is a sociological and historical explanation — the mathematical machinery he drew on, the work of Hilbert, the structure of Minkowski spacetime — and these are well documented [5]. But what Einstein went through internally to string those elements together, to embody a new concept deeply enough to expose it to the world — that generative process is far harder to record or audit. And it is the most important part of the story. The output was a theory. The creative act was the construction of a new generator: a way of seeing space, time, and gravitation that did not exist within any prior framework.

This is the pattern. We inherit the ideas but rarely the generators that produced them, and almost never the creative acts that built those generators in the first place.

Ioan P. Culianu described intellectual systems as ideal objects whose internal possibilities unfold historically [9].

This leads to a recursive structure:

• ideas are generated by intelligence
• intelligence operates through generators
• creativity modifies those generators

So the full system is:

Creativity → Intelligence → Ideas

Or more abstractly:

G₂ → G₁ → outputs

Why creativity looks like error

When a new generator appears, it often looks wrong.

Not because it fails within the current system—but because it does not belong to it.

Complex numbers were rejected. Non-Euclidean geometry was dismissed. Abstract mathematical systems seemed disconnected from reality.

But once the generator stabilizes, it becomes the new baseline.

What was once unintelligible becomes obvious.

AI and the boundary

Modern AI systems are extremely strong at intelligence in this framework.

They can:

• generate large numbers of ideas
• explore combinatorial spaces
• optimize within fixed representations

Consider AlphaFold: it is a remarkably powerful generator within a fixed representational space — amino acid sequences mapped to three-dimensional structures. It produces solutions that human scientists could not, but it does not invent new representations of protein structure. It operates entirely within a generator defined by decades of structural biology.

Large language models demonstrate a similar pattern. They excel at what Boden would call combinational creativity — recombining learned representations in novel ways, synthesizing across domains, generating fluent text in styles they have absorbed [7]. But they rarely perform transformational creativity: inventing a new representational space that makes previously inexpressible ideas expressible.

This distinction maps precisely onto the hierarchy. LLMs are sophisticated Level 1 systems — powerful generators within the representational space defined by their training. They do not typically modify the generator itself.

Gian-Carlo Rota described a “barrier of meaning” in mathematics — the moment when a structure is seen as something new, not merely rearranged within its existing interpretation [10]. Seeing something as something else is a generator-level operation, not an output-level one. It requires redefining the representational primitives, not just recombining them.

This is the boundary.

The difference between generating better answers and generating better ways of answering.

What this means for us

This boundary is not unique to machines.

It is easy to become comfortable in what has already worked.

And we should be proud of that. Every system we inherit—scientific, mathematical, cultural—represents real progress. These generators encode generations of effort.

But they are not final.

The mistake is to confuse stability with completeness.

The world will always pull us toward established structures. It rewards what fits. What is legible. What has precedent.

But the universe does not operate within our generators.

It continues to present forms that exceed them.

A functioning society needs intelligence—people who can operate effectively within existing systems.

But it also needs creativity—systems and individuals capable of generating new ones.

These are often harder to understand. They do not align with current expectations. They may appear incoherent.

But that incoherence may not be failure.

It may be a signal that the current generator is insufficient.

So the next time you feel the impulse to reject something because it does not make sense, pause.

Ask whether the failure is in the idea—or in the generator you are using to interpret it.

The same applies beyond ideas.

Even the smallest forms of life—things we dismiss as insignificant—are products of entirely different generative trajectories. They embody structures we did not produce, solutions we did not discover, generators we cannot yet replicate.

That is why we have each other.

Not because we think the same way—but because we don’t.

Society is not pure order. It is a system that must maintain structured variation. The more variation it can sustain without collapse, the more capable it is of adapting to an unpredictable universe.

Because the universe is not fixed.

It is open.

And if we can learn to embrace what is different—not just tolerate it, but engage with it—we transform uncertainty into curiosity, and the unknown into something we can approach.

Something we can play with.

…and maybe even shape.

Inspiration

This essay owes a deep debt to Jacob Foster’s talk “Ideal Objects”, delivered as part of a symposium on the platonic space. Foster — a professor of cognitive science and informatics at Indiana University Bloomington and an external professor at the Santa Fe Institute — articulates with remarkable clarity the distinction between exploratory, combinatorial, and transformational creativity, drawing on Boden’s framework and extending it into a theory of how intellectual systems unfold as ideal objects whose internal possibilities are explored historically.

His discussion of how AI excels at tinkering and combinatorial search but struggles with transformational creativity — and his framing of what it means to see a structure as something new rather than merely rearrange it — directly shaped the hierarchy presented here. The idea that generators have histories, that creativity operates over the structure of possibility rather than within it, and that the most important innovations look like errors from within the current paradigm: these insights crystallized while engaging with this talk.

I am grateful to Professor Foster for making these ideas publicly available, and I encourage anyone interested in the deeper structure of creativity, intelligence, and scientific discovery to watch the full lecture:

Experiments

The computational experiments supporting this essay are available in the project repository:

• Experiment 1 — Fitness Landscape Optimization: Intelligence (fixed strategy) vs. Creativity (meta-strategy) on a rugged landscape.
• Experiment 2 — The Generator Hierarchy: Agent-based demonstration of three levels — fixed behavior, weight evolution, and architecture evolution in a grid world.

References

1. Müller, G. B. (2007). “Evo–devo: extending the evolutionary synthesis.” Nature Reviews Genetics, 8(12), 943–949.
2. Feigin, C. Y., et al. (2018). “Genome of the Tasmanian tiger…” Nature Ecology & Evolution.
3. Thompson, D. W. (1917). On Growth and Form.
4. Kuhn, T. S. (1962). The Structure of Scientific Revolutions.
5. Riemann, B. (1868). Foundations of geometry.
6. Dirac, P. A. M. (1930). The Principles of Quantum Mechanics.
7. Boden, M. A. (2004). The Creative Mind: Myths and Mechanisms. 2nd ed. Routledge. See Ch. 3 for the distinction between exploratory, combinational, and transformational creativity.
8. Turing, A. M. (1936). Computable numbers.
9. Culianu, I. P. (1992). The Tree of Gnosis.
10. Rota, G.-C. (1997). Indiscrete Thoughts. Birkhauser. See the discussion of the “barrier of meaning” in mathematical understanding.
11. Hofstadter, D. (1979). Godel, Escher, Bach: An Eternal Golden Braid. On strange loops and tangled hierarchies as self-referential generators.
12. Holland, J. H. (1992). Adaptation in Natural and Artificial Systems. On the formal structure of evolutionary search operators.

This essay is part of the SHARC Theory research project.