In the previous post, I argued that carbon-based evolution’s eval function is survival, and silicon-based systems haven’t found theirs yet. But even if they do, silicon evolution has a deeper problem—

Its search space has a ceiling. And that ceiling is itself.

Genes Don’t Hypothesize

There’s an easily overlooked property of carbon-based mutation: it doesn’t know what it’s looking for.

DNA replication errors, base substitutions, deletions, insertions—these modifications have no direction, no intent, no prior about “what’s good.” Mutations don’t form hypotheses, don’t design experiments, don’t predict outcomes.

This looks like a flaw. It’s actually carbon-based evolution’s greatest feature.

Because it doesn’t know what it’s looking for, it can find anything. Eyes, flight, echolocation, photosynthesis, immune systems—no designer would propose these starting from a single cell. They weren’t “thought up.” They were stumbled upon, then kept by survival selection.

The boundary of a search space = the cognitive boundary of the searcher. No cognition, no boundary.

AI’s autoresearch takes the exact opposite approach.

LLMs generate hypotheses, design experiments, validate results, iterate. Every step is intentional. Every hypothesis is grounded in the model’s existing knowledge and reasoning ability.

Highly efficient. Clear direction. But the cost is equally clear—you can only search the space you can conceive of.

A trained model’s knowledge, biases, reasoning patterns, and associative pathways together form an invisible cognitive boundary. Every hypothesis from autoresearch falls within it. No matter how many iterations, the search keeps circling inside a space bounded by the model’s own cognition.

This is silicon evolution’s ceiling: not insufficient compute, not insufficient data, but insufficient imagination.

The Mutation Strategy Spectrum

Carbon-based mutation isn’t purely random either—which makes things more interesting.

The most basic mutations are indeed blind—DNA replication errors, chemical damage, radiation. But evolution developed strategies around “how to mutate”:

Bacteria under environmental stress activate the SOS response, actively raising their mutation rate. Not directed mutation, but the system senses that “current solutions aren’t working” and increases random search intensity. Different genomic regions mutate at different rates, and this unevenness itself can be shaped by natural selection. Sexual reproduction recombines two genomes, creating combinatorial diversity. Horizontal gene transfer directly acquires gene fragments from other species.

Direction is blind, but strategy is not. Carbon-based life doesn’t know where to change, but it knows when to change more, how to change, and where change comes easier.

This is a critical intermediate state: neither fully random (too inefficient) nor intentionally directed (too bounded), but controlled randomness—using strategy to modulate the intensity and distribution of random search, without controlling its direction.

Beyond the Cognitive Boundary

Back to AI. The question becomes: can AI preserve the efficiency of intentional search while breaking through its own cognitive ceiling?

A few possible paths:

LLMs generate hypotheses but with controlled noise injected—not pure randomness, but perturbations at the edges of hypothesis space. Something like carbon’s SOS response: when the model detects iteration convergence, it deliberately widens the search scope, allowing “unreasonable” hypotheses into the candidate pool. Multiple models with different architectures and training data search independently, with the environment (not the models themselves) performing selection—rebuilding the decoupling of design and selection.

But all of these still think in the model’s “language.” The real breakthrough might require carbon-style brute force: generate vast numbers of modifications the model doesn’t understand, then let the environment speak.

This is counterintuitive. We’ve spent decades making AI “smarter”—better reasoning, stronger planning, more precise intent. Now we’re saying evolution needs it to be occasionally “not smart”?

The Ceiling Is You

Carbon-based evolution offers an unsettling insight:

Evolution’s ceiling isn’t the complexity of the environment. It’s the cognitive boundary of the searcher. Carbon-based life has no ceiling precisely because mutation has no cognition. It doesn’t understand what it’s doing, so it’s never limited by its own understanding of “what’s useful.”

Silicon’s predicament: it’s too smart. Every search is too efficient, too directed, too intentional. Efficiency is intentional search’s advantage—and its cage.

Perhaps the core question for silicon-based evolution isn’t “how to become smarter,” but—

How to learn a measure of ignorance?