In a recent article here at Mind Matters News, “Horizontal Innovation Requires More Than Novel Outputs,” I argued that there is a fundamental difference between searching a predefined solution space and producing genuinely new functional outcomes. Search algorithms — even highly effective ones — operate within constraints that already define what counts as a valid solution. They do not create those constraints.

But what about randomness? Could an algorithm use randomness to produce genuine innovation?

The classic example is the “monkey with a typewriter.” We are told that, given enough time, a monkey randomly pressing keys will eventually produce the complete works of Shakespeare.

Recently, I watched Stephen Meyer discuss why randomness is unlikely to generate new functional information. He uses the analogy of guessing the combination of a lock; the more complex the lock, the longer random search is expected to take before finding the correct passkey. Likewise, a monkey randomly striking keys is extraordinarily unlikely to reproduce Shakespeare within any practical timeframe.

I agree with Meyer’s conclusion, but I believe there is a deeper issue that is often overlooked. Time is not the fundamental limitation — information is.

Searching for the unknown

Let us return to the monkey with a typewriter. How do we know the monkey has typed the first letter correctly? How do we know when it has finished? The answer is straightforward: we compare the output against Shakespeare’s original text. But that comparison quietly assumes that we already possess the very information the monkey is supposedly generating.

Viewed algorithmically, the process is simple. Generate a random character. Compare it to the corresponding character in Shakespeare’s text. If it matches, continue until all characters have been generated correctly. Otherwise, restart. The algorithm halts only because it has access to the complete target from the beginning. Remove Shakespeare’s text from the algorithm and the stopping condition disappears. The monkey continues typing forever — not because it lacks time, but because it lacks any criterion for recognizing success.

The same observation applies to brute-force password guessing. Random search succeeds only because the structure of the problem is already known: the number of digits, the allowed values, and the condition that signals success. Remove this external information, and random guessing ceases to be an algorithm with a goal. It becomes an unstructured stream of attempts with no objective and no termination condition.

Information is the key

The crucial ingredient is not randomness, but information: the structure that defines the target and determines success. Randomness merely explores a search space whose structure has already been supplied from the outside.

This reveals a deeper question. If every successful search process depends on external information that specifies what counts as success, where does that information come from? If every algorithm derives its specification from another algorithm, we are led to an infinite regress. At some point, the specification must enter the system from outside the algorithmic search process itself.

In other words, the monkey never finds Shakespeare unless Shakespeare is already embedded in the rules that define success.

Origin of information and artificial general intelligence (AGI)

The discussion above illustrates a broader principle. Randomness alone cannot generate genuinely new functional information unless that information is already implicitly encoded in the system performing the search. I refer to this kind of genuinely novel, functionally useful structure as functional information (or algorithmic novelty). This is the kind of information that enables humans to invent new algorithms and discover new functional capabilities that were previously unknown or impossible.

Whether creating literature, producing works of art, developing scientific theories, or inventing new technologies, we empirically observe that humans can generate new functional information that expands the space of what is possible. This ability is what allowed humans to progress from living in caves to building skyscrapers and eventually traveling into outer space.

This principle is central to the criterion I propose in my paper Turing Test 2.0: The General Intelligence Threshold  (2025), for detecting if a system has achieved artificial general intelligence (AGI). Instead of asking whether a system can imitate human behavior, the relevant question is whether it can generate new functional information that was not provided as input, hard-coded into its architecture, or implicitly specified by its training or search procedure.

If a system cannot produce such information, then what appears as “innovation” is ultimately exploration within a predefined search space. If it can, then we are dealing with a fundamentally different kind of computational capability — one that goes beyond standard notions of algorithmic search.

This naturally leads to an even deeper question: Can an algorithm ever generate genuinely new functional information—that is, can it demonstrate algorithmic novelty? More formally, is this remarkable human capability itself computable? I explore this question in my paper On the Computability of Artificial General Intelligence (2025).

Dr. Georgios Mappouras studied Electrical and Computer Engineering (ECE) at the National Technical University of Athens (NTUA), Greece. After graduating in 2014, he moved to the United States to pursue a Ph.D. in Computer Architecture at Duke University, which he completed in 2020. Since then, he has worked in the technology industry in Silicon Valley.