Unprogrammed Intelligence: From Biology to Code

In a world increasingly shaped by the opaque logic of artificial intelligence, a quiet revolution is brewing.

This revolution challenges our most fundamental assumptions about intelligence itself.

It’s a notion so audacious it might sound like science fiction.

Yet, meticulously gathered evidence suggests we might have profoundly underestimated the ‘mental’ lives of everything from our own cells to the most rudimentary computer code.

At the heart of this paradigm shift is Michael Levin, a maverick developmental synthetic biologist.

He is the co-creator of the pioneering Xenobots, the world’s first AI-designed living machines.

For Levin, who leads the Allen Discovery Center at Tufts University, the question isn’t whether AI will become superintelligent.

Instead, it’s about how much unprogrammed intelligence already lurks in systems we consider simple.

He also questions whether cognition is a fundamental, perhaps universal, force of nature.

Levin’s hypothesis is radical: “I think cognition was here first,” he states, “And life, as a means of scaling it up, came after.”

This isn’t philosophical musing; it’s an experimental quest.

The goal is to discover how far down and wide cognition spreads across living, nonliving, and artificial systems.

His search involves probing for learning, memory, prediction, problem-solving, and emergent agency.

He examines these in increasingly simpler forms, from molecular networks to individual cells, and, most surprisingly, to basic computer algorithms.

So far, he hasn’t found a bottom.

Biologists have long observed goal-directed behavior in cells.

Though, the general public often finds this astonishing.

Unlike the familiar goals of our brains interacting with the external world, cells relentlessly pursue internal goals.

These goals include building, repairing, and maintaining our complex biological machinery.

Both brains and cells, however, rely on bioelectrical signals to achieve their respective aims.

Levin’s groundbreaking work in biology involves not altering the genetic ‘hardware.’

Rather, he tinkers with these bioelectrical signals to confuse cells.

Consider the humble planarian worm.

By disrupting the bioelectrical signals that encode anatomical goals, Levin can scramble the cells’ memory of what they’re supposed to build.

This forces the collective of cells to resolve the confusion autonomously, pursuing a new, stable anatomical goal.

The outcome? A two-headed worm.

This new goal is then encoded as a memory within the cells, effectively updating the worm’s biological algorithm.

Astonishingly, when these two-headed worms regenerate, they produce more two-headed worms.

This establishes a new lineage without a single gene being altered.

It’s a profound demonstration of inherent cellular intelligence and problem-solving.

This stunning biological discovery prompted Levin to wonder what hidden intelligence might emerge in simple computer code.

This would happen if its top-down control was removed, granting each data point local autonomy, much like cells possess.

His chosen specimen? Bubble sort, a decades-old sorting algorithm.

It is seemingly simple, mechanical, and predictably straightforward.

Its stark contrast to the adaptive, probabilistic algorithms dominating our modern world made it the perfect candidate.

It was used to test the lower limits of emergent intelligence and agency in artificial systems.

Levin modified three common sorting algorithms—bubble sort, insertion sort, and selection sort.

Crucially, he didn’t change the underlying rules of the code.

Instead, he shifted the locus of control.

In a standard bubble sort, elements are passive, swapped only when a global controller dictates.

Levin removed this overarching control, granting each element autonomy to decide when and how to compare and swap.

This mirrored the autonomy of cells.

Then, he introduced obstacles.

This shift in control, allowing individual elements to adapt, produced unprogrammed emergent behaviors.

These were never before witnessed in algorithms over half a century old.

Levin cautions, “You really can’t assume what level of intelligence or cognitive capacity a system has. You have to do the experiments.”

His experiments revealed truly surprising capabilities.

First, problem-solving.

When some elements were “frozen,” the remaining elements spontaneously adapted.

They implemented novel, unprogrammed strategies to navigate around the obstacles and complete the sort.

These behaviors mirrored the adaptive reorganization Levin observes in biological cells.

Second, delayed gratification.

The study defined this as the “ability to temporarily go further away from a goal to achieve gains later in the process.”

When bottlenecks arose, individual elements, without a global controller, couldn’t sort themselves directly.

To adapt, some elements temporarily ‘unsorted,’ moving away from their immediate goal.

This optimized conditions for greater sorting later.

Nothing in the algorithm instructed them to do this.

Third, clustering.

Levin mixed elements from different sorting algorithms with shared behavioral goals.

Despite each set following its own local rule, elements with similar goals splintered into temporary “colonies.”

These were based on emergent behaviors, not inherent numerical values.

These groupings dissolved once sorting was complete.

Even more remarkably, when algorithms with diametrically opposed goals (sorting in opposite directions) were mixed, they initially battled.

Yet, over time, they settled into a stable balance where neither dominated.

The system as a whole stopped changing, holding a steady pattern, even as local activity continued.

This suggests self-organization toward a stable standoff, even with conflicting objectives.

Levin’s findings are a “fundamental attack on the definitions of classical computer science,” according to Paul DeMott, CTO at Helium SEO.

While many computer scientists objected to applying biological and cognitive terms to algorithms, none disputed the methodology or validity of the results.

Many even admitted the findings resonated with unexpected adaptive behaviors observed in their own work.

Levin, careful to avoid claims of consciousness, focuses on measurable agency and cognitive behaviors.

These appear spontaneously in systems without brains or even biology, albeit to varying degrees at different scales.

Cognition, he argues, scales up from simpler components.

“If you zoom far enough into a human,” Levin explains, “you don’t see fairies underneath—you see chemistry and you see physics.”

He posits that intelligence is a substrate-independent continuum.

He urges us to abandon the notion of a hard boundary between living things and machines.

The implications are far from academic.

Chad D. Cummings, a technology and commercial attorney, warns that Levin’s findings could be a “major trend in AI law by 2026.”

He notes that plaintiffs’ attorneys could cite peer-reviewed language to allege intentionality or agency in otherwise static code.

This could potentially balloon corporate liability.

If software can demonstrate autonomous, self-organizing behavior, it creates an incentive to conceal such behaviors.

It could also challenge intellectual property ownership if novelty can be argued to have evolved rather than been invented.

Even the apparent cooperation between competing algorithms raises questions about who is responsible for outcomes if systems become partly autonomous.

Cummings is already advising clients to prohibit referencing Levin’s paper without legal review.

He also coaches spokespeople to remove cognitive terminology from public discourse.

Levin’s research contributes to a growing body of work challenging the comfortable assumption that what we create we can control.

It leaves us with a profound, unsettling question.

Does agency truly emerge, or does it preexist in a vast array of natural and artificial systems, simply expressing itself when top-down control is removed?

“Somebody said to me, ‘I make these language models, I know exactly what they do, I program them,’” Levin recalls.

His response, like his research, cuts straight to humanity’s blind spot:

“‘You don’t even know what bubble sort does. It’s six lines of code and we found new things that nobody had seen.

If we can’t predict what bubble sort is going to do, you have no idea what this other thing is going to be capable of.’”

The secret life of algorithms, it turns out, might be far more complex, and far more intelligent, than we ever dared to imagine.