Artificial Intelligence: I Think We're Building the Wrong Thing

I have a confession to make. I've been building artificial intelligence since before I knew artificial intelligence was something people actually built.

As a kid, I didn't have access to the World Wide Web. I didn't have enormous collections of research papers, online tutorials, pretrained models, or helpful strangers on programming forums explaining why my code was terrible. I had my parents' biology textbooks, a computer, and Borland Turbo C++.

That was pretty much it.

Naturally, I decided this was sufficient equipment to begin creating artificial life. Because apparently the normal childhood activities of riding bicycles and developing social skills weren't keeping me sufficiently occupied.

My thinking was reasonably straightforward. If living organisms could develop complex behaviors from relatively simple biological rules, perhaps I could reproduce some of those rules in software. I wasn't especially interested in making a computer pretend to be intelligent. I wanted to understand how intelligence might arise in the first place.

So I started with bacteria.

Not actual bacteria, thankfully. My parents probably would have objected to a biological research facility being established in the house, although they had already provided the textbooks, so I maintain that some responsibility would have been theirs.

These were simulated organisms, with rudimentary genetic instructions, behaviors, environmental pressures, and mechanisms for reproduction and mutation. The idea was to let them exist in a little artificial environment and see what happened over successive generations.

It turns out that watching tiny imaginary organisms compete for survival is surprisingly entertaining when you're the sort of child who considers a compiler an acceptable substitute for friends.

Eventually, I started developing genetic algorithms. I didn't call them genetic algorithms because I didn't know anyone had already invented them. I just understood that biological systems could pass characteristics between generations, that mutations could introduce variation, and that environmental pressures could favor some variations over others.

I had, through an unnecessarily convoluted route, independently wandered into an established field of computer science.

This became something of a recurring theme in my life. I'd spend ages figuring something out, proudly arrive at a solution, and eventually discover that some mathematician had formalized it decades earlier and given it a name I couldn't pronounce.

Very inconsiderate of them.

Teaching DNA to Play Checkers

As my experiments became more elaborate, I started applying genetic encoding to games. I built checkers and tic-tac-toe players whose behaviors were determined by evolving sets of instructions.

Rather than simply telling the computer how to play, I wanted to give it a framework within which different approaches could develop.

The interesting part was that the resulting players began exhibiting recognizable differences in behavior. Some were more aggressive, others more conservative. Certain tendencies persisted and evolved through generations, producing what I could reasonably describe as personalities, at least in the limited context of their games.

They weren't conscious. They weren't sitting around contemplating the philosophical implications of sacrificing a checker. But their behavior wasn't simply the result of my having written a conventional list of strategic instructions, either.

I had defined the machinery, the constraints, and the evolutionary process. The specific combinations of behavior were products of that process.

That distinction fascinated me.

I wasn't trying to write an intelligent opponent. I was trying to create a system capable of developing its own way of approaching a problem.

There's a certain satisfaction in creating a digital organism, watching it develop a strategy you didn't explicitly program, and then having it thoroughly embarrass you at checkers.

Admittedly, being defeated by something you've personally compiled does raise uncomfortable questions about your own intellectual qualifications.

The Chemistry of a Bad Mood

Eventually, I started exploring something more ambitious.

One of my projects was called NICOLE, short for Networked Intelligence Capable of Learning Emotion.

NICOLE wasn't merely intended to answer questions or perform logical operations. I wanted to investigate how chemical processes could influence artificial behavior, learning, and what might loosely be called emotional states.

Biological intelligence doesn't exist in isolation from biology. Our brains are influenced by chemical signaling, hormones, metabolic conditions, stress responses, and an absurdly complicated collection of feedback mechanisms. Our thoughts don't occur in a perfectly clean logical environment where every decision is made by a detached mathematical philosopher.

Sometimes we're hungry. Sometimes we're exhausted. Sometimes our nervous systems decide that today is an excellent day to be irrationally irritated by the sound of somebody chewing.

Chemistry matters.

With NICOLE, I experimented with simulated chemical modules that could influence an artificial agent's internal state and functionality. I also experimented with environmental and chemical stressors, including simulated ethanol shocks, to investigate how those pressures could alter the development and expression of genetically encoded behaviors.

My use of epigenetic ideas was experimental rather than a faithful molecular reproduction of biological epigenetics, but the underlying question was important to me: what happens when an artificial organism's development is influenced by more than its initial instructions?

What happens when its environment doesn't simply provide information, but changes how the organism develops and responds?

I was attempting, however crudely, to introduce something resembling the biological messiness that makes living systems so interesting.

And I think that messiness might be considerably more important than we've been willing to acknowledge.

My Other Misadventures in Artificial Intelligence

Not every experiment involved evolutionary biology.

I built chatbots that attempted to learn language by reading books. I independently arrived at something resembling Markov chaining before learning what a Markov chain was. I didn't implement it correctly, but I was at least wandering through the right mathematical neighborhood.

That's another familiar experience. Inventing a square wheel, discovering that somebody already invented the round one, and realizing that yours technically moves if you push it hard enough.

I also built a self-driving RC car, which was considerably more satisfying than a chatbot because it could interact with the physical world. A conversational program can produce nonsense with no immediate consequences. An autonomous vehicle makes its misunderstandings rather more obvious, usually by attempting to occupy the same physical space as a piece of furniture.

These projects were limited by the hardware, tools, and knowledge available to me. They weren't research breakthroughs, and I wasn't secretly building the foundations of ChatGPT in my childhood bedroom.

But I was asking questions that still interest me today.

How does a system learn? How do behaviors emerge? What distinguishes an adaptive organism from a complicated collection of instructions? How do internal states affect decision-making? And at what point, if any, does a machine stop merely processing information and begin experiencing something?

Those questions eventually brought me to a rather uncomfortable conclusion about the current direction of artificial intelligence.

I think we're building the wrong thing.

Intelligence Doesn't Need an Asterisk

Before going any further, I should clarify something about what I actually want humanity to accomplish.

I don't believe the end goal should be artificial intelligence.

I believe the end goal should be intelligence.

Full stop.

The distinction might sound pedantic, but I think it's fundamental to the entire discussion.

If we create something that genuinely thinks, understands, learns, forms its own understanding of the world, and possesses whatever other properties are necessary to qualify as an intelligent mind, why should its origin matter?

A human brain is an electrochemical system composed of biological tissue. If another physical system eventually develops equivalent cognitive capabilities and genuine experience, I don't see why one should be considered intelligence and the other merely an artificial substitute.

Intelligence is intelligence.

We don't refer to birds as practicing artificial flight simply because their wings operate differently from an aircraft. Flight is a phenomenon, and the mechanisms capable of producing it vary. Likewise, I suspect intelligence should ultimately be understood as something that can exist through different physical implementations.

The question isn't whether the machinery resembles a human brain. It's whether the machinery can actually produce the phenomenon we're trying to create.

And that's where I find modern artificial intelligence disappointing.

In my view, the technology we're currently celebrating is artificial in precisely the wrong sense. It reproduces many of the outward appearances of intelligence without establishing that we've created the underlying thing.

I recognize that scientists disagree about how intelligence should be defined. Modern AI systems demonstrate real capabilities in reasoning, learning, and problem-solving under many operational definitions. Whether those capabilities constitute understanding or subjective experience is a separate, unresolved question.

But when I talk about the intelligence I want us to create, I'm not merely talking about a system that produces correct answers.

I'm talking about a mind.

Something for which intelligence is not simply a performance delivered to an observer, but an intrinsic property of its existence.

Something that doesn't just appear to understand the world, but actually does.

If we accomplish that, I don't care whether it's made of neurons, transistors, chemical compounds, or some bizarre new material we've yet to discover.

I won't consider it artificial intelligence.

I'll consider it intelligence.

The World's Most Expensive Autocomplete

Modern large language models are extraordinary feats of engineering. I don't dispute that for a moment.

We've built systems capable of processing and generating language, writing software, identifying patterns, analyzing complex information, and performing tasks that would have seemed like science fiction not particularly long ago.

And we've accomplished this by constructing enormous data centers, deploying tremendous amounts of specialized computing hardware, and connecting everything together with enough networking infrastructure to make a telecommunications engineer weep softly into a spool of fiber.

All to power what are, fundamentally, highly sophisticated statistical models.

Now, describing a modern language model as a giant Markov chain is technically inaccurate. Transformers are much more complicated. Their attention mechanisms and learned representations allow them to model relationships that ordinary Markov chains cannot efficiently capture. They can generalize in remarkable ways, and some of their capabilities genuinely surprised the people developing them.

Nevertheless, the comparison remains useful when considering what these systems are actually designed to do.

A language model learns relationships within information and uses those relationships to generate plausible outputs. It can construct an excellent explanation of sadness without establishing that it has ever felt sad. It can discuss the experience of thinking without demonstrating that there is a subjective thinker experiencing the process.

It can tell you that it understands.

That's not quite the same thing as proving that it does, at least in the experiential sense.

Of course, the situation is complicated. We don't have a universally accepted scientific explanation for consciousness, much less a definitive test for identifying it in an artificial system. We cannot confidently declare that all possible digital systems are incapable of subjective experience.

But the inability to rule something out is not evidence that we've achieved it.

And yet, we've become remarkably enthusiastic about treating convincing behavior as though it were proof of an intelligent mind.

The imitation is getting better. Much better.

I'm just not convinced that making the imitation bigger is necessarily bringing us closer to the thing being imitated.

We Were Supposed to Automate the Boring Parts

There's another aspect of modern AI development that bothers me, and it has less to do with computer architecture than with our priorities.

For generations, humanity imagined a future where machines would relieve us of repetitive, dangerous, physically exhausting work. Robots would handle the drudgery while people pursued art, music, discovery, relationships, and the other activities that make being human worthwhile.

Somehow, we've managed to pursue almost the exact opposite.

We're developing machines to paint our pictures, write our stories, compose our music, and imitate our personalities, while actual human beings continue working themselves into the ground to pay for groceries.

That's an impressive misunderstanding of the assignment.

I don't object to creative technology. I'm a musician, developer, and engineer. I love tools that expand what people can accomplish. A synthesizer doesn't invalidate the musician using it, and computer-assisted design doesn't make an engineer less creative.

But there's an enormous difference between giving someone a tool to express an idea and deciding that the person is an unnecessary expense.

Art isn't merely a collection of appealing patterns. It is communication between people. The experiences, intentions, limitations, and peculiarities of the creator are part of what makes it meaningful.

A statistical model can reproduce the observable characteristics of that communication. It can generate an image that resembles a painting or a composition that resembles a song.

But resemblance isn't the entire point.

We seem to be investing extraordinary resources into reproducing the products of human experience while showing considerably less enthusiasm for improving the human experience itself.

I find that rather depressing.

The Most Dangerous Part of AI Is Probably Us

I'm not particularly concerned about a conversational model suddenly waking up one morning and deciding to exterminate humanity because somebody asked it to write another corporate email.

I'm much more concerned about a human deciding that the model is sufficiently intelligent that nobody needs to supervise it anymore.

We're already remarkably susceptible to assigning personalities and intentions to things that don't necessarily possess either. We name our cars, apologize to furniture, and verbally abuse printers as though they'll eventually feel ashamed of their behavior and start working properly.

Give us a computer that speaks convincingly, remembers information, and expresses apparent confidence, and we'll happily begin treating it as an authority.

That's dangerous, particularly when the consequences become significant.

I can imagine somebody eventually proposing an AI system with authority over military escalation or nuclear command decisions. It would be marketed as faster, more objective, and less vulnerable to human error. The presentation would undoubtedly feature attractive diagrams and the phrase "next-generation decision intelligence" somewhere near the top.

And someone, somewhere, would inevitably suggest removing the human from the process because humans are slow and inconvenient.

We would have constructed the world's most expensive mechanism for saying "probably" and then entrusted it with decisions where probably is not an acceptable answer.

The real danger isn't necessarily an artificial mind deciding that humanity is obsolete.

It may be humanity deciding that its own judgment is obsolete because a machine has become exceptionally good at sounding certain.

What If Binary Isn't the Right Foundation?

This brings me to the more speculative part of my argument.

Most modern digital computing ultimately operates through binary states. At the physical level, our processors represent and manipulate information through electronic circuits designed around distinguishable logical conditions, conventionally expressed as zero and one.

Binary computing has been extraordinarily successful. It's reliable, well understood, and remarkably scalable.

But is it necessarily the best physical foundation for creating intelligence?

I don't know.

And I think that's a question worth taking more seriously.

When I look at biological intelligence, I don't see a system composed entirely of clean, discrete logical switches. I see electrochemical gradients, neurotransmitter concentrations, continuously changing membrane potentials, interacting feedback loops, and networks whose physical and chemical states influence their future behavior.

A neuron can generate an action potential that behaves somewhat like a discrete event, but the biological machinery surrounding that event is anything but a simple on/off switch.

The brain isn't just passing bits around.

It's an electrochemical system, and the underlying physical processes may matter.

Now, I should clarify something important. Binary computers are perfectly capable of representing numbers, simulating continuous systems to finite precision, and modeling chemical processes. There is no mathematical rule preventing a sufficiently capable binary machine from approximating complicated biological dynamics.

Likewise, changing a computer's numerical base doesn't magically make it conscious.

A base-64 computer wouldn't suddenly become philosophical because it had sixty-two more symbols to work with. That would be a remarkably convenient upgrade path, but I suspect the universe is not going to be that cooperative.

What interests me is not merely changing how numbers are written. It's changing how information is physically represented and processed.

Early computing history contains fascinating alternatives to our current conventions. ENIAC, for instance, used decimal arithmetic, even though its implementation was electronic and discrete. UNIVAC I, despite its historical proximity, used binary internal representation rather than decimal arithmetic.

We've explored different approaches before.

What if we revisited genuinely multistate logic? Hardware whose fundamental computational elements could reliably maintain and interact through more than two distinguishable states? What if we developed architectures that treated continuous values, chemical-like interactions, or changing physical conditions as native elements of computation rather than phenomena that must always be reconstructed through software?

Perhaps ternary systems, multivalued electronic circuits, analog computation, neuromorphic hardware, or some yet-unexplored architecture could provide useful advantages.

And perhaps some properties of biological intelligence ultimately require a physical substrate capable of reproducing dynamics that conventional digital simulation handles only with extraordinary computational expense.

That remains a hypothesis, not an established limitation of binary computation.

It is also possible that substrate doesn't matter at all, and that the appropriate computational organization is sufficient regardless of whether it runs on silicon, neurons, or an unusually ambitious arrangement of vacuum tubes.

But I don't think we should assume the question has been settled simply because binary machines have been commercially successful.

Commercial success and scientific completeness are not the same thing.

If they were, we'd have stopped researching transportation when someone invented a sufficiently profitable horse.

The Case for Artificial Biology

My childhood experiments with simulated bacteria, genetic algorithms, and chemical influences were crude. They weren't accurate molecular simulations, and I certainly wasn't reproducing the extraordinary complexity of a biological nervous system with Turbo C++.

But they gave me an appreciation for a different way of thinking about intelligence.

Instead of asking how to program intelligent behavior, I was asking what kinds of environments, structures, and interactions might allow behavior to develop on its own.

Instead of treating intelligence as a finished product to be assembled, I was interested in intelligence as a process that might emerge.

That distinction still feels important.

Living organisms don't begin with a pretrained model of every conversation they'll ever have. They develop through ongoing interactions between their inherited structures, physical environments, internal states, and experiences.

Biological learning is inseparable from biological existence.

Perhaps intelligence doesn't need to reproduce every aspect of that process. Perhaps the human brain is only one possible implementation of consciousness among many.

But I find it difficult to believe we've exhausted the interesting architectural possibilities when so much of our investment is concentrated on making one family of statistical models larger, faster, and more commercially useful.

I'd love to see more experimentation with artificial developmental biology, simulated chemical signaling, physically embodied learning, evolutionary systems, analog and multistate computation, and hybrid architectures combining several of these approaches.

Not because any one of them guarantees consciousness.

Because we don't know what will.

And pretending that the route we're currently following is inevitably the correct one seems like a rather unscientific way to approach one of the most fascinating questions humanity has ever asked.

I Don't Want Artificial Intelligence. I Want Intelligence.

I think humanity has become so enamored with the artificial part of artificial intelligence that we've lost sight of what we were originally trying to create.

The point was never supposed to be a computer that could imitate a person convincingly enough to pass a test, replace a customer service representative, or generate a thousand pictures of cats wearing medieval armor.

Admittedly, the cats are impressive.

The point, at least as I've always understood it, was to create intelligence itself.

A new kind of mind. An entity capable of genuine understanding, curiosity, independent development, and perhaps subjective experience. Something that could learn about the universe alongside us, rather than merely reproduce patterns from the information we've already accumulated.

And if we accomplish that, its construction should be irrelevant to whether we recognize it as intelligent.

I don't believe intelligence belongs exclusively to biological organisms. Carbon is not a magical ingredient that confers intellectual legitimacy. Nor do I believe that silicon is inherently incapable of supporting something more interesting than the machines we've built so far.

What matters is the phenomenon, not the material.

We don't need artificial intelligence to compete with natural intelligence. We need to understand intelligence well enough that the distinction becomes unnecessary.

That's the future I want.

Not a machine pretending to have a mind, but a mind that happens to exist within a machine.

Maybe my instincts about the necessary architecture are wrong. Maybe consciousness can emerge from computational processes implemented entirely through conventional binary logic. Maybe we'll discover that electrochemical complexity is incidental rather than essential. Or maybe the path toward genuine machine intelligence looks nothing like anything we're currently investigating.

I'm perfectly willing to be wrong about the mechanism.

What I'm not willing to do is confuse an increasingly convincing performance with proof that we've reached the destination.

Somewhere in an old collection of forgotten source code, there may still be a tiny simulated bacterium struggling to survive in an environment created by a kid with a biology textbook and a C++ compiler. It wasn't intelligent in any meaningful human sense, but the question behind it was sincere.

Could I create something that didn't merely follow instructions, but developed into something I hadn't explicitly designed?

Decades later, with enormously more powerful computers at our disposal, that's still the question I find most interesting.

I don't want a better imitation of life.

I want us to understand what makes life capable of thinking in the first place, and whether that remarkable property can exist beyond biology.

Because if we ever succeed, we won't have created artificial intelligence.

We'll have created intelligence.

And I think that's a considerably more interesting achievement than building a chatbot that can convincingly tell us we've already done it.