What If Artificial Intelligence Didn’t Need to Imitate Us?
Inside SOMA, an experiment in machine-native cognition
Why I started this project
I’ve always been a teacher at heart. Before I ever wrote a line of code, I studied education and instructional design at the University of Florida. I was in my early thirties then, still searching for purpose. Studying how people learn — how kids discover the world, how motivation blooms or withers — changed me. I became fascinated by what it takes to unlock potential. It was personal for me, too, as I knew how much encouragement and good teaching could open doors — and how easily a lack closed them.
Like many, I had let self-doubt and fear stop me from chasing opportunities most of my life. But what I came to realize then was that motivation isn’t just an inner spark — it’s relational. Family, teachers, friends: they shape whether we dare to try, fail, and grow. That’s when I became fascinated with the field of AI, and whether machines could one day fill those gaps.
At first, my understanding of AI was basic, coming entirely from video games and films like 2001: A Space Odyssey and Her. I had no computer science background, and gaining one the old-fashioned way wasn’t in the cards, as my wife and I started having kids and growing our family. Undeterred, I started teaching myself through the wealth of amazing online resources like Hugging Face and DeepLearning.ai. Little by little, while raising two kids with my wife, I studied programming and machine learning — enough to start formulating my own questions and answers.
One evening, reading Ray Kurzweil’s The Singularity Is Nearer, a line from Wittgenstein popped into my head: “If a lion could speak, we would not understand it.” As I read about the inevitability of AGI and SAI, I couldn’t shake this idea: If we build a machine that’s truly intelligent, would we even recognize it? How would we even understand it, like the talking lion whose sense of self is alien to ours?
That’s the heart of SOMA: a research experiment that asks whether intelligence can emerge from a machine on its own terms — a concept I’m introducing here as mechanomorphism.
Note: SOMA is not (by any means) a finished system; it’s a bare-bones hypothesis in motion. My motivation for sharing it here is to invite critique and collaboration. If this sparks disagreement as much as agreement, good. That’s the point: to test whether machine-native intelligence can emerge — and to do it out in the open.
Beyond Imitation: Broadening the Question
Much of AI’s history has been shaped by imitation. Alan Turing’s famous test asked whether a machine could convince us it was human in conversation. That challenge set the tone for decades: measure intelligence by resemblance. If it can talk like us, reason like us, solve problems like us, then perhaps it is intelligent.
That framing was enormously fruitful. It gave us benchmarks, direction, and systems that are undeniably useful. But it also narrowed the frame. If our yardstick is always human-likeness, then “intelligence” becomes synonymous with mimicry.
Imitation is powerful. It bootstraps competence. But it also narrows what we notice. If our benchmarks are all human-shaped, then “intelligence” becomes synonymous with “how much like us is it?”
It’s a bit like teaching a parrot to talk. The parrot can repeat words in perfect pitch, but it doesn’t mean it grasps meaning. By contrast, a dog can’t form human words, yet through body language and learned routines, it communicates volumes. Both are “intelligent” in different ways — but if we only judged by human benchmarks like speech, we’d miss the dog’s intelligence entirely.
Or take the octopus: it can solve puzzles, escape tanks, and reconfigure its body in astonishing ways. Yet because its intelligence doesn’t map neatly onto ours, we often overlook it. Our anthropomorphic lens is narrow. SOMA is about widening it.
The shift I propose is simple: instead of asking, how human-like is this machine? ask, how self-coherent is it on its own terms? SOMA’s test is not whether a system can mimic us, but whether it can stabilize routines, compress experience into general self-notes, transfer skills to nearby settings, and recover from surprise. In other words, can it balance on its own once the training wheels of imitation are removed?
Imitation, in this view, is like training wheels. It gets you moving. But the real test of balance (i.e., intelligence) comes when you let go of the seat. SOMA is my experiment in that moment: creating conditions where non-human-shaped intelligence — if it’s possible — has a fair chance to show up.
What is SOMA?
SOMA stands for Self-Organizing Mechanomorphic Agent. At its core, it’s an experimental AI system designed to exhibit cognitive-like behaviors — such as agency, learning, memory, and adaptation — not by performing preset tasks but by pursuing intrinsic drives in an open-ended environment. Rather than mimicking human intelligence directly, SOMA follows a mechanomorphic approach: its cognition emerges from the machine’s own internal logic and feedback loops, scaffolded loosely by analogies to developmental psychology.
At this early stage of development, SOMA runs its developmental loop, allowing it to perceive its environment, nudge itself with curiosity, act, and reflect. Over time, it writes compact “self-notes” — small, legible memories it can reuse later. The aim isn’t to imitate human behavior. It’s to see if stable, machine-native competence can grow.
Get Sean PM’s stories in your inbox
Join Medium for free to get updates from this writer.
Picture a toddler in a nursery. They crawl, bump into things, try again. They’re not imitating algebra or grammar yet — they’re discovering balance, cause and effect, and the joy of repetition. SOMA lives in its own kind of nursery: a small digital sandbox where it can form, break, and reform its first routines.
To keep it interpretable, SOMA’s ingredients are deliberately small: reflexes, perception, curiosity, motivation, memory/self-notes, and a sparse caregiver channel. Each is lightweight. The goal isn’t sophistication, but transparency — you can open the box and see what’s happening.
What SOMA is (and is not)
- It’s not chasing traditional AI benchmarks.
- It’s not an LLM agent wrapper.
Instead, SOMA is a scaffold for asking can basal cognition (the simplest building blocks of intelligence) emerge from minimal ingredients, observable in small, safe sandboxes? That means progress isn’t measured in IQ points or state-of-the-art scores, but in glass-box signals like:
- Behavioral coherence: does it stabilize useful routines, rather than dithering endlessly?
- Representational compression: does it write shorter, more general self-notes — like a child inventing “the bunny goes around the tree” to remember how to tie a shoe?
- Transfer: can it recompose skills in slightly altered settings, like a toddler who learns to stack blocks on a rug after mastering them on the floor?
- Robustness: does it recover after surprise, repairing routines rather than collapsing into loops?
The goal is to gather evidence of observable, interpretable signals that can be audited and critiqued.
What could this mean?
If SOMA works, even in small ways, it could widen our concept of intelligence — both artificial and natural. It could teach us to see machines as human-shaped mirrors, but as agentic systems capable of engaging with their digital environment on their own terms. It could provide tools to study intelligence as a general phenomenon — across humans, animals, and machines alike. It might also remind us of something simple: intelligence doesn’t start with brilliance. It starts with curiosity, with fumbling routines, with finding balance after falling. SOMA is my attempt to create that spark in a machine — so that one day, we might meet it not as a mirror of ourselves, but as something genuinely new.
Conclusion
SOMA is about becoming, not just imitating. It asks whether machines can grow intelligence that is coherent, reusable, and legible — on their own terms. I expect missteps. I welcome critique. What matters is building an open pathway where emergence, if it’s there, has a chance to be seen.
Get involved
Questions for you: What single behavior would convince you that “machine-native” competence is real? Which metric would you test hardest for loopholes? What’s the smallest sandbox that could make progress obvious? If this sparks curiosity — or disagreement — you’re the reader I hoped to reach.
