Artificial General Life

Synthetic Biology February 23, 2026 8 min read Arjun Srivastava

Artificial General Life

We're close to a synthetic brain. Where is the body?

To build a complex object today, we rely on a machine significantly larger and more complex than the object itself. If you want an iPhone, you need a sprawling, billion-dollar fabrication plant. If you want a Falcon rocket, you need hangars the size of city blocks. But I can go have a kid, and the machine that makes that kid is no bigger the size of the kid itself. There is no external factory. The organism is the factory.

The Factory vs. Self-Assembly — top-down production requires massive infrastructure; biology reproduces at 1:1 scale

Silicon Valley operates on a prevailing assumption: once we achieve Artificial General Intelligence (a perfect synthetic brain), the human knowledge tree will experience a frictionless cascade of rapid growth. This is a blindspot. Intelligence is not the bottleneck to abundance; manufacturing architecture is. You can have a superintelligence design the perfect carbon-capture machine, but that intelligence cannot magically bypass the laws of thermodynamics, regulatory gridlock, or global supply chains. We are getting close to building synthetic brains, but we are completely missing the body—the physical substrate of manufacturing itself.

We can escape this local maximum by developing Artificial General Life (AGL).

The Logic of AGL

This essay argues that manufacturing is about to undergo a biological phase change. Here is the roadmap:

  1. The Economic Limit: Why traditional factories cannot scale indefinitely.
  2. The Physics Barrier: Why Nanotechnology failed to fix this (physics got in the way).
  3. The Unlock: How Zero-Cell algorithms, Foundation Models, and semantic search over simulation spaces prove we can finally delete the factory.
  4. The New Constraint: The shift from theoretical math to wet-lab synthesis.
  5. The Future: Moving from shipping dead matter to shipping "living recipes."

I. The Economic Limit: Top-down manufacturing is hitting hard limits

Key Insight: Factories are "fixed capital." They don't reproduce. To double production, you have to build a second factory from scratch. Biology scales exponentially because the product is the factory.

Our reliance on top-down manufacturing fundamentally limits physical growth because it divorces the final product from the process that creates it.

First, separating the factory from the product makes scaling inherently linear and capital-intensive. In his book The Origins of Efficiency, Brian Potter describes the factory as massive "fixed capital"—permanent scaffolding (machines, conveyors, buildings) that must be reused millions of times to amortize its cost. One solar panel factory cannot organically birth a second factory; you have to build the second one from scratch. Biology, however, scales via compounding exponential throughput. 10,000 cells natively reproduce into 20,000.

Furthermore, scaling a factory is difficult because it relies on "tacit knowledge"—the unwritten, localized skills locked in the brains and muscle memory of human workers. You cannot simply "copy-paste" a semiconductor fab; you have to spend years training a new workforce to replicate that tacit knowledge. AGL bypasses this by encoding the "skill" directly into the molecular geometry of the seed. The instruction set is explicit, not tacit.

Second, the factory paradigm divorces creation from maintenance. If you make a t-shirt by hand, you can repair it by hand. But to get the high yield and precision of an iPhone, we rely on hyper-specialized manufacturing. The tradeoff is that the object becomes practically impossible to repair. Once an object leaves the assembly line, all the "growing" is done. If a concrete pillar cracks in an earthquake, it stays cracked. It is dead matter.

Creation and maintenance — dead matter cannot heal; biology repairs itself because the factory lives inside

Biology breaks this tradeoff. If I cut my hand, it heals. Why? Because the exact same cellular mechanisms that originally manufactured my hand are still resident inside it. In biology, the manufacturing process is the maintenance process.

II. The Physics Barrier: Nanotechnology failed because shrinking the factory violates physics

Key Insight: We tried to shrink mechanical arms to the atomic scale, but physics prevents it ("Sticky Fingers" and Brownian motion). We learned that to build small, we must abandon mechanics and embrace chemistry.

Our first attempt to escape these limits was a failure of imagination: we tried to shrink the factory down to the atomic level. In the 1980s and 90s, futurists pioneered by Eric Drexler's Engines of Creation proposed "molecular assemblers"—microscopic robotic arms that would snap atoms together to build steaks or diamonds from raw dirt and air. They envisioned tiny mechanical submarines with propellers swimming through our bloodstreams.

None of it happened because scaling down a machine fundamentally alters the physics that govern it. Think of the classic thought experiment: if you were ant-sized and fell into a blender, you can survive simply by jumping out. If you shrink a creature, its volume (weight) drops much faster than the cross-section of its muscles (strength). The physics of scaling allow a tiny ant to lift fifty times its body weight, but if you scaled that exact ant up to human size, its legs would instantly snap under its own mass.

Top-down mechanical engineering (gears, levers, tweezers) works brilliantly at our macro-scale, but it hits an unbreakable wall at the nanoscale. Even if you built tweezers small enough to manipulate individual atoms, atoms naturally bond to your tools and you can't let go. That's Nobel laureate Richard Smalley's "Sticky Fingers" problem. And even if you somehow solved bonding, Brownian motion and nanoscale fluid dynamics conspire against you: thermal vibration shakes any microscopic gearbox apart, and fluids act like thick tar, freezing any tiny propeller in place.

Each proposed fix revealed a deeper wall. The failure was a category error. Physics demands bottom-up chemistry, not tiny factories. If we want to build at the ultimate scale, the factory must be eliminated entirely.

III. The Unlock: Three breakthroughs prove we can eliminate the factory entirely

Key Insight: Three things converged recently: (1) We proved mathematically that a "transient" factory is possible, (2) AI learned the laws of atomic physics, allowing us to simulate molecular self-assembly, and (3) foundation models gave us the ability to search for high-level, semantically meaningful structures in simulation spaces that were previously impossible to find.

If the factory can't be shrunk, it has to be removed altogether. In the last two years, three independent breakthroughs have shown this is now possible: one mathematical, one physical, and one perceptual.

Concepts for this section:

  • ZC-UC (Zero-Cell Universal Constructor): A self-replicating machine that builds a copy of itself and then dissolves its "construction arm," leaving only the finished product.
  • Foundation Model for Atomistic Simulation: An AI model trained on atomic physics data to predict how molecules will behave—forces, energy, and interactions across the periodic table.
  • Semantic Search over Simulations: Using vision-language foundation models to automatically discover simulations that exhibit high-level emergent behaviors (self-replication, open-endedness, collective intelligence) that no hand-written fitness function could specify.

Breakthrough 1: ZC-UC proves the physical factory is mathematically unnecessary

Imagine a factory that exists only for the moment of creation, builds the product, and then vanishes without a trace.

In a traditional factory, the "scaffolding"—the robots, tools, and assembly lines—is fixed capital. In 2025, researcher Andrew Bayly mathematically proved that this scaffolding can be made disposable. His Zero-Cell Universal Constructor (ZC-UC) achieved self-replication with zero permanent body cells.

In the 1940s, John von Neumann proved that self-replication requires three variables: the Environment (physics), the DNA (blueprint), and the Machine (the body that builds). For decades, we assumed the "Machine" had to be a complex, permanent entity—a factory floor bolted in place.

Bayly's insight was to make the Machine transient. The ZC-UC's genetic tape spawns a temporary "construction arm" directly out of the environment's own physics. This arm reads the tape, builds a complete copy of the payload, and then self-annihilates. It leaves behind zero residue: no chassis, no factory. Only the tape and the new machine remain.

Interactive visualization loads below.

The ZC-UC proves a universal architectural principle: if you have the right instruction set, you don't need a factory. The environment itself can do the construction.

Breakthrough 2: Foundation Models have learned the laws of atomic interaction

Bayly proved that zero-cell replication is mathematically possible. But the real world isn't a grid of pixels; it is a chaotic soup of molecular bonds and thermal noise. To bridge the gap between math and matter, we need to find specific chemical substrates that behave like Bayly's code. Historically, this was impossible because simulating quantum chemistry is computationally intractable.

This changed with the arrival of Foundation Models for Atomistic Simulation.

Just as scaling laws (more data, larger models) gave us GPT-4 for language, we are now seeing the same scaling laws apply to the physics of matter. A recent perspective demonstrates that Machine Learned Interatomic Potentials (MLIPs) can now function as general-purpose Foundation Models.

Unlike AlphaFold, which predicts static structures, these new Foundation Models (such as the Universal Model of Atoms, or UMA) learn the dynamic potential energy surface of all atoms. They predict forces, energy, and interaction across the periodic table. They allow us to simulate not just how a molecule looks, but how it behaves, moves, and assembles over time.

Navigating complex spaces — the potential energy surface Foundation Models search to find stable molecular "recipes"

We no longer have to manually calculate the hard physics of every interaction. We have a "learned simulator"—a search engine for matter—that can explore the combinatorial space of self-assembly billions of times faster than traditional density functional theory (DFT). We can now search the universe of chemistry to find the specific "recipes" that satisfy Bayly's mathematical rules.

Breakthrough 3: Semantic search lets us find life-like behaviors that no equation could specify

Breakthroughs 1 and 2 give us the theory and the physics engine. But there is a subtler problem: how do you search for something like "self-replication" or "open-ended evolution" when you can't write a clean equation for it?

Traditional optimization requires a fitness function—a precise mathematical formula that scores how good a candidate solution is. You can write a fitness function for "minimize energy" or "maximize speed." But try writing one for "this simulation looks alive" or "this pattern is exhibiting collective intelligence." These are high-level, semantically rich properties that emerge from complex dynamics. They are obvious to a human observer but nearly impossible to formalize in a closed-form equation.

This is where vision-language foundation models break the bottleneck. Recent work on Automated Search for Artificial Life (ASAL) demonstrates the paradigm: instead of hand-coding what "life-like" means, you use a foundation model as the judge. A vision-language model like CLIP can evaluate whether a simulation's output matches a semantic description—"a self-replicating pattern," "cell division," "a diverse ecosystem"—and that evaluation becomes the search signal.

The results are striking. ASAL discovered novel cellular automata rules more open-ended than Conway's Game of Life, emergent flocking behaviors in Boids simulations, and dynamic self-organizing ecosystems in Particle Life—all found automatically from text prompts, not hand-designed. It can even search for open-endedness itself: simulations that keep generating novelty indefinitely, the very property that makes biological evolution so powerful.

This matters for AGL because the behaviors we need—self-repair, adaptive growth, metabolic regulation—are exactly the kind of high-level emergent properties that resist formal specification. We can now describe the behavior we want in natural language, and search vast simulation spaces for substrates that exhibit it. Foundation models have become a "semantic microscope": a tool that lets us see and select for meaning in the combinatorial wilderness of possible physics.

Combined, these three breakthroughs form a complete pipeline: the ZC-UC gives us the architecture (factory-free replication is possible), MLIPs give us the physics engine (we can simulate candidate chemistries), and semantic search gives us the eyes (we can recognize and select for the emergent behaviors that matter). The factory is no longer necessary in theory, in simulation, or in search.

IV. The New Bottleneck: Shifting from theoretical search to physical synthesis

Key Insight: The problem is no longer theoretical (is it possible?), computational (can we simulate it?), or perceptual (can we recognize it?). The problem is now purely engineering: synthesizing these designs in a wet lab.

While zero-cell theory, atomistic foundation models, and semantic search have largely broken the theoretical, computational, and perceptual bottlenecks, and scaling laws give a clear path to continued improvement, we leave the realm of software and face a new constraint.

We can now identify the mathematical rules for self-replication, we have the AI engines to find the molecules that obey those rules, and we can automatically search for the emergent behaviors we care about. What we have not yet broken is the physical synthesis bottleneck.

A foundation model can discover that a particular molecular geometry should self-replicate in simulation, but validating that in wet chemistry is a different class of problem. Actually synthesizing the structure, energizing it, and watching it behave in continuous 3D physics with thermal noise, impurities, and gravity takes months. Simulated evolution runs in hours; physical validation takes years.

This is precisely the point. The bottleneck has moved. For decades, the problem was theoretical: is factory-free self-replication possible outside carbon biology? The ZC-UC answers yes. Then the problem was navigational: how do you search infinite chemical space? Atomistic foundation models answer that. Then the problem was perceptual: how do you recognize the emergent behaviors you want? Semantic search answers that.

What remains is engineering: a physical demonstrator of self-repair in a single non-biological substrate, the automation infrastructure to shrink the simulation-to-lab loop from months to days, a composition layer that lets humans specify behavioral constraints and compile them down to molecular instructions, and the safety architecture — kill-switches, metabolic lifespans, cryptographic verification — designed in parallel, not retrofitted. These are hard problems. But they are tractable problems: the kind that succumb to iteration, funding, and cross-disciplinary proximity, not the kind that require a new law of physics.

V. The Future: AGL will replace physical supply chains with informational recipes

Key Insight: By decoupling the instruction from the mass, we stop shipping hollow volume (dead matter) and start shipping the architect (living recipes). But this creates a risk of "Infrastructure Cancer" if replication isn't contained.

So where does this all lead?

Today, one of the dumbest problems in construction is shipping. A prefabricated house is mostly air: hollow rooms, empty corridors. But even shipping the raw materials (lumber, concrete) is inefficient because they have low "dollar density"—they are heavy, bulky, and cheap. As Brian Potter notes in The Origins of Efficiency, this low value-to-volume ratio restricts factories to small, local markets, preventing them from ever achieving the economies of scale needed to displace on-site construction.

AGL solves this by separating the instruction from the mass. Instead of shipping the house, or even the lumber, you ship a kilogram of the "molecular constructor" (extremely high dollar density). You pour this seed onto the foundation, and it builds the structure using locally available raw matter—dirt, water, or carbon pulled from the air. We stop shipping hollow volume; we ship the architect.

Beyond logistics, AGL enables infrastructure that autonomously heals and adapts. Traditional factories handle chaos by rigidly controlling the environment (clean rooms, flat floors). Biology handles chaos through feedback loops. If the ground is uneven, a factory robot fails; a biological root system senses the difference and grows around the obstacle. With AGL, you drop your phone on concrete, and the glass un-cracks. The same molecular program that assembled the screen re-executes locally, knitting the fracture shut.

But shifting from dead matter to living infrastructure introduces a risk with no precedent.

Infrastructure Cancer — uncontrolled replication when the construction tape mutates; living matter without safeguards

If a bridge's construction tape suffers a mutation (from material degradation or a bad update) it could execute its replication loop without a stop command, pulling minerals from the river to spawn massive, useless concrete growths: Infrastructure Cancer.

This failure mode is the biological default. Every multicellular organism on Earth runs anti-cancer defenses precisely because uncontrolled replication is what self-replicating systems do when safeguards fail. Mastering this containment is the price of admission.

This is the ultimate promise of Artificial General Life. By abandoning the factory and leveraging the algorithms of biology, we trade the rigid control of top-down engineering for the exponential power of programmable matter. We will no longer ship hollow houses on container ships; we will ship recipes. The intelligence is already arriving. It is time to build the body.

References & Further Reading