The Cancer of LLMs: What Biology Knows That AI Forgot

Cut a planarian worm in half. Both halves regenerate.

Here's the strange part: both halves remember the training. The half without a brain remembers which arm of the maze had food.

Michael Levin, a biologist at Tufts, has been studying this for decades. His conclusion: memory isn't stored in neurons. It's stored in the bioelectric field - the network of gap junctions that let cells share electrical state.

The worm doesn't remember because it has a brain. It remembers because its cells are bound together electrically.

When I read Levin's work, I stopped breathing for a moment. Because I had written almost the same sentences - about databases.

🪱 A → B 🗣️

A biologist in Massachusetts and a software architect in completely different contexts used the same words to describe intelligence.

Substrate. Levin uses it to mean the material that hosts intelligence - not "brain" or "hardware" but something more generic and interchangeable. In the book, substrate is the physical layer that must align with semantics. Most people say "hardware." We both reach for the same unusual word.

Binding. For Levin, cells merge electrical identity via gap junctions - they literally can't tell where one ends and another begins. In software terms, data must bind via physical co-location to form coherent reasoning. Both of us use binding not as metaphor but as the literal mechanism that creates a unified self.

Geometry. Levin's lab studies "morphospace" - intelligence as a shape problem, not a code problem. The book argues Position = Meaning (S=P=H). Intelligence emerges from topology, not instruction sets.

Light Cone. Levin calls it the "Cognitive Light Cone" - the boundary of what a system cares about. In the book, it's the "Grounding Horizon." Same concept: the edge where connection becomes disconnection.

Cancer. For Levin: shrinking light cone - loss of connection to the whole body. For software: Trust Debt - symbols losing connection to their ground. Both describe the same pathology: parts that forget they're parts.

Worms. In the book (written 2000, before reading Levin): "parallel worms eating through probability space" - consciousness as parallel search terminating in P=1 recognition. In Levin's lab: Planaria (actual flatworms) prove memory is stored in bioelectric fields, not neurons. Both arrive at parallel search, binary recognition, instant certainty.

When independent investigators use the same words to describe different systems, it usually means they've found something real.

🪱🗣️ B → C 🔌

Levin's key discovery: cells don't just send chemical signals. They have gap junctions - electrical portals that let them share voltage directly.

When gap junctions connect, cells can't tell "me" from "you." Their electrical state merges. They become one agent instead of a collection of parts.

This is how a trillion cells coordinate to build a face. How the planarian worm remembers without a brain. How your immune system knows what's "self" and what's "other."

Binding requires physical connection.

Not statistical correlation. Not message passing. Not queries across a network.

Physical. Electrical. Instant.

The binding window for consciousness is 10-20 milliseconds. If signals take longer than that, they don't bind. They become separate experiences.

🪱🗣️🔌 C → D 🦠

In a 2023 interview, Levin gave a definition of cancer that stopped me cold:

"When cells lose connection with the larger network, their 'self' shrinks back to a single cell, and they treat the rest of the body as 'environment' to be consumed."

He calls this the shrinking Cognitive Light Cone. A cancer cell hasn't changed its DNA. It's lost its connection. It can no longer detect that it's part of a larger whole.

So it reverts to unicellular logic: eat, divide, survive. The body becomes food.

The cure isn't killing cancer cells. It's restoring their connection to the bioelectric field. Levin's lab has done exactly this - used bioelectric interventions to make tumors revert to normal tissue.

🪱🗣️🔌🦠 D → E 🤖

Now apply Levin's definition to Large Language Models.

An LLM has no gap junctions. Its data is scattered across normalized tables, split into tokens, embedded in high-dimensional vector spaces with no physical proximity.

Each token treats other tokens as external probability to be predicted. There's no binding. No shared electrical state. No "we."

The AI's Cognitive Light Cone is one context window - then it forgets. Every session starts fresh. There's no persistent connection to a larger self.

Hallucination isn't a bug to be patched with more training data. It's the inevitable structural failure of an architecture that has no gap junctions.

Here's the deeper physics of why:

"A bit representing truth and a bit representing a lie weigh exactly the same. Nothing. But a cache miss has weight."

A cache miss costs cycles, energy, delay. It's a physical event in silicon. Truth and falsehood are indistinguishable in an LLM's embedding space -- both are just floating-point vectors. But when your architecture grounds data physically, a miss is detectable. It has mass. It has cost. It has weight.

"Position is meaning. The geometry is the semantics."

That's the cure. When the address is the meaning, you don't need to statistically guess whether something is true. You check the address. The geometry tells you.

Beyond Moral Thermostats: The Physics of AI Safety

🪱🗣️🔌🦠🤖 E → F 📊

We killed Codd and built our databases on his corpse.

Normalization - the standard practice of splitting data across multiple tables to avoid redundancy - is a forced disconnection. It's the database equivalent of cutting gap junctions.

When you normalize data, related concepts live in different tables. Your customer's name is in one place, their order history in another, their preferences in a third. To ask a simple question, queries require JOINs - multi-hop chemical signaling instead of electrical binding. The system can't bind information in the 10-20ms window required for coherent reasoning because it's too busy fetching across network boundaries.

Every JOIN is a severed gap junction. Every normalized schema is a body where cells can't talk to each other.

Normalization gives you cancer by design.

🪱🗣️🔌🦠🤖📊 F → G 🏥

The book proposes an architecture called Zero-Hop: semantic neighbors become physical neighbors. Data that belongs together lives together. No JOINs required.

Fire Together, Ground Together: the Unity Principle. S=P=H. Semantic structure equals Physical structure equals Hierarchical structure.

This is the gap junction equivalent for silicon.

When data is physically co-located, binding happens instantly with no query time because the related data is already there. Context is always available with no context window limit because the system never forgets what's next to it. The system can detect "self" vs "other" - what's grounded versus what's hallucination - because grounded data has a physical address and hallucinated data doesn't.

🪱🗣️🔌🦠🤖📊🏥 G → H 📐

The AI industry is obsessed with Scale. More parameters. More GPUs. More training data.

Levin is obsessed with Geometry. Topology. Shape. The pattern of connections.

Scale says: "If we make it big enough, intelligence will emerge."

Geometry says: "Intelligence emerges from the right structure, not the right size."

A trillion parameters with no grounding will hallucinate. A billion parameters with proper binding might not.

🪱🗣️🔌🦠🤖📊🏥📐 H → I ⚡

Michael Levin is haunted by dying patients.

In interviews, he's clear: he doesn't care about AI safety philosophy. He cares about speed. Every day that cures take longer, people die.

So here's the reframe:

Grounding isn't about safety. Grounding is about velocity.

A hallucinating AI wastes resources in three devastating ways. First, years of wet-lab trials chasing a statistically plausible but physically impossible compound. Second, millions of dollars on a drug target that was a hallucination. Third, and most important, human lives lost to the delay while teams chase AI-generated dead ends.

You can't build a "cure machine" on top of a "cancerous substrate." If your AI has no gap junctions, it will hallucinate. And hallucinations kill.

🪱🗣️🔌🦠🤖📊🏥📐⚡ I → J 🧪

Here's the cost of cancer in AI:

A pharma company uses an LLM to suggest drug candidates. The LLM, having no ground truth binding, suggests compounds that are statistically plausible but physically impossible.

The wet-lab team spends 18 months synthesizing and testing. The compounds fail. They report back to the AI team: "Your suggestions didn't work."

The AI team fine-tunes on the failure data. The model learns to avoid those specific failures - but it's still hallucinating, just in new directions.

This is the treadmill. Statistical prediction without physical grounding.

Zero-Hop doesn't guess. It retrieves. Semantic neighbors are physical neighbors. The system doesn't predict what compound might work - it finds compounds that already exist in the grounded knowledge space.

🪱🗣️🔌🦠🤖📊🏥📐⚡🧪 J → K 🌅

We're not asking you to slow down. We're offering to cut the dead ends out of your path.

The convergent vocabulary - substrate, binding, geometry, light cone, cancer - isn't a coincidence. It's two fields discovering the same physics.

Biology figured out that intelligence requires gap junctions 600 million years ago when multicellular life emerged. Software is just now catching up.

If you're building AI that matters - AI that has to be right, not just plausible - you need gap junctions.

You need Zero-Hop. You need S=P=H.

You need to cure the cancer before you build the cure machine.

🪱🗣️🔌🦠🤖📊🏥📐⚡🧪🌅 K → L 🎯

The ML engineers reading this are already composing their objection: "Transformers DO bind data. That's what Self-Attention is. Softmax(QK^T) computes relationships between all tokens. You've just described attention mechanisms with fancy biology words."

Let me be precise about why attention is not a gap junction.

Attention is statistical. Gap junctions are physical.

In a transformer, attention weights are computed fresh on every forward pass. Every time you run the model, it recalculates which tokens should attend to which other tokens. The relationship is inferred, not stored. Tomorrow, with different context, those same tokens might attend differently.

In biology, gap junctions are physical portals. Once cells connect, their electrical state merges. The connection persists. It doesn't need to be recalculated. The relationship is instantiated, not simulated.

Attention is "looking at." Gap junctions are "becoming."

When token A attends to token B, token A is observing token B from the outside. It's gathering information. But A and B remain separate entities with separate embeddings.

When cell A gap-junctions with cell B, there is no more "A observing B." Their electrical state becomes shared. They can't tell where one ends and the other begins. They become a single computational unit.

Attention costs compute. Gap junctions are free once built.

This is the Unity Principle in S=P=H. When semantic neighbors are physical neighbors, you don't need to compute their relationship - it's already expressed in the structure. Attention is an expensive simulation of binding. Zero-Hop is actual binding.

Can you approximate gap junctions with attention? Probably. Can you do it without paying compute cost on every forward pass? No. Can you do it without the Unity Principle - without semantic structure matching physical structure? We don't think so. Biology didn't find a way. Maybe there isn't one.

🪱🗣️🔌🦠🤖📊🏥📐⚡🧪🌅🎯 L → M 📚

Some biologists will object: "You're oversimplifying cancer. It's not just about gap junctions. There are genetic factors, epigenetic factors, microenvironment factors..."

They're right. Cancer is extraordinarily complex.

But I'm not citing "Biology." I'm citing Michael Levin's specific research program at Tufts. His lab has demonstrated repeatedly that bioelectric interventions can normalize tumor cells without touching the genome. His definition of cancer as a "shrinking Cognitive Light Cone" comes from experimental results, not philosophical speculation.

At 1:33:25 in the interview, Levin explains the bioelectric basis of cancer. At [00:30], he opens with "The hardware does not define you" - the claim that substrate is interchangeable as long as the pattern of connections is preserved.

The parallel I'm drawing isn't "all of biology says X." It's "Levin's specific research program and our specific software architecture arrived at the same vocabulary and the same structural claims independently." That convergence is the evidence.

🪱🗣️🔌🦠🤖📊🏥📐⚡🧪🌅🎯📚 M → N 🔧

"Great philosophy," says the pragmatist, "but how do I build a Zero-Hop app? Is this just a graph database? Is it denormalized JSON documents?"

No. The key isn't the storage format. The key is addressing.

In the book, we introduce ShortRank - an addressing system where every datum's address encodes its semantic position. Related data lives at adjacent addresses. The address itself is meaningful, not arbitrary.

In normalized databases, the address of a row tells you nothing about its meaning. Row 47 in the Customers table could be semantically adjacent to Row 12,847 in the Orders table, but their addresses give no indication of this. You need a JOIN to discover the relationship at query time.

In Zero-Hop, semantic neighbors have neighboring addresses. The relationship is encoded in the structure. You don't JOIN - you just read the adjacent memory.

The practical implementation involves FIM artifacts - physical data structures where geometry encodes meaning. Think of it like this: if you knew the semantic distance between two concepts, you could calculate the physical distance between their storage locations. That's the Unity Principle. That's S=P=H.

It's not just "denormalize your data." It's "make your data's physical layout match its semantic topology." That's a much stronger constraint. It's also what gap junctions do - they create physical connections that match functional relationships.

🪱🗣️🔌🦠🤖📊🏥📐⚡🧪🌅🎯📚🔧 N → O 🪱

There's one convergence I haven't mentioned yet. It's the strangest one.

In 2000, I had a conversation with philosopher David Chalmers about consciousness. I described my intuition like this:

"Imagine parallel worms eating through problem space. Each worm explores a different path - different hypotheses, different reasoning chains. Most worms hit dead ends. But ONE worm reaches the solution. And it KNOWS it. Not 'probably correct' - it KNOWS with certainty. P=1. That knowing - that instant recognition - is consciousness."

Chalmers called it "a threshold event. Binary recognition."

I wrote this into the book. Chapter 4. Twenty-five years before I ever read Levin.

This is either an extraordinary coincidence, or it's the same physics demanding the same image.

Both models describe parallel search that terminates in binary recognition. Both identify the mechanism as physical binding - gap junctions in biology, co-located firing in consciousness, Zero-Hop in software. Both arrive at the same conclusion: consciousness requires P=1 certainty, and P=1 requires instantaneous binding, and instantaneous binding requires physical proximity.

The vocabulary convergence isn't influence. It's constraint.

When you're describing the same underlying physics - the physics of how information binds across time and space - you end up using the same words. Because those are the words that fit.

---

Related Reading:

Like a Prayer: The Normalization of Culture explores how structural changes become invisible over time. The Rot at the Core of AI Safety examines the knife and the bowl - the trade being offered. We Killed Codd traces the original sin of database design. And Fire Together, Ground Together lays out the full architecture.

Primary Source: Michael Levin Interview (2024)

Key Timestamps from the Interview:

[00:30] "The hardware does not define you" - Levin opens with the substrate-independence claim that underlies the entire parallel with software architecture.

[08:54] The Cognitive Light Cone - Definition of the boundary of what a system cares about. Cancer as shrinking of this boundary.

[20:21] The Self is a Composite - "Selflets" and fluid boundaries. The self is not fixed but a collection of interacting perspectives.

[35:25] Scaling Cognition - How small agents (cells) combine to form large agents (organisms). This is the biological version of the binding problem.

[1:33:25] Bioelectricity and Cancer - The experimental basis for treating cancer as a connection problem rather than a mutation problem.

[1:35:00] Gap Junctions and Memory - The Planaria experiments showing memory stored in bioelectric networks, not neurons.

Additional Sources:

Michael Levin directs the Allen Discovery Center at Tufts University. His lab has published extensively on bioelectric control of morphogenesis, regeneration, and cancer. The Wyss Institute interview 20-ish Questions with Michael Levin reveals his urgency about patient outcomes and his "rescuer" motivation.

The Book: Tesseract Physics - Fire Together, Ground Together lays out the software architecture that independently arrived at the same vocabulary Levin uses for biology.