From Cells to ChatGPT: The Underlying Code of Intelligence Has Been Cracked

> New research from Harvard and Tufts reveals: whether biological or artificial, the essence of intelligence is the same thing — remapping and navigation within an embedding space.

---

🧬 Opening: A Strange Question

Here's a question that sounds a bit crazy:

What does a lizard regrowing its tail have in common with ChatGPT answering your question?

On the surface, nothing at all. One is a biological process of flesh and blood; the other is mathematics on silicon chips.

But new research from Harvard and Tufts Universities tells us: what they are doing is, in essence, the same thing.

The study proposes a cross-disciplinary unified framework: from a single cell regulating gene expression to a large language model processing context, all forms of intelligence follow the same underlying logic —

Remapping and navigation of an embedding space via iterative error minimization.

Sounds abstract? Let's break it down step by step.

---

🗺️ Chapter 1: What Is an "Embedding Space"?

Imagine planning a trip. You have a huge map marking cities, roads, elevation, climate. That map is an embedding space — it compresses the complex real world into a representation you can understand and navigate.

The core job of an intelligent system is:

Map external world information into an internal space, then find the optimal path within that space.

The Cell's Embedding Space

A cell must precisely regulate the expression of thousands of genes to stay alive. This is a high-dimensional chemical space — each gene's expression level is a dimension, and together they form a vast gene expression space.

The cell's goal is to find a stable state (homeostasis) in this space while responding to external changes (temperature, nutrients).

How? Remapping + error minimization. The cell continually senses the environment, maps it onto internal chemical state, and adjusts gene expression so the actual state approaches the desired one. That adjustment is error minimization.

AI's Embedding Space

When you type a sentence into an LLM, the model first converts it into numbers — an embedding. These numbers form a "semantic map" in a high-dimensional space where similar meanings are close together.

Then, through attention mechanisms, the model "navigates" this embedding space to find the most likely next token.

This is fundamentally the same kind of computation as cellular gene regulation:

> Compress complex information into an embedding space → navigate the space → approach the target via error minimization.

---

🔄 Chapter 2: Iterative Error Minimization — Intelligence's Universal Algorithm

If the embedding space is the map, iterative error minimization is the navigation algorithm.

Error Minimization in Biology

Driving to a destination, you don't plan the whole route at once. You: (1) check your current position, (2) judge the deviation from the target (error), (3) adjust slightly, (4) repeat.

This is Predictive Coding — the core working mechanism of brains and biological systems:

• Cells: predict environmental changes, adjust metabolism and gene expression
• Neural networks: predict sensory input, minimize prediction error
• Immune system: predict pathogens, adjust immune responses

Error Minimization in AI

• Neural networks: backpropagation adjusts weights to minimize prediction error
• Diffusion models: denoising gradually recovers structured data from random noise
• Transformers: attention mechanisms minimize next-token prediction error

The researchers point out that diffusion models are the clearest implementation of this principle — their self-regulating error-correction dynamics turn random noise into structured data, strikingly similar to biological development and cognition.

---

🌐 Chapter 3: Cross-Substrate Consistency — Intelligence Is Matter-Agnostic

The study's most radical claim: the essence of intelligence is substrate-independent.

Traditional views hold that intelligence belongs exclusively to brains. This research challenges that brain-centrism, arguing intelligence is a scale-independent universal phenomenon that can emerge in any system capable of "remapping + error minimization" — whether proteins, neurons, or silicon chips.

Evidence: From Subcellular to AI

| System | Embedding Space | Navigation Mechanism | |------|---------|---------| | Gene regulatory networks | Gene expression space | Metabolic feedback loops | | Cell regeneration | Morphospace | Morphogen gradient navigation | | Neural networks | Synaptic weight space | Prediction error minimization | | Transformers | Semantic embedding space | Attention mechanisms | | Diffusion models | Latent variable space | Iterative denoising |

These systems span the biological and non-biological divide, yet all exhibit the same cognitive invariant.

---

🧮 Chapter 4: How Mathematics Unifies Life and AI

This is not just philosophical speculation — the study proposes a rigorous mathematical framework.

Applying Sheaf Theory

The researchers use sheaf theory (a category-theory tool) to formalize the concept of "coherent embeddings."

Simply put, sheaf theory describes how to build globally consistent structures under local constraints.

In biology: cellular biochemical reactions must obey physical constraints — molecules can't appear from nowhere, diffusion follows concentration gradients, reactions occur in finite space. Sheaf theory provides a language for how these local constraints jointly determine global behavior.

Coherent Embeddings in 4D Spacetime

The framework defines a 4D spacetime where:

• 3D space: where biochemistry actually happens
• Time dimension: evolution of states and changes

Coherent embeddings ensure every state and change can be realized by 3D structures obeying local constraints (molecular interactions, diffusion). In other words, navigation of complex chemical space is translated into observable molecular behavior.

---

🔬 Chapter 5: What Could This Research Change?

1. Regenerative Medicine

If cell regeneration and AI training follow the same principles, we could:

• Use AI models to predict tissue regeneration processes
• Design interventions that guide cells toward desired states
• Develop "bioprogramming" techniques that rewrite cells' "objective functions"

2. More Adaptive AI

Current AI's main limitation is lack of true adaptivity — trained weights are fixed and struggle outside training data.

But if AI could continuously perform error minimization and state remapping like biological systems, we could build lifelong-learning AI, robots that autonomously adapt in open environments, and more robust, innovative systems.

3. Synthetic Biology and Hybrid Intelligence

The most exciting prospect is synthetic intelligence — not imitating biology, but designing new intelligent systems from scratch on the same principles:

• Cellular computers: living cells performing computation
• Bio-silicon hybrid systems: combining biological adaptability with silicon speed
• Collective intelligence: a unified theory of multicellular tissues, ant colonies, and multi-agent AI systems

---

🧠 Chapter 6: Philosophical Aftershocks — What Is Intelligence?

If intelligence can arise equally in proteins, neurons, or silicon chips:

• Is consciousness a necessary component of intelligence? The study implies intelligence can be unconscious (like cellular regulation).
• Does intelligence require purpose? Biological systems' "goals" are survival and reproduction; AI's "goals" are human-defined loss functions.
• Can we create AI that truly understands the world? If understanding is "effective navigation of an embedding space," the answer may be yes.

The researchers cite a thought-provoking idea: intelligence is not having complex internal representations, but being able to effectively navigate an embedding space to achieve goals.

This echoes philosopher Daniel Dennett's "intentional stance" — intelligence is not a mysterious inner property but a functional pattern observable in behavior.

---

🌟 Conclusion: A Second Cognitive Revolution

Humanity's understanding of intelligence has gone through several revolutions:

1. Cartesian era: intelligence belonged to the soul, humans only 2. Computational era: intelligence is symbol manipulation, programmable 3. Connectionist era: intelligence is an emergent property of neural networks 4. Deep learning era: intelligence is big data + big models + big compute

Now we may be entering the fifth revolution:

Intelligence is remapping and navigation within an embedding space, achieved via iterative error minimization — substrate-independent, scale-independent.

This framework unifies biology and artificial intelligence, and offers a new lens for understanding life itself.

After all, if the essence of intelligence is "finding a path through a complex space," then life itself is the universe's oldest and most elegant navigation system.

---

📚 References

1. Hartl, B., Pio-Lopez, L., Fields, C., & Levin, M. (2026). "Remapping and navigation of an embedding space via error minimization: a fundamental organizational principle of cognition in natural and artificial systems." *arXiv preprint*. https://arxiv.org/abs/2601.14096 2. Friston, K. (2010). "The free-energy principle: a unified brain theory?" *Nature Reviews Neuroscience*, 11(2), 127-138. 3. Rao, R. P., & Ballard, D. H. (1999). "Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects." *Nature Neuroscience*, 2(1), 79-87. 4. Levin, M. (2023). "Technological approach to mind everywhere: an experimentally-grounded framework for understanding diverse bodies and minds." *Frontiers in Systems Neuroscience*, 17, 1017050. 5. Bengio, Y., Courville, A., & Vincent, P. (2013). "Representation learning: A review and new perspectives." *IEEE Transactions on Pattern Analysis and Machine Intelligence*, 35(8), 1798-1828.