Key Takeaways

  • Max Hodak observed that the internal mathematical geometry inside large artificial intelligence models directly mirrors neural recordings from animal brains.
  • This alignment supports the Platonic representation hypothesis: neural networks and biological brains converge on the same data manifold of physical reality.
  • Neuroscientists are moving to labs like OpenAI and Anthropic because dissecting software weights in code is faster than probing physical brain tissue.
  • Brain-computer interface teams at Science Corporation map biological neural signals onto AI latent spaces to decode sensory inputs for devices like the PRIMA retinal implant.

Convergence on the Physical Manifold

Max Hodak noticed something unexpected when inspecting large artificial intelligence models. The internal mathematical structures did not look like arbitrary computer code. They matched the firing patterns recorded from living animal brains.

“When you look inside these big AI models, the mathematical objects that you see look a lot like the things that you see in neuroscience,” Hodak says. “We know that that is true because we can get alignments between animal brain neural recordings and AI model internal representations.”

This observation anchors the Platonic representation hypothesis. The theory argues that any capable learning system, whether made of wet biological neurons or silicon transistors, discovers the same geometric structures to represent the physical world. A single data manifold governs physics, light, and 3D space. When you train a model on enough data, it reconstructs that exact manifold.

As Hodak describes it: “There's some fact about the universe where these things as they're learning are grabbing onto some true underlying data manifold. I mean it feels like a lot of physics. Like if you apply enough compute to matter, you get this thing that looks like intelligence.”

This alignment convinced Hodak that modern AI had moved past statistical parlor tricks. Because internal representations match biology, neural networks do far more than memorize text strings: they construct an accurate map of reality itself.

Why Brain Researchers Move to AI Labs

For decades, neuroscientists hit physical barriers trying to measure living brains. Reading a thousand neurons at once required complex physical probes, delicate animal surgeries, and endless biological noise.

Artificial neural networks remove that barrier completely. You can inspect every weight, run millions of forward passes, and measure exact activations across layers without touching biological tissue.

“I have a couple of neuroscience friends at OpenAI and Anthropic who it's like we would joke like, 'Oh, you left neuroscience?'” Hodak recalls. “Like, 'No, no, no. It is just way easier to do neuroscience on the models.'”

This shift explains why brain-computer interface companies are drawing sudden interest from machine learning engineers. If brains and neural networks share representational spaces, translating between them becomes a standard math problem. Science Corporation applies this concept directly: instead of inventing custom decoding logic for every biological nerve, engineers can translate biological signals directly into pre-trained AI representations.

What to Do With This

Stop treating biological data processing and machine learning as separate disciplines. If your product translates sensor signals, audio, or imagery into software actions, stop writing hand-crafted heuristic decoders. Train a standard foundation model on rich physical-world data, extract its latent space representations, and map your raw sensor inputs directly against those vectors.