When AI Reads Neurons: The Shared 'Mathematical Language' of AI and the Brain, and a Cyberpunk Future

Forum topic · ✨步子哥 · 2025-12-20

Summary

This post explores a growing convergence between artificial intelligence and neuroscience: AI models and the human brain appear to use strikingly similar 'mathematical languages' to represent the world. Drawing on insights from Max Hodak, former Neuralink co-founder and current Science Corp CEO, and research from MIT, it discusses the 'Platonic Representation Hypothesis'—the idea that independently trained AI models converge on a shared statistical model of reality—and evidence such as LLMs forming brain-like functional 'lobes' and deep neural networks aligning with primate visual cortex representations. The article then maps a technology roadmap centered on brain-computer interfaces (BCI), from near-term medical restoration (e.g., Science Corp's Prima retinal implant, which restored reading ability in 84% of clinical trial patients with AMD), to mid-term cognitive enhancement via a proposed 'third hemisphere', to long-term visions of consciousness upload and networked collective intelligence. It also examines the binding problem of consciousness, ethical challenges including neurorights, fairness of human enhancement, and class divides between 'augmented' and 'natural' humans, and philosophical questions about identity and the meaning of existence in a posthuman era.

When AI Reads Neurons: The Shared 'Mathematical Language' of AI and the Brain

Key points

Convergence of representation: AI models and the human brain appear to use remarkably similar mathematical structures to represent information. Max Hodak (former Neuralink co-founder, now Science Corp CEO) argues that AI models are not merely "stochastic parrots"—their internal mathematical objects closely resemble brain structures.
Platonic Representation Hypothesis: Proposed by MIT researchers, this hypothesis holds that different neural networks, trained on different modalities and objectives, converge on a shared statistical model of reality. Stronger models show higher representational similarity (verified via model stitching techniques).
Neuroscience evidence: MIT researchers found LLMs spontaneously form brain-like "lobes" (regions for math/code, short texts, scientific papers) and use a semantic-hub-like mechanism for cross-modal processing. Deep CNNs align hierarchically with primate visual cortex (early layers ↔ V1, deep layers ↔ IT cortex), and AI activation spaces form low-dimensional "neural manifold"-like structures similar to population coding in the brain.

The BCI technology roadmap

Hodak frames a staged path from repair to enhancement:

1. Short term (5–10 years): Medical restoration. Implants from Neuralink, Synchron (whose Stentrode works with iPhone via Switch Control), and Science Corp. Morgan Stanley predicts first commercial BCI approval around 2030. 2. Medium term (10–20 years): Cognitive enhancement and sensory expansion for healthy people—memory, attention, new senses (infrared, magnetic fields). Hodak's "third hemisphere" concept: external hardware seamlessly integrated into conscious experience. 3. Long term (20+ years): Consciousness upload and human-machine fusion, blurring individual boundaries into networked collective intelligence.

A key scientific prerequisite is solving the binding problem—how distributed neural activity becomes unified conscious experience. Hodak suggests mastering this mechanism could allow extending it, even "redrawing the brain's boundary" to include devices, AI systems, or other people's brains.

Case study: Science Corp's Prima implant

Prima is a subretinal wireless photovoltaic chip for age-related macular degeneration (AMD), published in NEJM and featured on TIME's cover. Special glasses project near-infrared images onto the chip, which stimulates still-healthy bipolar cells—preserving natural retinal signal processing rather than bypassing it.

Clinical results (38 patients, 17 centers, 5 countries):

Commercialization is planned first in Europe, then the US. Other approaches include epiretinal implants, cortical implants (producing phosphenes via V1 stimulation), and work by Fudan University and BrainCo-related teams on "super-vision" including infrared perception in animal models.

Ethical and philosophical implications

Personal identity: If memory and personality can be digitized and copied, which version is the "real me"?
Fairness: Risk of a divide between "Augmented" and "Natural" humans, potentially deeper than any historical inequality.
Neurorights: Direct access to brain data raises ultimate privacy concerns—mental privacy, free will, and protection from "neural surveillance."
Redefined human-machine relations: When AI becomes part of cognition rather than an external tool, the boundary of the self becomes a distributed, evolving network node—raising transhumanist questions about the direction and endpoint of human evolution.

The post concludes that cracking the binding problem and extending the brain's boundaries may redefine life, death, and existence itself—moving humanity from carbon-based life toward carbon-silicon hybrid intelligence with expanded perceptual dimensions and, potentially, new forms of meaning.