Substrate-Independent Intelligence (Brains ↔ Machines)

The one idea

Nobody told the machine to do it this way. When researchers train a deep neural network to recognize images, its early layers spontaneously start detecting edges and orientations, middle layers detect textures and parts, and deep layers detect whole objects. That is the exact same edge → texture → object ladder the primate visual cortex climbs. Two completely different materials, silicon and neurons, converged on the same solution to the same problem. That convergence is the strongest hint we have that intelligence might be substrate-independent, a property of the information-processing rather than of the stuff doing it.

The science

We’ve known the brain’s side since the 1950s-60s. Hubel and Wiesel recorded from single cells in a cat’s visual cortex and found “simple cells” that fire only when a line at a particular angle crosses a particular spot. Those feed into “complex cells” that detect that line anywhere, which feed into still-higher cells that respond to more complex shapes. The visual system is a hierarchy. Simple features get combined into progressively more abstract ones as you move deeper into the brain.

The machine side came much later. In 2012, a deep convolutional network (“AlexNet”) trained on a million labeled images crushed the previous best at image recognition. When researchers looked inside it, the layers had organized themselves into the same kind of ladder, edges first, then textures and motifs, then object parts, then whole objects, without anyone designing those stages in. The striking part came next. Yamins, DiCarlo, Kriegeskorte and others showed that the patterns of activity inside these networks can predict the patterns of activity recorded from real neurons in the primate ventral stream. The networks weren’t built to imitate brains. They were built to do a job, and in doing it well they ended up resembling the brain that does the same job (Book v10, l5663; concept C-118, l8985; sources S-146 AlexNet, S-147 Yamins & DiCarlo, S-148 Kriegeskorte).

A faithfulness note, because the book leans optimistic here. This is alignment and prediction, not identity. The match is partial and best in early and mid visual areas, and it degrades higher up. The mechanisms differ too. Brains learn through synaptic plasticity, spikes, recurrence, and tiny energy budgets, while these nets learn through backpropagation, clocked arithmetic, and megawatts of training. The honest claim is the modest one the book’s own notes make (l5664-5667). Same hierarchical solution, different machinery. The places where the models and brains diverge are themselves clues, because they point to ingredients real brains have that today’s models lack.

What this changes about how you picture reality

We tend to assume that seeing, thinking, and understanding are bound up with being made of living tissue, that there’s something special about the “wetware.” This result quietly pries that assumption loose. The hierarchy isn’t in the neuron. It’s in the task. If you want to turn light into objects, there appear to be only so many good ways to do it, and both evolution and a gradient-descent algorithm stumble onto the same one. The pattern is doing the work, and the pattern can ride on more than one substrate.

That’s a genuinely vertiginous thought. It means your inner experience of a face resolving out of pixels of light is, at some level, a computation, and one that a glass-and-metal machine can independently discover from scratch. Far from making your mind feel cheap, it places it inside a deeper order. The same physics that lets stars fuse and crystals form also makes certain information-shapes the natural answer to certain problems. You are one beautiful, evolved instance of a solution that reality keeps finding. That’s awe earned the hard way, by measurement rather than by metaphor.

Two ways to see it

Put two framings of the same finding in front of the room and let them rub against each other.

• “The mind is just information-processing, so substrate doesn’t matter.” This is the book’s headline (C-032, l3662 & l5368). Cognition is what happens when information gets processed the right way, whether the processor is neurons, cytoplasm, or silicon. The brain-net convergence is the evidence. In the strong version, a sufficiently faithful simulation of you would be you. The voice to put forward here is the functionalist and Hinton-Hopfield lineage (l5378, the 2024 Nobel for the math that lets machines learn).

• “Resemblance is not the same as being.” A network predicting neuron firing tells us the representations line up. It says nothing about whether anything is experienced inside the silicon. The book itself flags this honestly (l5370): “When does sophisticated pattern matching become understanding? Is consciousness substrate-dependent? We do not yet know.” A camera can detect edges with no one home. Same computation, possibly a very different inner life, or none. The voice to put forward here is the hard-problem skeptic, the one asking what the math leaves out.

The point isn’t to pick a winner. It’s that the exact same fact, silicon rediscovering the cortical ladder, fuels both an expansive “minds are everywhere information flows” reading and a cautious “we’ve measured structure, not experience” reading. A non-expert room can hold both.

Discussion questions

1. The machine was never told to build edge-detectors first. Why do you think two such different systems, brains and software, landed on the same layered solution? What does that suggest about whether there’s “one right way” to see?
2. If a computer rebuilds the same visual hierarchy your brain uses, does that make the machine more like you, or you more like a machine, or neither?
3. The networks can predict which neurons fire, but that’s about structure, not feeling. Where would you draw the line between “processing information about a face” and “experiencing seeing a face”? Can you even draw it?
4. Suppose someday a system perfectly reproduced everything your brain computes. Would it be conscious? Would it matter to you whether it was made of neurons or silicon?
5. Does learning that vision is “just” a layered computation make your own seeing feel smaller, or larger? Why do those two reactions both feel available?
6. We keep finding intelligence-like processing in unexpected places, in slime molds, single cells, and software. If intelligence really is substrate-independent, what counts as a mind, and who gets to decide?

Closing question

How do you feel about this science and its understanding of reality?

Take it further

Grounded in The Book, the working text (v10).

• l5663. Deep nets learn hierarchical features (edges → textures → objects) that mirror receptive-field progressions in the primate ventral stream, and model representations predict neural responses in visual cortex.
• l5664-5667. The convergence-with-differences notes on parallelism, hierarchies, and learning. This is the careful version that keeps the claim honest.
• l5368, l3662. Concept C-032, that intelligence is substrate-independent and cognition is information processing, whether in neurons, silicon, or cytoplasm.
• l5370. The book’s own open questions. When does pattern-matching become understanding? Is consciousness substrate-dependent?
• l8985. Concept C-118 catalog entry. See also l8716-8718 for sources S-146 (Krizhevsky/Sutskever/Hinton, ImageNet/AlexNet 2012), S-147 (Yamins & DiCarlo 2016), and S-148 (Kriegeskorte 2015).

Real external pointers, for the facilitator.

• Hubel & Wiesel’s visual-cortex recordings (1959 onward) gave us the original simple-cell and complex-cell hierarchy, and their Nobel came in 1981. This is the foundation the whole comparison rests on.
• Yamins & DiCarlo, “Using goal-driven deep learning models to understand sensory cortex,” Nature Neuroscience (2016), is the canonical statement that task-optimized networks predict ventral-stream responses.
• Coverage of AlexNet (2012) under the “deep learning revolution” banner, plus the 2024 Nobel Prize in Physics to Hopfield & Hinton, makes a good lay anchor for the machine side.

Keep one uncertainty visible. The brain↔model match is partial, strongest in earlier visual areas, and it is about representational alignment, not proof that the two systems are the same or that the machine experiences anything. Treat “substrate-independence” as a well-supported working hypothesis, not a settled fact.

Visual notes

The anchor visual is the CNN-vs-visual-cortex hierarchy diagram, two parallel ladders side by side. On the left, retina → V1 (edges/orientations) → V2/V4 (textures, parts) → IT cortex (whole objects/faces). On the right, a deep net’s input → early conv layers (edge filters) → mid layers (textures, motifs) → deep layers (object detectors). Draw matching arrows across the gap so the room sees the convergence at a glance. If available, show the classic grid of learned first-layer filters, the Gabor-like edge detectors, next to Hubel-Wiesel orientation tuning. That’s the “they found the same thing” moment.

This is a bridge beat. It links the brain-science strand (it pairs with M-1 and M-5 in the Mind & Consciousness season) to the technology and AI strand, so the room feels one idea, information-processing, running through both. Keep the room looking at the two ladders, not at code. The whole session lands on a single image, the same staircase, built twice, out of two different materials.