Biology and AGI: What Was Really Copied From the Brain, and What Is Just Metaphor

The phrase is everywhere: «neural networks, inspired by the workings of the human brain». It shows up in product launches, press releases and headlines. And it is not false. The problem is that it papers over the gap between two very different claims: «this was inspired by that» and «this works like that».

The first is verifiable history, with documents and dates. The second is, almost always, decoration. Learning to separate them changes how you read any announcement in this industry, today and five years from now.

The real borrowing, and its three dates

It begins in 1943, when Warren McCulloch and Walter Pitts publish «A logical calculus of the ideas immanent in nervous activity» in the Bulletin of Mathematical Biophysics. Their idea: a neuron can be modelled as a unit that sums its inputs and fires once a threshold is crossed. From there, in a straight line, comes the artificial unit repeated millions of times inside any model today.

It continues in 1959 and 1962, with David Hubel and Torsten Wiesel's work on the cat visual cortex, published in The Journal of Physiology and now freely available in PubMed Central. They described neurons that respond to edges at a specific orientation — they called them simple cells — and others that hold that response even when the edge moves: complex cells. Local detection first, tolerance to position second.

And it culminates in 1980, with the link that turns all of this into engineering. Kunihiko Fukushima publishes the Neocognitron in Biological Cybernetics, and his own abstract states that the network, once self-organized, «has a structure similar to the hierarchy model of the visual nervous system proposed by Hubel and Wiesel». This is not a later reconstruction by historians: the author says so, in writing, in the document. His layers are called S-cells and C-cells, lifted straight from simple and complex. That scheme — detect, then tolerate displacement — is exactly the convolution-and-pooling alternation of modern convolutional networks.

That is the borrowing, and it is solid. Now the other half.

Where the analogy breaks: the watt

The human brain runs on about 20 watts. The figure has an open, citable source: physicist Vijay Balasubramanian gives it in PNAS (2021), noting that the brain «manages to produce poetry, design spacecraft, and create art on an energy budget of ∼20 W». In the same piece he sets that against the 80 watts of the computer he was writing on.

There is a physiological detail that rarely gets told and that dismantles the processor intuition. Marcus Raichle and Debra Gusnard documented in PNAS (2002) that the brain is around 2% of body weight and consumes «about 20% of the oxygen and, hence, calories consumed by the body» — and they stress that this metabolic rate is remarkably constant even though mental activity varies enormously. Thinking harder does not spike consumption the way load does on a chip. If the analogy were literal, that should not happen.

Where it breaks again: the gradient

The second crack is more technical and more revealing, because it explains why biological fidelity is not free.

There is a family of networks that does imitate the real mechanism of the neuron: spiking neural networks, which communicate with discrete pulses in time rather than continuous numbers. They are more biologically plausible. And they have spent decades failing to displace conventional deep learning.

The reason is in an open review published in Frontiers in Neuroscience (2023): «the intricately discontinuous spike mechanism brings difficulty to the optimization» of these networks. All of modern deep learning rests on gradient descent, which needs differentiable functions in order to know which way to nudge each parameter. A pulse that either happens or does not is not differentiable. Without a gradient, the entire training machinery stops applying as-is.

The conclusion is worth reading twice: these networks lose not for lack of biological plausibility, but precisely because of it. Copying the brain more faithfully moves them away from the mathematical tool that makes the competition work.

What is genuinely neuromorphic, and with which numbers

Hardware does pursue this path seriously, and it is worth looking at it through its makers' own figures.

Intel introduced Hala Point on 17 April 2024: 1,152 Loihi 2 processors, 1.15 billion neurons and 128 billion synapses, 140,544 neuromorphic cores, deployed at Sandia National Laboratories, drawing a maximum of 2,600 watts.

One note of honesty, because the temptation is obvious: setting those 2,600 watts beside the brain's 20 is our comparison, not Intel's. Intel compares its system with conventional GPUs and CPUs, not with a brain, and it is right to: Hala Point sits orders of magnitude below the human brain in neuron count. Both figures are true; putting them side by side without saying this would be a trick.

IBM took another route with NorthPole, published in Science in October 2023. The paper is paywalled, but IBM Research's official blog gives the numbers in the open: 256 cores, 22 billion transistors, and roughly 25 times better efficiency in frames per joule than common 12-nanometre GPUs on the ResNet-50 benchmark. The key is not imitating the neuron: it is removing off-chip memory and interleaving compute and memory, which is what nervous tissue does.

A real name, and why the name matters

Much of this hardware grew out of a public programme with an exact name: Systems of Neuromorphic Adaptive Plastic Scalable Electronics, shortened to SyNAPSE, run by DARPA. Its official page states the goal — low-power neuromorphic computers that scale to biological levels — and marks it as completed. It started in 2008, and its most visible result, IBM's TrueNorth chip, was announced by DARPA on 7 August 2014: more than 5 billion transistors, one million electronic neurons, over 250 million synapses, and power draw under 100 milliwatts.

The acronym is a pun on the word synapse. And here is a trap that circulates: invented expansions of real acronyms, built backwards so they sound technical, turn up regularly in secondary write-ups. An acronym that sounds like a laboratory is not a source. It takes thirty seconds to check whether an official programme page with that exact name exists — and when it does not, the answer is not to fix the name: it is to delete it.

The capability: three questions for «brain-inspired»

1. What exactly was copied, and who says so? The gold standard is Fukushima's: the author himself states in his own document which work he took the structure from. If the inspiration appears only in marketing material and not in the technical documentation, it is ornament.

2. Does biological fidelity buy anything measurable? Energy per inference, latency, accuracy. If imitating the neuron more closely improves no number, the resemblance is aesthetic.

3. Does the name point to a real document? A programme, a paper, an official registry. If the institution's page does not come up, the term does not get used.

The deep end, and an anomaly worth knowing

There is something telling in this piece's source list. Hubel and Wiesel's two neuroscience papers, from 1959 and 1962, are freely accessible in PubMed Central. The founding paper of neural networks — McCulloch and Pitts, 1943 — is behind a paywall at Springer: the PDF costs 39.95 euros, and I found no authorized institutional copy in the open.

In other words: the cornerstone of the field everyone is talking about is out of reach for the average reader, while the physiology that inspired it is not. Worth keeping in mind when someone quotes it from memory in a slide deck.

A practical note on the two Hubel and Wiesel papers: they are free but digitized as scanned images, with no indexed abstract. You can read them, not search them. Anyone wanting to quote them verbatim will have to open the PDF and find the page by hand — and that, too, is part of working with primary sources.

The capability you leave with: faced with any «brain-inspired» claim, ask what was copied, who asserts it in a document of their own, and what that resemblance buys in measurable numbers. With those three questions, the difference between a scientific borrowing and a sales metaphor stops being a matter of opinion.