2026-08-25
Jonas and Kording run connectomics, lesions, tuning, LFP and NMF on all 3,510 MOS 6502 transistors. Plots look like brain data; they never recover adders or the instruction cycle.
A common claim in neuroscience is that the field is mainly data-limited. Get a denser connectome, a wider recording, a bigger simulation, run the usual algorithms, and the computation will show itself. A real brain has no ground truth. Nothing tells you whether a recovered "module", "rhythm", or "causal edge" is the thing the tissue actually computes.
Eric Jonas at UC Berkeley and Konrad Kording at Northwestern turned that into a control experiment. They picked a machine whose function is known at every layer: the MOS 6502 (and the nearly identical 6507), the chip inside the Apple I, the Commodore 64, and the Atari VCS. The Visual6502 project had already stripped the epoxy, imaged the die, and produced a transistor-accurate netlist of 3,510 enhancement-mode transistors. A cycle-accurate simulator then exposes every wire voltage and every transistor state. The point is not that a chip is a brain. The question is whether off-the-shelf neuroscience tools, applied naively, recover the kind of understanding an electrical-engineering undergraduate would accept: registers, an ALU, instruction decode, adders.
"Behavior" is the boot sequence of three games: Donkey Kong, Space Invaders, and Pitfall. Each run lasts about 10 seconds and more than 250 frames. The simulator dumps roughly 1.5 GB of state per second and advances about 1,000 clock cycles per wall-clock second. Off-to-on transistor transitions are treated as spikes.
The analyses climb resolution:
Each step uses tools a 2017 neuroinformatics course would teach. Electrical-engineering priors are deliberately withheld.
Clustering splits out clocked stateful transistors, grounded inverters, and a group that gates the X/Y/S registers onto a special bus. That looks like cell typing. Physically there is one transistor type. The clusters mix layout with local circuitry and never surface "this is an adder".
Lesions are more misleading. Some transistors take down only Donkey Kong, which in neuroscience language sounds like a Donkey Kong neuron. They implement generic logic such as full adders. Necessity for one game means that boot code happened to touch them.
Tuning curves are pretty and wrong. A few transistors show simple or complex luminance curves while computing nothing about brightness; the correlation likely tracks game-stage changes. Weak pairwise and strong global correlations also mimic neural population data. LFPs yield roughly power-law spectra and region-specific "rhythms". The chip has one element type. Those frequencies are byproducts of computation, not evidence of communication channels, and they are a poor reason to model the processor as coupled oscillators.
Granger causality gets some arrows right: registers really influence the accumulator, the decoder really influences status bits. It also invents decoder-to-decoder talk in Donkey Kong and drops accumulator-to-register influence in Pitfall, which the architecture requires. NMF's six components line up with the two-phase clock and the read/write line RW. Finding the clock is not understanding the instruction cycle.
Unlimited data plus standard analyses can manufacture figures that could pass as neuroscience. They do not yield a hierarchical account of computation.
For people who work on interpretability, connectomics, or the story that more recording will be enough, this is a cheap sanity check. Brains are harder than a 6502: neurons are stochastic, diverse, plastic, and their input-output functions are not reconstructed. The chip is smaller, deterministic, single-typed, and arbitrarily lesionable. Tools that fail on the easy object have no default claim on the hard one.
The discussion already flags deep-network visualization. Even human-trained classifiers remain partly opaque. If cortex does anything like gradient descent on a cost, watching and factorizing activity will be even less sufficient. The constructive suggestion is narrow: new methods should first recover known artifacts. If they cannot find an adder, they are not ready to claim they found grid cells.
Connectomics is not wasted here. Without the netlist the simulation does not exist. What is missing is the jump from anatomy to computational hierarchy.
The authors list the obvious mismatches. A brain is analog, massively parallel, and evolved; a 6502 was designed by a small team in a few years. Behavior is a passive boot, not play, so the input space is a handful of switches, close to the reaching tasks that dominate motor neuroscience. Binary spikes discard analog voltages on wires. The NMF rank of 6 was chosen by hand for interpretability, not by cross-validation. Ten seconds of three games leave most instruction paths unexcited.
A sharper caveat sits in the success criterion. "Understanding" is defined as the textbook module diagram. If someone thinks neuroscience should not chase replaceable modules, the failure verdict does not apply. The paper also does not benchmark methods that arrived after 2017: nonlinear state-space models, differentiable circuit search, mechanistic interpretability tooling. It shows that the then-standard toolbox is not enough. It does not show that no data-driven method can work.