← Back to Home

Bitflips in the Brain

April 2026

Cosmic ray particles flip bits in computer memory at measurable rates. Neurons are much noisier than that, and the argument here is that the brain evolved under the same kind of unshieldable noise environment, where deterministic precision was never an option. Once noise tolerance was forced on the substrate, evolution found uses for the noise itself.

Apollo astronauts saw flashes of light with their eyes closed. About once every 2.9 minutes on average, in the dark, with no light source involved 1. The mechanism turned out to be cosmic ray particles ionizing molecules in the retina, or producing Cherenkov radiation in the vitreous humor, on their way through the head. The brain treated the resulting signal as light because at the receptor level it had no way to tell it wasn't.

Astronaut Ron Evans wearing the ALFMED light flash detector during Apollo 17.
Astronaut Ron Evans wearing the ALFMED light flash detector during Apollo 17.

The same kind of ionization happens at sea level, just much less. About one muon per square centimeter per minute reaches the ground, which is roughly ten muons per second through a head-sized volume 2. That flux is going through everyone, continuously, and a small subset of those particles deposit enough energy somewhere sensitive to flip a bit of state.

In computers this is well documented. Ziegler and Lanford described the mechanism for cosmic-ray-induced bit flips in DRAM in 1979 3. Modern estimates put the rate at roughly one flip per 256 MB per month at sea level, scaling with altitude (about an order of magnitude higher on a mountaintop, about 300x on a plane) 4.

Simulated cosmic ray air shower from a 1 TeV proton hitting the atmosphere 20 km above the Earth.
Simulated cosmic ray air shower from a 1 TeV proton hitting the atmosphere 20 km above the Earth.

There is a well-documented case study of one of these in the wild, from a 2013 Super Mario 64 speedrun race. The runner DOTA_Teabag was running the 70-star category live on stream against another player, MidBoss, when his Mario teleported straight upward through a ceiling and ended up on the floor above 5. The warp saved time. More importantly, nothing in the game's state at that moment could have produced it. Pannenkoek2012, a SM64 analyst who has spent years documenting the engine's quirks, posted a $1,000 bounty for anyone who could reproduce the glitch in normal play. It went unclaimed for years.

The leading explanation is a cosmic ray bit flip in the N64's RAM. A single bit in the high byte of Mario's vertical position flipped from 1 to 0, changing the byte from 0xC5 to 0xC4 and his height in memory from 0xC5837800 to 0xC4837800, which by coincidence corresponded almost exactly to the height of the floor above him. Pannenkoek confirmed this was plausible by manually flipping that bit at the same moment in the run and reproducing the warp.

Neurons are very noisy components

If you record from a single cortical neuron and present the same stimulus 100 times, the responses are not identical. Spike timings vary by milliseconds, spike counts by tens of percent, membrane voltages fluctuate continuously at rest, and synaptic vesicle release is intrinsically stochastic. The standard reference is Faisal, Selen and Wolpert's 2008 review Noise in the nervous system 6, which argues that this kind of variability is unavoidable in any system built from molecular components at the nanometer scale. Their list of noise sources is almost entirely intracellular: thermal motion, ion channels, synaptic vesicles, biochemistry.

Evolution

A brain in a noiseless universe could afford to be deterministic, because nothing would ever flip its state from outside. A brain in our universe cannot, because something is always flipping its state from outside. Whether the dominant flips are cosmic, thermal, or biochemical is a quantitative question. The qualitative answer is the same: deterministic precision was never an option, and a brain architected for digital reliability would have been outcompeted long before it could do anything interesting.

So selection favored substrates that work despite ongoing corruption: distributed representations, population codes, soft thresholds, anything that prevents a single flipped bit from cascading into a wrong decision. The architecture is built to assume corruption is a constant of the environment, because in this environment it is.

Once noise tolerance is in place, evolution can repurpose the noise itself. The stochastic-dynamics literature in cognition makes this case directly 7: trial-to-trial variability prevents attractor networks from deadlocking, supports probabilistic inference, and enables stochastic resonance, where adding noise to a sub-threshold signal makes it detectable.

There is also some genuine error correction in the brain, just not at the level of individual cells. Sreenivasan and Fiete (2011) showed that grid cells in entorhinal cortex implement an analog error-correcting code for spatial position 8. The hexagonal firing pattern that initially looked like a computational quirk is a redundant population code with strong noise resistance. There is probably more of this hiding in cortex that hasn't been characterized yet.


Announcing my new company: Blips (Brain Flips). We launch satellites that beam collimated muons through the meeting room of your choice, flipping one or more bits somewhere in your counterparty's brain at the moment of decision. We are working on the targeting. We cannot tell you which bits. We cannot tell you which way they flip. Outcomes are stochastic in the technical sense. Accepting angel checks.

Footnotes

  1. Cosmic ray visual phenomena - Wikipedia. The Apollo Light Flash Moving Emulsion Detector experiment on Apollo 16 and 17 confirmed the link by matching cosmic ray tracks to reported flashes.

  2. Muons - Lawrence Berkeley Laboratory. ~1 muon/cm²/min at sea level; about ten per second through a volume the size of a human head.

  3. Single-event upset - Wikipedia.

  4. Soft error - Wikipedia. Flux roughly doubles every 1,000 m of altitude.

  5. Cosmic Ray Flips Bit, Assists Mario 64 Speedrunner, Hackaday, 2021. Pannenkoek2012's original video on the bounty: SM64 - TTC Upwarp $1000 Bounty.

  6. Faisal, Selen & Wolpert, Noise in the nervous system, Nature Reviews Neuroscience, 2008.

  7. Deco & Rolls, Stochastic dynamics as a principle of brain function. For the noise-as-feature angle specifically, Stochastic resonance - Wikipedia.

  8. Sreenivasan & Fiete, Grid cells generate an analog error-correcting code for singularly precise neural computation, Nature Neuroscience, 2011.

July 26, 2026

Out of band

Visible light covers about one octave of the electromagnetic spectrum. Everything else, from mains hum up through radio and microwave, crosses the skull continuously and leaves no trace, and over the last century we filled most of that range with our own traffic.

The head is not opaque to any of it. Kirschvink's group reported in 2019 that rotating an Earth-strength magnetic field around a shielded, seated human reliably drops occipital alpha power, the response a real sensory stimulus produces, while nobody reports feeling anything 1. Pulsed microwaves are audible outright: Frey's subjects in 1961 heard clicking with no acoustic source, each pulse heating tissue by about 10^-5 °C and sending a pressure wave to the cochlea 2.

Ambient RF in ordinary European microenvironments runs 0.33 to 1.72 mW/m², around 0.6 V/m in air, and neurons respond to imposed fields from 0.1 V/m upward 3 4. On amplitude alone WiFi should be doing something to us. It does not, because a patch of membrane is an RC circuit with a time constant near 10 ms, which corners it around 16 Hz and rolls off as 1/f above that, so 2.4 GHz arrives eight orders of magnitude down. Per quantum it is worse: 10 µeV, 2,700 times below kT at body temperature, enough to heat and not enough to break a bond. A muon leaving 2 MeV per centimetre of track is the opposite kind of object, which is why one argument survives and the other does not.

How much of an imposed field a membrane actually feels, against the frequency of that field.
How much of an imposed field a membrane actually feels, against the frequency of that field.

Mains at 50 Hz is the only thing we run that lands in band, and the only case where exposure limits are set by neural coupling rather than by heating. Residential exposure of around 0.1 µT induces microvolts per metre in tissue, four orders under the limit 5.

Engels et al. showed in 2014 that European robins in unscreened wooden huts on the Oldenburg campus lost magnetic orientation entirely, and that screening the huts across 50 kHz to 5 MHz brought it back, at noise levels a thousandth of the WHO limit 6. A bird sitting in the same room as you is jammed by the building's own wiring, because its compass runs on a radical-pair mechanism that holds coherence in exactly the band we fill with AM broadcast and switching-supply hash, and nothing in human cognition listens in that band at all.

European robin. Its compass stops working inside an ordinary building on an ordinary university campus.
European robin. Its compass stops working inside an ordinary building on an ordinary university campus.

The environment therefore divides into noise that arrives as an oscillation, which the membrane filters out of existence regardless of how much of it we generate, and noise that arrives as a single concentrated deposit, which lands in full and has to be tolerated. Evolution only ever had to solve the second kind, and nothing we transmit has ever belonged to it, so the robustness the original post argues for was never a defence against anything we built.

Footnotes

  1. Wang, Shimojo, Kirschvink et al., Transduction of the geomagnetic field as evidenced from alpha-band activity in the human brain, eNeuro, 2019. The same group found biogenic magnetite in human brain tissue, at least 5 million single-domain crystals per gram: Magnetite biomineralization in the human brain, PNAS, 1992.

  2. Microwave auditory effect - Wikipedia. Frey's original tests used 50 Hz pulse repetition with 10-70 µs pulse widths.

  3. Sagar et al., Radiofrequency electromagnetic field exposure in everyday microenvironments in Europe: a systematic literature review, Journal of Exposure Science & Environmental Epidemiology, 2018.

  4. Anastassiou & Koch, Ephaptic coupling to endogenous electric field activity: why bother?, Current Opinion in Neurobiology, 2015. Slice thresholds for DC and low-frequency AC fields sit around 0.5 V/m.

  5. ICNIRP guidelines for limiting exposure to time-varying electric and magnetic fields (1 Hz to 100 kHz), 2010. The basic restriction for CNS tissue in the head is 0.02 V/m at 50 Hz.

  6. Engels et al., Anthropogenic electromagnetic noise disrupts magnetic compass orientation in a migratory bird, Nature, 2014.

For any thoughts, comments, or questions, you can reach me at: mkalcher ethz.ch.

← Back to Home