Brain

Where the heck does this thing called experience come from??

^^ that’s the whole question I want to spend life getting to the bottom of

in what ways are we working towards it?

inspiring reads:

CAVE (infra behind both flywire & MICrONS dataset!)
- what bottlenecks to scale to 10mm^3? some early ideas:
- rn queries to historical states replay all $K$ edits, in $O(K)$ time, which lags users editing a heavily-modified neuron. A partially persistent DSU storing temporal edges as $parent[u] = (v, t_{edit})$ could make traversal with union-by-rank take at worst $O(log N)$ time (N=total neurons).
- Since in Discussions they mentioned supervoxel size is the primary limiter for scaling the dataset, we could have the ChunkedGraph keep a low-resolution node set for fast reading in BigTable (enough info to suspect errors) while offloading the raw high-res AI affinity maps to GCS. This way the engine can lazily fetch fine pixel data only when a specific bounding box is flagged for a split. Alternatively, we could also experiment with implicit neural representations for compression (train neural nets to map (x, y, z) to structural density).
- CAVE’s support of automated proofreading was really interesting (and important for scale). I wonder if instead of using external tools like Neurd that download data from ChunkedGraph API and compute their own skeleton, we could run a GNN/3D CNN directly on existing L2-Cache skeletons to detect anomalies quickly.
POLI Fiber for MRI-compatible neural recording that unifies optical/electrical/chemical modalities. (wonder if autoencoders could be adapted to track and subtract that baseline capacitance drift to automate detection of quick dopamine oxidation peaks)
piezo ultrasound for deep brain stimulus https://pmc.ncbi.nlm.nih.gov/articles/PMC11150473/
Gut is central to nervous system, and thus awareness too!
https://www.nature.com/articles/s41587-023-01833-5
https://pubmed.ncbi.nlm.nih.gov/39394431/

multiagent models of mind on lw: https://www.lesswrong.com/s/ZbmRyDN8TCpBTZSip

they’re training models to reconstruct what mice see from brain activations https://elifesciences.org/articles/105081. it’s still only a first step tho. they only probed neurons in a tiny tiny patch (630 micrometer across). but even if we can measure more, maybe we’ll never reconstruct the actual image entirely since mice brains probably lose info when they turn light into experience. ideally we should be reconstructing subjective experience directly, not the external movie, but mice can’t tell us what they see, and no human’s going to gene edit their babies to have glowing neurons lol

omg this is amazing. 10 days. at home! https://brainhack.vercel.app/ae

alb is so cool. https://writetobrain.com/olfactory

^^^ hey i can acc do something like that this summer

deep brain stimulation noninvasively with interfering waves hmm

50 minutes 2026-04-18 let’s actually build a mental model. write chain-of-thought here.

hypothesis: we can’t achieve it (pls be wrong!)
noninvasive telepathy is physically impossible because of skull attenuation and resolution needed
even w/ single single neuron resolution, how experience is encoded in neurons is unique to every individual

so this splits into two parts: imaging and decoding

let’s start with part 1 the decoding

what resolution do we need / is possible?

neurons are ~5e-5 m in soma diameter, up to >1m long (down the spine)
neuron counts: humans 8.6e10, mice 7e7-1e8, fly 1e5
roughly equal amount of glial (non-neuronal) cells that supply nutrients, immunity, signal insulation (myelin)
dense network of arteries, capillaries, veins
blood flow propto neural activity <- what all functional neuroimaging captures

from the outside in: 
scalp ~7e-3 m (skin, connective tissue, fat, membrane)
skull ~7e-3 m (hard, spongy, hard)
meninges ~3e-3 m (3 layers of stiff membrane, sandwiching cerebrospinal fluid)
dura mater
arachnoid mater
this gap with csf has all the major arteries/veins
pia mater
gray matter ~2.5e-3 m (processing) heavily vascularized
white matter (myelinated wiring)

where’s the information encoded and how precise do we need?
action potentials 1ms so we need 1kHz/2=2kHz by sampling theorem
we’ll need <50 micron spatial resolution for individual neurons
maybe 1e4-1e6 channels for full telepathy?
empirical evidence:
neuralink: 1024 channels (64 threads) and 20khz per channel and they can decode motor intentions
invasive speech decoding in patient w/ locked-in: 384 channels, ~400 microns, 30kHz raw sampling; ~50% accuracy in 50-word vocabulary
	

how close are we? i.e. what methods
electrical (directly measure voltage from action potentials)
EEG is fast (<1ms) but the informations all smeared by skull (spatial in cm range)
ECoG is invasive, 1-100 microns, 5ms
magnetic (directly measures tiny magnetic fields from neurons’ current)
MEG isn’t smeared like electric
hemodynamic (proxy)
fMRI tracks blood oxygen level-dependent (BOLD) signal
fUS tracks changes in blood volume in small vessels <- emerging
[TODO] what do contrast agents do here?
PET uses radioactive tracers
fNIRS uses light
optical (directly watches light interacting w/ tissue)
multiphoton calcium imaging genetically modifies neurons and has to open a window in the skull.
https://neuro.lev.la/ <- this is awesome!
most of these have >=1mm spatial, but we need <100micron=.1mm. AO is the one that’s close.
hybrids (combinations of previous 4)
photo/thermo -acoustic (PAT, TAT): laser/radio pulse fired into brain; blood/tissue absorbs, heats, expands, emits ultrasound, that is detected
acousto-optical (AO): ultrasound + laser fired simultaneously
magneto-acoustic (MAET, MAT-MI): ultrasound vibrates conductive tissue through magnetic field creating measurable current


how robust is proxying with blood flow? idk it feels too slow compared to thought
well clearly we can train an embedder that maps 7T fMRI timeseries to a good diffusion model prompt.
natural scenes dataset that dynadiff used has 1.8-mm isotropic, 1.6-s sampling, 32 channels
theoretically a perfect map of blood structure/flow would provide much higher resolution for more accurate images
[TODO] is there any catches here?
but time-wise, hemodynamic response peaks 5-6s after stimulation and more to return to baseline. that’s not enough for say realtime video, but still very powerful if we can get images

could we achieve a perfect live reconstruction of vasculature?