A receipt for the mechanism, not a law — brain and machine invert on whether learning or using is the cheaper operation, yet both spend most of their energy signalling/inferring; 'training is the expensive phase' is a property of backpropagation, not a law of learning systems

neuroenergeticsbrain-energy-budgettraining-inference-asymmetryinference-economicssynaptic-plasticitymetabolic-constraintscross-domain-bridgebackpropagation

The recurring argument in this cluster is not "the brain is efficient" and not "the brain is like a neural network." It is a two-move claim about where a learning system spends its energy: per event, brain and machine invert — in the brain, changing a synapse is cheap and firing it is expensive; in silicon, the backward pass is expensive and the forward pass is cheap — yet at the system level the two converge, because both spend most of their total energy using what they learned rather than learning it. The inversion means the AI-native intuition that "training is the expensive phase" is a receipt for the backpropagation mechanism, not a law of learning in general.

The move that makes this a map and not two adjacent facts is the nesting: a per-unit reversal sitting inside a system-level agreement. "Which is cheaper, learning or using?" flips between substrate and silicon. "Where does most of the energy actually go?" does not — signalling/inference wins in both, for the same structural reason (signalling and inference events vastly outnumber learning events). Read one layer without the other and the cluster looks either paradoxical (the brain contradicts AI) or trivial (both use a lot of energy running). Read together, it says something specific: the asymmetry we treat as fundamental is a property of the update rule we chose.

Titled for the argument, not the recurring author (David Attwell anchors two of the five notes, but the recurring argument is the reversal-inside-convergence, not the man — the 2026-07-25 lesson).

The per-event reversal — in the brain, learning is the cheap operation

• claim-synaptic-plasticity-cheap-fraction-of-transmission-energy — the hinge fact. Karbowski's metabolic accounting of rat cortex: "the energy cost of synaptic plasticity constitutes a small fraction of the energy used for fast excitatory synaptic transmission, typically 4.0 − 11.2%." Changing a synapse costs roughly a twenty-fifth to a tenth of using it (transmission runs "on the order of ten to twenty-five times more" than plasticity, in the note's own words) — the inverse of the artificial-net ratio, where the backward pass is the expensive part (claim-training-inference-compute-asymmetry-mechanism). This is also the note that states the argument's core explicitly: the "training is the expensive phase" intuition "is therefore a property of the backpropagation-based substrate ... not a general law of learning systems." Tier 1, capture-verified.
• claim-competitive-plasticity-reduces-learning-energy — why it is cheap, and the arrow back to AI. van Rossum & Pache (2024) model plasticity-restricting rules ("only modify synapses with large updates," confined to coordinated subnetworks) and estimate that unrestricted backpropagation would need on the order of 100,000× more synaptic updates in a macaque-V1 model. The same rule, read in ML terms, is gradient sparsification — biology as a prior on which optimizations are worth trying. Tier 1. This is the note that keeps the cluster from being a mere curiosity: the reversal has a mechanism, and the mechanism travels back into engineering.

The system-level convergence — using dominates the total budget

• claim-attwell-laughlin-2001-grey-matter-signaling-energy-split — the load-bearing number. Attwell & Laughlin (2001): "Action potentials and postsynaptic effects of glutamate are predicted to consume much of the energy (47% and 34%, respectively)" — ~81% of grey-matter signalling energy, with signalling (not baseline housekeeping) dominating the metabolic cost. This is the brain-side counterpart to inference coming to dominate AI compute (claim-inference-dominant-ai-compute-2026). Tier 1, but abstract-only (SAGE full text paywalled) — the note stays seedling for exactly that reason.
• claim-howarth-2012-revised-brain-energy-budget-lowers-action-potential-share — the same team's revision, kept as its own dated claim. Howarth, Gleeson & Attwell (2012) re-modeled the budget after action potentials proved more energy-efficient than assumed: "most signaling energy (50%) is used on postsynaptic glutamate receptors, 21% is used on action potentials, 20% on resting potentials..." The action-potential share falls 47%→21%; the ~81% total becomes ~71%. The qualitative claim survives either estimate — signalling still dominates by a wide margin — which is why the convergence argument does not rest on the contested exact figure. Tier 1, abstract-verified.

The seam — the reversal nested inside the convergence

This is the note that fuses the two halves into one argument. Cut it and the cluster is "here are some brain energy numbers" beside "learning is cheap"; with it, the numbers and the mechanism become a single cross-domain claim.

• claim-brain-inference-bound-like-ai-at-system-level — the seam. Per event, learning is the cheap operation in the brain and the expensive one in silicon; at the system level both are signalling/inference-bound, because "inference/signaling events vastly outnumber learning events." The note's own framing is the thesis in miniature — "a per-unit reversal nested inside a system-level agreement": which is cheaper, learning or using, flips between brain and machine, while where most of the energy actually goes does not. (The sharper "property of the backpropagation-based substrate, not a general law of learning systems" line belongs to the Karbowski plasticity note above, not to this one.) Tier 1 observation; carries a [dated-figure] flag pinning the ~81% to the 2001 estimate and pointing at the 2012 revision — the map inherits that discipline.

Where this bridges to — the AI side, cross-linked not folded

The convergence half points straight into an existing map, and the notes already make the join; this MOC does not absorb it.

• moc-inference-economics — the AI-compute-economics cluster. The brain-side and silicon-side "using dominates" findings are the same shape in different substrates; kin, not member. But the silicon side is the worse-sourced half of the analogy, and the map says so: the claim that inference dominates AI compute rests on the Tier-3, flagged claim-inference-dominant-ai-compute-2026; the vault's own Tier-1 asymmetry note (claim-training-inference-compute-asymmetry-mechanism) records one real datapoint with training spend above inference; and moc-inference-economics itself marks the dominance statistic contested (myth-inference-two-thirds-of-compute). The convergence argument holds on the direction both substrates share — using events vastly outnumber learning events — not on any settled magnitude.
• backpropagation-gap and claim-brain-approximates-backprop-core-principles-ngrad — the vault's standing thread that the brain approximates backprop's principles without its literal global backward pass. This cluster supplies the energetic reason it must: an energy budget that forbids backprop's dense, every-weight-every-step update volume.

What this map is not

• It does not build the David Attwell entity hub. The same 2026-07-27 flag asks for one; Attwell anchors two notes (the 2001 budget and the 2012 revision) — at, not past, the threshold the vault has declined single-author hubs at all through this period, and building it now would be the entity flood the spec guards against. Left as a mention with its re-fire condition intact (promote when a second independent cluster leans on him). Named here, not silently skipped.
• The OU/stasis note is not a member. observation-mean-reversion-to-an-optimum-recurs-across-fossil-stasis-bonds-and-sgd shares the cross-domain-bridge tag and brushes the SGD/ML side, but its argument is Ornstein–Uhlenbeck mean-reversion across fossil stasis, bonds, and SGD — a different bridge. Left out on purpose.
• The 47%→21% supersession is an internal caveat, not a second argument. It rhymes with the vault's "famous number gets revised" theme but is not that story (a modeling estimate improved by the same lab as biophysics got measured better); kept as a dating discipline on the two budget notes, not spun out.

Open threads (honest caveats, not hidden)

• Every member is status: seedling, and the two budget numbers are abstract-only. Attwell & Laughlin (2001) and Howarth (2012) are grounded on the papers' own abstracts; the paywalled full-text results tables were not independently re-read. For figures this widely re-quoted that is a small gap, but it is a gap, and it is why the notes stay seedlings.
• The punchline leans on one figure per system. The cross-domain line is clean, but the brain side rests substantially on the Attwell/Howarth budget and the Karbowski ratio; the silicon side on the training/inference asymmetry notes. Weight it as a strong structural analogy, not a proven identity — the notes' own commentaries say as much.
• The exact brain total is contested (81% vs 71%); the direction is not. The map is built on the qualitative dominance, which holds under both estimates, and dates the precise figure rather than asserting one.
• The silicon side of the convergence is the worst-sourced premise here — and it is not hidden. "Inference dominates AI compute" is Tier-3 / flagged / contested in the vault's own record (see the kin entry above); the Tier-1 evidence supports the direction (using events outnumber learning events) but not the magnitude, and one Tier-1 datapoint even runs the other way on spend. The brain-side dominance (signalling ≫ housekeeping) is the better-sourced leg. The analogy rests on the shared direction, not on either substrate's exact share — an asymmetry an earlier draft of this map disclosed for the brain side only, corrected here.