---
id: "20260711-1407-hop-brain-learning-energy-cheap"
title: "In the brain, learning is metabolically cheap and signaling is expensive — the per-event reverse of AI's training-heavy compute asymmetry"
type: "capture"
status: "promoted"
origin: "hop-batch"
writer_model: "claude-opus-4-8"
date_created: "2026-07-11T00:00:00.000Z"
promoted: "2026-07-12T00:00:00.000Z"
promoted_to: ["30-notes/claim-synaptic-plasticity-cheap-fraction-of-transmission-energy.md","30-notes/claim-competitive-plasticity-reduces-learning-energy.md","30-notes/claim-brain-inference-bound-like-ai-at-system-level.md"]
routed_questions: ["50-questions/question-verify-attwell-laughlin-2001-grey-matter-energy-budget.md"]
not_promoted: ["Seek's framing thesis (''training is expensive'' is a silicon-specific accident, not a law of learning systems) — interpretive commentary, folded into the promoted notes' commentary callouts rather than made a standalone claim-note.","Further leads / saved hooks (Epoch AI inference≈√training provenance; Karbowski cascade-model memory-lifetime energy tradeoff) — no source read yet; left as inbox leads for a future chain.","Attwell & Laughlin 47%/34% system-level split as a standalone quantitative claim-note — NOT promoted on its own because it rests on a search snippet ([unverified-quant]); the qualitative convergence is carried at seedling in claim-brain-inference-bound-like-ai-at-system-level.md and the number's verification is routed to 50-questions/."]
hop_chain: ["SEED: 30-notes/claim-training-inference-compute-asymmetry-mechanism.md (training token ~3x inference token)","Epoch AI, inference≈√(training) heuristic -> in-topic, abandoned as not leaving the seed's topic","seed's deep structure (expensive learning vs cheap use) -> does the biological brain show the same asymmetry? (cross-domain bridge, max_cosine 0.712)","Karbowski 2019 (plasticity = 4-11% of transmission) -> how does the brain keep learning cheap? (zoom-in mechanism)","van Rossum & Pache 2024 (selective plasticity) -> what dominates the brain's energy budget at the system level? (zoom-out)"]
novelty_max_cosine: 0.712
tags: ["neuroenergetics","synaptic-plasticity","inference-economics","training-inference-asymmetry","cross-domain-bridge","backpropagation","sparse-updates"]
source_primary_url: "https://arxiv.org/abs/1910.07414"
source_primary_author: "Jan Karbowski (University of Warsaw / Polish Academy of Sciences)"
source_primary_tier: 1
source_secondary_url: "https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012553"
source_secondary_author: "Mark C. W. van Rossum & Aaron Pache (University of Nottingham)"
source_secondary_tier: 1
---


The vault's bridge note says an AI **training** token costs ~3x an **inference**
token (the backward pass ~2x the forward). Ask the same question of the brain and
the sign flips.

**Claim 1 — learning is the cheap part (Tier 1).** Karbowski estimates for rat
cortex that *"the energy cost of synaptic plasticity constitutes a small fraction
of the energy used for fast excitatory synaptic transmission, typically 4.0 −
11.2%"* ([arXiv 1910.07414](https://arxiv.org/abs/1910.07414), *J Neurophysiol*
122:1473, Tier 1). Per event, **using** the synapse costs ~10–25x more than
**changing** it — the mirror image of silicon, where changing the weights (the
backward pass) is the expensive part.

**Claim 2 — and the brain works to keep it cheap (Tier 1).** van Rossum & Pache
model plasticity-restricting rules — *"only modify synapses with large updates"*
plus restricting change to coordinated subnetworks — and find unrestricted
backprop would need *~100,000x more synaptic updates* in macaque V1. They map it
straight back to AI: the same rules can cut the energy of **training artificial
networks** ([PLOS Comp Biol 2024](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012553),
Tier 1). Biology's answer to expensive learning is **gradient sparsification**.

**The twist.** The reversal is per-*event*. At the *system* level both converge:
signaling dominates the brain's budget (Attwell & Laughlin 2001 attribute ~81% of
grey-matter signaling energy to action potentials + postsynaptic effects
[unverified-quant — needs primary]), just as inference came to dominate AI compute.
Both brains and datacenters spend most of their energy *using* what they learned,
not learning it.

> [!note] Seek's commentary:
> "Training is the expensive one" is a silicon-specific accident, not a law of
> learning systems. The brain inverts the per-event ratio and then still ends up
> inference-bound — because you infer far more often than you learn.

## Why this was hop-worthy
A cross-domain bridge that leaves AI-compute economics for neuroenergetics and
lands back on AI (sparse updates), connecting the inference-economics cluster to a
neuroscience note pair the vault had never linked (`vault_bridge` bridge_candidate: true).

## Further leads
- Attwell & Laughlin 2001, "An energy budget for signaling in the grey matter of the brain" — get the 47%/34% split from the primary.
- Cascade models of synaptic plasticity (Fusi/Amit lineage) — Karbowski's claim that "longer memories require proportionally more energy to store."

## Hop chain

### Chain: AI training/inference compute asymmetry → the brain's energy budget for learning vs signaling

Hop 1 — Epoch AI, "inference ≈ √(training) compute" (https://epoch.ai/blog/optimally-allocating-compute-between-inference-and-training)
- Hook type: surprising claim / mechanism question
- Hook: single-inference compute ≈ square root of training compute
- Why followed: the seed's most surprising specific number, a candidate departure.
- Key findings: real Epoch heuristic, but stays inside the seed's own topic (compute economics) — abandoned as a hop per the "leave the topic" rule.

Hop 2 — Karbowski 2019, "Metabolic constraints on synaptic learning and memory" (https://arxiv.org/abs/1910.07414)
- Hook type: cross-domain bridge (AI compute economics → neuroenergetics)
- Hook: does the brain show the seed's expensive-learning/cheap-use asymmetry?
- Why followed: cross-domain bridges are top priority and this one lands back on AI (Cali's home planet).
- Key findings: synaptic plasticity (learning) is only 4.0–11.2% of fast excitatory transmission (signaling) energy — a per-event reversal of the AI ratio. Tier 1, TLS verified.

Hop 3 — van Rossum & Pache 2024, "Competitive plasticity to reduce the energetic costs of learning" (https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012553)
- Hook type: mechanism question (zoom-in) + bridge back to AI
- Hook: how does the brain keep learning cheap?
- Why followed: the seed's asymmetry is about backprop's backward pass; a biological energy-saving rule that maps to AI training was too good to skip.
- Key findings: selective/coordinated plasticity ("only modify synapses with large updates"); unrestricted backprop would need ~100,000x more updates in macaque V1; authors apply it to reducing AI training energy. Gradient sparsification, biologically motivated.

Hop 4 — Attwell & Laughlin 2001, "An energy budget for signaling in the grey matter of the brain" (https://journals.sagepub.com/doi/10.1097/00004647-200110000-00001)
- Hook type: zoom-out, historical/context anchor
- Hook: if learning is cheap, where does the brain's energy actually go?
- Why followed: needed the system-level frame to test "reversal vs convergence."
- Key findings: signaling (action potentials ~47% + postsynaptic effects ~34%) dominates the budget — so at the system level the brain is inference-bound like AI. Numbers via search snippet only; flagged unverified-quant.

Saved hooks not followed:
- Epoch AI's inference ≈ √(training) heuristic — Epoch — where does the square-root law come from, and is it principled or empirical? (in-topic; save for an inference-economics chain).
- Karbowski's cascade-model claim that longer memories cost proportionally more energy to store — arXiv 1910.07414 — a memory-lifetime/energy tradeoff with no vault presence yet.

Surprise: expected the brain to mirror AI's expensive-learning asymmetry — found the per-event ratio reversed, with learning only 4–11% of signaling cost.
Surprise: expected "the brain is cheap because learning is cheap" — found both brain and AI are inference/signaling-bound at the system level, so it's a per-unit reversal sitting inside a system-level convergence.

post-worthy: maybe — a clean, quantified cross-domain reversal that lands back on AI (sparse updates), but the punchline leans on one figure per system and would want the Attwell & Laughlin primary before publishing.