the expensive energy principle

could a computer scientist pay for an embodied brain

August 21, 2026

just read a Quanta article.

incorrect same way llms often are.

original paper here.

todos are left for some future me to read; he'll be a smarter person anyway.

help, I need some body

5% extra thinking tokens.

that's what you spend.

the rest is body.

[we maximize info transmission per ATP spent,] under 10 hertz. Our neurons have an average firing rate of 4 hertz, 50 to 60 times less than what is optimal for [maxxing raw Shannon capacity].

info trasmit optimality refs (50x4Hz = 200) might come e.g. from here. todo but that has some weird assumption like 100% firing rate.

many synaptic transmissions fail: Even when an electrical signal is sent to the synapse, priming it to release molecules to the next neuron, it will do so only 20% of the time.

wtf, now that seems awfully energy inefficient.

todo. seems like this is hierarchical optimization at play (ref).

the energy expense here would be

~21% of signaling energy → action potentials
~50% → postsynaptic glutamate-receptor currents
~5% → presynaptic transmitter release itself
~4% → transmitter recycling.

so synaptic failure saves most of the cost.

you also do not strictly need the full information.

todo.

g-pretty-please

let's take as fact.

performing an externally directed cognitive task changes human brain's ongoing energy consumption by ~5%.

todo. going further back, seems a consistent 5% is achieved simply by turning off specific networks.

my brain is tired today, so borrowing some gpt compute (edited for brevity) gives:

An adult human brain runs at roughly 17–20 Wh. [Over 24h it's] ~ 350–410 dietary kcal (...) Goal-directed cognition appears to increase whole-brain metabolism by only about 5%. “hard thinking” is an extra ~ 0.85–1 W.

ok so what about llms

it's good google already did it