The Human Shannon Number, and Why It Is Not a Boast

Shannon’s famous number is not about information theory. It is about chess: he estimated the number of possible games at around 10^120, which is a way of saying far more than you could ever enumerate, and he used it to argue that brute-force search would never solve the game.

The Human Shannon Number borrows that move and applies it to knowledge instead of moves. Let each domain have a number of distinguishable knowledge units, k, weighted by how much that domain interacts with others, alpha. Then the total space of possible knowledge states is a product:

H = product over domains of k_i ^ alpha_i

Written in logs, because otherwise it is unreadable:

log H = sum of alpha_i log k_i

This is not a measure of what we know. It is the size of the space of possible human mental states and their structured outputs, and it is designed to be astronomically larger than anything reachable.

Why it is not a boast

The number gets quoted as though it were a claim of near-omniscience — “humanity has explored less than a millionth of a percent of all possible knowledge.” That reading is backwards, and it is worth being blunt about it.

The space is mostly not useful. It contains every true but pointless statement, every redundant formulation, every structure that cannot be reached, built, or recognised by anything. Its size is a fact about combinatorics. Finding it large tells you nothing about what remains worth doing.

What matters is the gap between the whole space and the accessible region. The empirical work suggests the accessible region in many classical domains is largely filled, while the total space is untouched. Both statements are true at once, and only the first one has consequences.

The number is a monument, not a target

I have come to think of H as a boundary marker. It is useful for one job: expressing how little of the space is reachable, so that “there is plenty left to discover” can be answered precisely rather than rhetorically. It is not a roadmap and it is not computable in practice — k_i and alpha_i are not measurable in any clean way, and the product is dominated by whatever you assume for the largest term.

That weakness is why the more recent work in this programme abandons the formula and goes after something computable instead. If you want a quantitative account of discovery limits, you need a quantity you can measure from data rather than a quantity you assert from a taxonomy.

The candidate is description length, and it turns out to make the whole thing sharper: novelty becomes the rate at which a better model compresses the same observations. That is the subject of a later post in this archive, and it is the part of the programme I would defend hardest.