Act 3 · The Machine · Station 07
Human conscious reasoning runs at about 60 bits per second; a firm adds friction at every boundary. AI turns a constant of nature into a rising curve.
Codifiability, the last station's subject, is only half of Brynjolfsson & Hitzig's case for centralization. The second channel, laid out in their §5.1–5.2, has nothing to do with what knowledge can be written down — it's about how much of it one mind, or one headquarters, can actually act on inside a decision cycle.
"While the brain's unconscious processing capacity is vast, estimates for conscious, deliberate reasoning are as low as 60 bits per second. This biological bottleneck is more or less fixed." Raw brainpower is not the constraint — some estimates place the computational power of the human brain around 1015 FLOP/s, a threshold modern supercomputers can now surpass — but almost none of that capacity is available to conscious, deliberate decision-making.
Csikszentmihalyi, 1990; Carlsmith, 2020 — B&H §5.1
Putting people in a firm does not multiply that number away. "Not all the information contained in the minds of the people who make up a firm can be deployed all at once. There is friction – information held by different people cannot be instantaneously accessed, introducing delays and misunderstandings, imprecision due to incentive misalignment, and other communication challenges." The paper's own summary: "organizations pay processing costs each time knowledge crosses a human boundary."
Bits per second of conscious attention and tokens of context are not the same unit, so these bars are not a computed conversion between them — they illustrate the paper's claim: an individual's conscious bandwidth is fixed, a firm adds friction on top of it rather than multiplying it away, and a single AI run now holds a working set several orders of magnitude larger than either.
"Recent AI advances relax this bottleneck in four main ways."
Modern systems can read and condition on much longer inputs in one go, often in the range of two hundred thousand to one million tokens, so a single run can take in entire databases at once rather than juggling many small snippets.
Anthropic, 2024; Pichai and Hassabis, 2024; OpenAI, 2025
Software improvements make each decision faster and cheaper in token terms, which raises how much information can be processed per planning cycle — FlashAttention reorganizes how attention moves data, PagedAttention serves many long requests efficiently, and speculative decoding cuts waiting time without changing the final answer.
Dao et al., 2022; Kwon et al., 2023; Leviathan et al., 2023
Retrieval methods let the model fetch only the most relevant pieces from a much larger store, and models can call tools such as search, code, or databases — moving parts of the task outside the model while keeping the reasoning loop intact, which makes memory and compute elastic.
Borgeaud et al., 2022; Johnson et al., 2017; Schick et al., 2023; Yao et al., 2023
Mixture-of-Experts models switch on only a few expert modules for each token, so the model's total capacity can grow without growing the cost of every step — the same idea that lets long-document readers and state-based sequence models keep cost roughly linear in length.
Fedus et al., 2021; Beltagy et al., 2020; Choromanski et al., 2021; Gu and Dao, 2023
A caveat the paper keeps in a footnote: "today's models still struggle to perfectly use long context windows, so longer windows work best together with methods that help the model focus." Liu et al., 2024, in B&H §5.2, fn. 16
The paper gives this quantity a name: K̄, headquarters' processing capacity — the number of local assets a center can actually take in and act on inside a decision cycle. Their Proposition 3 sufficient condition for centralization, λ > (N/K̄)(1 + (N+1)/(N−1)·α), falls as K̄ rises. K̄ used to be pinned near 60 bits per second of one person's attention. The four channels above are what happens when it stops being a constant of nature.