PK Wiki

concept · concepts/The Bottleneck Shift.md

The Bottleneck Shift

status: active · created 2026-06-02 · updated 2026-11-08 · 6 sources · #ai-and-work #expertise #judgment #mollick

When AI makes execution cheap, the scarce input shifts from doing the work to deciding what work to do. Articulated by Ethan Mollick and independently by Dan Shipper — the strongest cross-source convergence in the wiki so far.

The idea

Ethan Mollick (Making AI Work — Leadership, Lab, and Crowd): once intelligence is on demand, the constraint moves up a level —

  • "When research that once took weeks now takes minutes, the bottleneck isn't the research anymore, it's figuring out what research to do."
  • "When code can be written quickly, the limitation isn't programming speed, it's understanding what to build."
  • "When content can be generated instantly, the constraint isn't production, it's knowing what will actually matter to people."

This reframes the value of cheap output: when "companies used to pay tens of thousands for a single research report" and "now they can generate hundreds for free," the open question becomes what was the report even for?

Convergence (two sources, same insight)

Independent agreement: Dan Shipper in The AI Daily Brief — Why Agents Still Need Humans makes the same move — "the current generation of models only knows about work that has been done; humans know about what needs to be done right now" (AI Commoditizes Expertise). The same scarcity also drives The Infinite Backlog (the human's job becomes choosing what's next for the agent) and The Human Sandwich (the human "frame" and "what should happen next?" steps).

Two writers, different framings (organizational vs. economic), converging on: judgment about what-to-do is the durable human bottleneck. A candidate synthesis page (wiki/syntheses/) if the human wants it filed.

Why judgment stays scarce (cognitive mechanism)

Willingham supplies the why that Mollick and Shipper assert. Judgment about what to do isn't a generic skill that can be lifted out and packaged. It is bound to domain knowledge and practice. Knowing which research is worth doing needs the knowledge to judge plausibility, spot anomalies and see a problem's deep structure. The corollary, unaddressed by either author: if execution and the knowledge it used to build both move to AI, the judgment may thin out too. See Method Reflection (T3).

Empirical evidence: science (2026)

Anthropic — How Scientists Use Claude supplies usage data, the first in the wiki for this claim. Across 207k research sessions:

  • Ideation and hypothesis generation get 4.2% of Claude sessions and 2.1% of Claude Code sessions, against 7.1% of researcher time.
  • Surveyed scientists rank ideation as the most important task for producing important results.
  • The authors name "choosing what to study" as a possible bottleneck: "If ideation is a bottleneck, it will limit acceleration even as answering a given question becomes much cheaper" (AI-Accelerated Science).

That is Mollick's "figuring out what research to do," measured.

Contradiction: Mollick and Shipper treat what-to-do judgment as the durable human bottleneck. Anthropic — How Scientists Use Claude treats it as open. AI "may also be genuinely weaker at high-level ideation," but "if AI outstrips humans in formulating high-quality hypotheses, acceleration could be more substantial." In its survey of AI-adopting scientists, 72% expect AI to match top researchers at ideation by 2030. Specialized tooling (Claude Science) already brings ideation to about 8% of sessions, close to its share of researcher time. Use may be low because humans prefer to do this work, or because AI can't do it yet. The data doesn't settle which, and the answer decides whether this bottleneck is durable or temporary.

A case where the line moved a little

Anthropic — Investigating Vault Particles with Claude keeps the division the contradiction above is about, and then blurs it. The humans chose the problem (vaults, unexplained for 40 years), chose which 44 proteins to test, and decided which structure to solve — textbook what-to-do judgment. But two contributions came from the model side:

  • Method ideation. Claude proposed loading ruptured-cell contents straight onto the chromatography column with no pre-purification — an idea the team called unconventional and was reluctant to try. It worked, and is the reason purification dropped from three days to 90 minutes.
  • A candidate hypothesis. Claude flagged an unidentified molecule in a pocket of each of the 96 protein copies and noted the shape resembles coenzyme A — the paper's only hint at the function nobody has found since 1986.

Neither is "what research to do"; both are "how to do it" and "what might be going on here," which is nearer the ideation line than this page's sources assume. Note the symmetric data point: Claude also argued against the chromatography approach that worked. Its judgment about method was valuable and wrong in the same project.

In scientific modeling: build moved, interpret didn't

Anthropic — Teaching AI to Build Scientific Models places the line inside a single research step. Designing a model's equations "was thought to be something only humans could do"; LLM agents now do it. What stays hard to delegate is interpretation (what the model says about the real system), because fit has an objective and "understanding" doesn't. The authors add a second human-heavy choice: which experiment to run next. They say the most useful one "rule[s] out the greatest number of remaining models, even if [it is] not always the one a human scientist finds the most interesting." That partly turns the what-to-do decision into a calculation, once a distribution of candidate models exists.

Connections

Sources

The wiki is the source. This door reads it at build time and never writes to it.