Taste as Bottleneck
The cost of a competent paragraph fell roughly two orders of magnitude between November 2022 and early 2026. The cost of knowing which paragraph to ship did not move. When generation gets cheap, selection becomes the job — and selection is the part nobody figured out how to teach at scale.
The case
Every craft has two costs: producing a candidate, and choosing which candidate is worth shipping. For most of industrial history, producing dominated. A junior copywriter spent four hours on a draft; a senior spent twenty minutes deciding which of three drafts to ship. The senior's twenty minutes looked like overhead because the four hours were the visible work.
Reverse the ratio. GPT-4-class models can produce 30 competent drafts in the time it takes to read one. Cursor and Claude Code can produce a working pull request in the time a senior engineer takes to write the issue. Midjourney v6 can produce 200 logo candidates in an afternoon. The four hours collapsed. The twenty minutes did not.
This is the bottleneck shift. Pre-2022 the constraint was can you produce it. Post-2024 the constraint is can you tell which one is right. The first is a skill you can hire for at a known price. The second is harder to name, harder to test, and — the contested claim — possibly harder to teach.
Where it shows up
| Field | Old bottleneck (~2020) | New bottleneck (~2026) |
|---|---|---|
| Software | Writing the function | Knowing which function not to write |
| Copy | Drafting the post | Picking the one of forty that won't sound generic |
| Design | Producing the comps | Killing the comps that almost work |
| Research | Summarising the paper | Knowing which question to ask the model |
| Buying (fashion) | Working a range plan in Excel | Knowing which 30 styles of 3,000 generated options to PO |
In every column the right-hand cell describes a smaller number of people doing higher-stakes decisions. The pyramid inverts. This is the structural fact behind the 2024-2026 layoffs at the bottom of knowledge-work pyramids and the simultaneous bidding war for senior judgment.
What's contested
Two camps, both sincere.
Camp A — taste is craft, and craft is teachable. The argument: senior engineers got senior by reading 100,000 lines of other people's code and watching what shipped well versus what shipped badly. Give a junior the same volume of curated examples plus an LLM to accelerate the loop and they reach senior judgment in three years instead of ten. The bottleneck is pedagogy, not biology. See concept deliberate practice for the underlying claim.
Camp B — taste is a long apprenticeship in failure, and the loop just got shorter without getting easier. The argument: judgment comes from owning the consequence of a bad call. A junior who never shipped a broken migration, never lost a $50k buy, never had a customer cancel because the copy was off — that junior has nothing for the LLM to amplify. The accelerated loop trains pattern-matching, not consequence-bearing. See concept skin in the game.
The empirical question — does Camp A or Camp B describe the next decade — is unresolved as of mid-2026 and probably won't resolve before 2030. Both camps have working examples. Neither has a clean controlled study.
Why this has to do with other realms
The taste-as-bottleneck story echoes a much older pattern from fashion: when sewing machines made garment construction cheap in the 1850s, the bottleneck moved from making the dress to knowing which dress would sell. The buyer profession was invented to fill that gap. A buyer's edge is exactly what an AI-era engineer's edge is becoming: high-volume exposure, internalised customer model, willingness to commit and own the call.
Industrial revolutions don't eliminate work. They relocate the scarce skill upstream of the newly cheap step. The 1850s sewing machine made a tailor's hand-stitching obsolete and made a buyer's eye load-bearing. The 2020s transformer made a coder's typing obsolete and made an architect's judgment load-bearing. Same shape, different century. See concept inditex playbook for where this logic gets industrialised.
An open question
If taste is the bottleneck, and taste comes from owning consequence — who pays the apprenticeship cost in a world where junior roles get automated away before the consequence arrives? The 2026 hiring graphs show the gap is already widening. The question is whether the answer is institutional (firms invest in slow apprenticeships again), individual (juniors self-fund their own consequence-bearing through side projects), or structural (a missing generation of senior talent shows up in 2035).
Key sources
- The Coming Wave by Mustafa Suleyman (2023) — frames the diffusion-and-containment problem; the labour-pyramid argument is implicit but consistent.
- to verify: McKinsey "Economic potential of generative AI" report (June 2023) — the original numerical claim about white-collar task automation share, frequently misquoted.
- Skin in the Game by Nassim Taleb (2018) — the canonical statement of Camp B's intuition about consequence and judgment.
- to verify: Anthropic / OpenAI internal evaluations on coding-task automation rates 2024-2026 — public summary versions appear in their model cards; precise numbers shift quarter to quarter.
- Where Good Ideas Come From by Steven Johnson (2010) — argues judgment is downstream of exposure volume; useful for steel-manning Camp A.
Further reading
- concept deliberate practice — Ericsson's framework, the strongest version of "judgment is teachable".
- concept skin in the game — Taleb's framework, the strongest version of "judgment requires consequence".
- The Cult of the Amateur by Andrew Keen (2007) — pre-LLM but exactly the wrong shape of argument; useful to read against.
- Karpathy's "Software 2.0" essay (2017) — the earliest sharp statement of generation getting cheap.
- Stratechery archives 2023-2026 (Ben Thompson) — the most consistent operator-side commentary on where the bottleneck moves quarter by quarter.
Abhishek's take
The honest answer is that taste is partly teachable and partly not, and the un-teachable part is the part that scales. You can teach someone to recognise a bad PR; you cannot teach them the gut-check that says we shouldn't be building this at all. The first is pattern-matching on 10,000 examples. The second is some compound of having lost money, having been wrong publicly, and having read outside your field for ten years — and the AI-era pyramid is structured so fewer people will ever get there.
Where I've used this
In the buying floor tools I shipped 2024-2026, the AI does the volume work: generating range options, scoring styles, drafting POs. The judgment calls — which 30 of 300 candidates to actually buy, which vendor to push back on, when to override the model — stay human and stay senior. The tools made my juniors faster at producing; they did not make them faster at deciding. That gap is the entire story.
Tags: #taste #judgment #ai-leverage #scarcity #hiring #craft
See Also
- concept deliberate practice — the strongest version of the teachable-taste argument.
- concept skin in the game (cross-realm: risk/philosophy) — why consequence-bearing might be load-bearing for judgment.
- concept inditex playbook (cross-realm: retail-ops) — what happens when an industry industrialises around taste-as-bottleneck.
- concept jacquard loom (cross-realm: history-of-tech) — the 19th-century precedent for cheap generation reshaping a craft hierarchy.
- concept bus factor (cross-realm: distributed-cognition) — the failure mode when load-bearing judgment lives in too few heads.