notessamuele onelia

AI doesn’t make you replaceable, it makes everyone the same

Idea

Because every model is trained on the same residue of past human work, AI’s default output converges to sameness (“slop”), which is exactly why the live, present-moment judgment only a human has is what stays valuable.

Notes & Verified Quotes

Curation, relay-an-idea. Source: Dan Shipper (cofounder/CEO of Every), “After Automation.” AI-locked framing: the differentiation payoff is a consequence of how LLMs are built and distributed, not a generic “be different” lesson.

The mechanism — models are trained on the exhaust of finished work, then hand it back cheaply to everyone:

QUOTE: “AI commoditizes the residue of human expertise—whatever can be made explicit enough to train on” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation

What that produces by default when everyone uses the same model:

QUOTE: “Slop is visible sameness, repeated ad nauseam.” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation

The flip — sameness manufactures demand for difference:

QUOTE: “the work that doesn’t fit the pattern becomes the rare, valuable, and high-status thing” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation

Why the difference has to come from a human (the load-bearing distinction): models know the done, humans know the needed-now.

QUOTE: “The current generation of models only knows about work that has been done. Humans know about what needs to be done, right now, at this moment.” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation

QUOTE: “Once a situation has been reduced to text, once it has become corpus, it is a corpse.” — src: scratch/sources/after-automation.txt | source: https://every.to/p/after-automation

My angle (to develop): the fear is “AI will replace me.” The real shift is “AI makes the default version of me identical to everyone else’s default.” The defense isn’t working faster, it’s the present-moment judgment that can’t be in the training set yet.

Research dossier (added via /ae-research)

1. The sameness isn’t a vibe, it’s measured. Shipper asserts AI defaults to sameness; empirical work backs it. A preregistered study (Kibum Moon, Adam Green, Kostadin Kushlev, Georgetown) compared thousands of real college-admissions essays to ChatGPT-written ones and found each extra human essay added more new ideas than each extra GPT-4 essay — the effect grew with sample size and survived prompt/parameter tweaks.

QUOTE: “the widespread use of LLMs could diminish the collective diversity of ideas” — src: scratch/sources/llm-homogenization-kushlev.txt | source: https://www.kushlev.com/latest-news/2024/9/12/do-large-language-models-llms-have-a-homogenizing-effect-on-creative-diversity

This is the mechanism made concrete: independent humans spread across the idea-space; independent model samples cluster. Shared corpus → narrow distribution → sameness. (Hard data turns “slop” from an aesthetic complaint into a structural claim.)

2. The named defense is “taste.” Peter Deng (former OpenAI product head, ex-Instagram/Uber/Facebook; now GP at Felicis) frames the next wave as taste over models — the same move as Shipper’s “demand for difference,” from the builder’s side.

QUOTE: “it’s not going to be about who has the best model. It’s going to be about who has the best taste” — src: scratch/sources/peter-deng-taste.txt | source: https://www.youtube.com/watch?v=uOeFDTHBJCE QUOTE: “the model is not the differentiator. It’s the workflow. It’s the taste and it’s the choices in the product” — src: scratch/sources/peter-deng-taste.txt | source: https://www.youtube.com/watch?v=uOeFDTHBJCE

Why the floor drops and the premium moves to taste: when building is trivial, sameness is the default and difference is the scarce thing.

QUOTE: “the barrier to building is going to be super low” — src: scratch/sources/peter-deng-taste.txt | source: https://www.youtube.com/watch?v=uOeFDTHBJCE

3. The standards-ratchet corroborates Shipper directly. Shipper: a new model floors you, then “a few months later they feel ordinary.” Deng independently:

QUOTE: “our bar as humans keeps on increasing. What we find interesting today, we may not find interesting tomorrow.” — src: scratch/sources/peter-deng-taste.txt | source: https://www.youtube.com/watch?v=uOeFDTHBJCE

The human bar rising is why the demand for difference never gets satisfied — every wave resets it. This is the present-moment judgment point in motion.

Second Brain corroboration (no new quote needed): The Market Rewards Asymmetrical, Scarce Insight (Yossi Levi / Car Dealership Guy) is the same thesis from the creator side — “if a hundred other creators can produce your post, the post is worthless.” Plus the already-linked If a Stranger Could Have Posted It, Delete It and The N of One - 80th Percentile in Three Things.

Candidate angles for the outline:

  • A — Reframe the fear: the threat isn’t replacement, it’s homogenization; lead with “AI doesn’t replace you, it averages you,” then taste/judgment as the way out. (Strongest; matches headline.)
  • B — Proof-first: open with the homogenization study (humans spread, models cluster), then Shipper’s mechanism, then the taste defense. More argumentative/data-led.
  • C — Builder’s mirror: Deng’s “best taste, not best model” + Shipper’s “demand for difference” → for knowledge workers, taste = your present-moment read on what matters now. Practical “here’s the moat” close.

Where evidence is thin: the BCG/GPT-4 “less conceptual variation” finding showed up in search summaries but I didn’t capture a verifiable primary source, so it’s omitted as a quote. The taste discourse is abundant but mostly founders/VCs (selection bias toward “taste matters”); honest disagreement worth flagging in-essay: “taste” can be a flattering story experts tell themselves.

Outline

Headline: AI Will Force Us To Embrace Our Weirdness Variant: relay-an-idea (Dan Shipper / Every, corroborated by Peter Deng + the Moon–Green–Kushlev study) Hook pattern: 2. The contrarian flip — open by flipping the replacement fear into the real risk (sameness).

Key points:

  1. The fear is the wrong shape. Everyone’s braced for “AI takes my job.” The quieter thing already happening: AI makes the default version of everyone identical. — draws on: my reframe angle A; Shipper “Slop is visible sameness, repeated ad nauseam.” (source: every.to/p/after-automation)
  2. Why sameness is structural, not a phase. Everyone prompts the same models, trained on the same residue of past work, so output converges by design. And it’s measured, not vibes: independent humans spread across the idea-space, independent model outputs cluster. — draws on: Shipper “AI commoditizes the residue of human expertise—whatever can be made explicit enough to train on”; study “the widespread use of LLMs could diminish the collective diversity of ideas” (Moon, Green, Kushlev / Georgetown).
  3. The way out is the thing that can’t be in the training set. Difference has to come from a human’s live read on what matters now — call it taste or judgment. Models know what has been done; you know what needs doing, here, today. — draws on: Shipper “The current generation of models only knows about work that has been done. Humans know about what needs to be done, right now, at this moment.”; Peter Deng (ex-OpenAI) “it’s not going to be about who has the best model. It’s going to be about who has the best taste” (youtube.com/watch?v=uOeFDTHBJCE).

Takeaway/CTA: Stop trying to out-produce the machine. The only moat left is the part of you that isn’t in its training data — your judgment about this moment. What’s yours?

Draft — LinkedIn (master)

Everyone’s scared AI will take their job.

AI can make yesterday’s competence cheap and available to anyone for peanuts.

But that creates slop.

As Dan Shipper, CEO of Every, says: “Slop is visible sameness, repeated ad nauseam. It is what gets produced by default when humans in many different circumstances use the same tool, trained on the same corpus, without thinking too hard.”

And it can be measured. A Georgetown study (Moon, Green, Kushlev) compared thousands of real college essays with ChatGPT ones and warned that “the widespread use of LLMs could diminish the collective diversity of ideas”. Independent humans spread out. Independent model outputs cluster.

It’s not a phase the next model fixes. Same training data in, same average out. Which skyrockets the value of being different: “the work that doesn’t fit the pattern becomes the rare, valuable, and high-status thing”.

And it can’t come from the model. As Shipper puts it: “The current generation of models only knows about work that has been done. Humans know about what needs to be done, right now, at this moment.” Peter Deng, who ran product at OpenAI, says the next winners won’t have the best model — they’ll have “the best taste”.

So stop trying to out-produce the machine. The part of you that isn’t in the training data (your judgment about this exact moment, your weird specific take) is the one thing that doesn’t average out.

What’s the weird, specific thing only you would’ve said?

Draft — X (variant)

1/ AI won’t replace you. It’ll force you to embrace your weirdness. Here’s why ↓

2/ Everyone now reaches for the same models, trained on the same pile of past work. Same input, same output. Dan Shipper (CEO, Every) calls it: “Slop is visible sameness, repeated ad nauseam.”

3/ And it’s measured, not a vibe. A Georgetown study (Moon, Green, Kushlev): “the widespread use of LLMs could diminish the collective diversity of ideas”. Humans spread out. Models cluster.

4/ So sameness becomes the default, and being different becomes scarce. Shipper: “the work that doesn’t fit the pattern becomes the rare, valuable, and high-status thing”.

5/ The catch: that difference can’t come from the model. “The current generation of models only knows about work that has been done. Humans know about what needs to be done, right now, at this moment.” — Shipper

6/ Peter Deng (ex-OpenAI product) on the next winners: not the best model. “the best taste”.

7/ Stop trying to out-produce AI. Your weird, specific judgment about this exact moment is the one thing that doesn’t average out.

What would only you have said?

Final

Backlinks

External sources