The world is too loud. Read what matters.

AI & I

AI makes generating answers cheap; what's scarce is your perspective

When everyone gets the same default answer from Claude, what's truly valuable isn't the model's power—it's the different answer you extract by questioning with your responsibility and perspective.

AI & automationFuture of workAllocation economyClaudeTeam collaborationAI evaluation

The video won't play here. Listen to the audio instead:

This episode skips abstract automation threats. It uses one week of real work to examine which jobs AI actually replaces—and the argument is airtight.

The argument · tap a timestamp to hear it

4:15

In three years, AI went from losing to experts to crushing them in evaluation

OpenAI's GDPval benchmark gives models real professional tasks to complete, then blind expert evaluators judge human vs. AI versions. In 2024 with GPT-4o, AI won 12.4% of the time; by 2025 that rose to 47.6%; by 2026 with newer models it reached about 83%. Dan uses these numbers to show that the curve looks dramatic when you aggregate results—but he'll argue the curve is misleading.

— Dan Shipper
6:18

One Claude subscription replaced five professional roles in a single week

Dan walks through his actual work last week: analyzing quarterly customer research, building a team guessing game, vibe coding his own CRM and using it to contact customers. By 2020 standards, that would require a data scientist, game developer, product marketer, engineer, and BDR—roughly $200,000 in labor. He did it all in fragmented time with one Claude subscription. He calls the jolt of watching a model do your work an "agency rupture".

— Dan Shipper
11:28

The same data, seen through different lenses, tells completely different stories

One customer research report: Dan saw an opportunity to pitch Every to the team; Yash, leading growth, spotted high churn after month one for new users; Douglas, marketing lead, noticed readers felt lonely; editor Jack turned it into a family Q&A game show. Default Claude? Generic conclusions anyone can get: "Writing is the business, bundling drives retention, concentrated risk and too much email".

— Dan Shipper
14:46

Generating answers is cheap; finding the right one in thousands is hard

Mike Taylor, the evaluation lead, ran an experiment: can Claude predict what Dan will say on Slack? If given 400 candidate answers, Dan's actual message appeared 93% of the time. But ask Claude to rank those candidates, and Dan's actual message lands in the top 10% only about 60% of the time. The bottleneck isn't generating content—it's ranking through enormous possibility space to find what actually matters.

— Dan Shipper
18:01

Comparing people to AI misses the point; compare how people deploy it

Someone asks: once Claude's ranking improves, won't 2027's Claude make Yash obsolete? Dan says the comparison itself is wrong. By 2027, Yash becomes "a 2027 Yash using 2027's Claude", so it's really "2027 Yash + Claude" vs. "default prompt + 2027 Claude". Claude cancels out. What you're comparing is "someone using the model with responsibility and perspective" against "someone using a default prompt"—and he picks the first every time.

— Dan Shipper
21:10

Model power sets a rising floor; your perspective is what increments beyond it

Model defaults get stronger with each update, but since everyone has access, they just become a rising floor—constantly pushed higher, distributed everywhere, no longer valuable in themselves. Real alpha is what you extract by pushing the model through your particular lens and responsibility. Even if tomorrow absorbs what you found into a higher floor, you're still upgrading alongside new models instead of getting frozen with old ones.

— Dan Shipper
23:13

Stop evaluating AI on standard answers; grade it on perspective-driven work instead

In a 2024 article, Dan proposed the allocation economy: the future judges you not on what you know but on how well you deploy and manage AI as a resource to finish work. He extends this to a counterintuitive point: we benchmark AI using problems with agreed answers—math, code—but those are cheap precisely because they're verifiable and sharable. What we should test is whether the model can execute "work from my perspective"—tasks with no consensus answer.

— Dan Shipper

In their own words · checked verbatim

This is in 2020 terms like a data scientist, a game developer, a product marketer, an engineer, a BDR, and like 200, maybe 200K.

Dan Shipper6:18

And you think about that, you really think about that, you look at a table like this and you are like, we are so fucked.

Dan Shipper6:18

Generation is really cheap, but ranking is hard.

Dan Shipper15:55

I pick Yash every day. I pick Yash every day, every time.

Dan Shipper20:06

the floor keeps moving, but the floor keeps getting distributed to everybody. And so what's valuable, the alpha, is where you push the model using your perspective and your responsibility for a particular area of the world.

Dan Shipper21:10

you won't be judged on how much you know, but instead on how well you can allocate and manage the resources to get work done, the resources of intelligence.

Dan Shipper23:13

Figures

AI win rate in GDPval (2024)12.4%4:15
AI win rate in GDPval (2025)47.6%4:15
AI win rate in GDPval (2026)about 83%4:15
Every team size309:25
Claude accuracy predicting Dan's Slack message (400 candidates)93%14:46
Probability of top-10% ranking after sorting candidatesabout 60%14:46

Glossary

GDPval
OpenAI benchmark measuring model performance on real professional tasks through blind expert evaluation versus human work
agency rupture
The disorienting shock of seeing an AI model replicate your own work
allocation economy
An economy where value comes from how you deploy AI resources, not from your own knowledge

How to listen

Who it's for

Founders and product leaders wanting concrete answers on what people should do after automation, plus knowledge workers figuring out how to position themselves in the AI era.

Skip

The conference background in the opening (0:00-1:51) is skippable without breaking the core argument.