What got cheap, what got scarce

The principle behind the list#

AI doesn't know everything. It knows what has already been written down, and it produces something close to the average of it. That one fact explains the whole list: whatever is written, public and average got cheap. Whatever is unwritten, specific, verified or trusted got scarce.

The 10 shifts#

1. Writing → checking#

A model writes 2,000 lines in minutes, but humans still read at the same speed, and bugs are subtler because the code looks right. Every generated line is a line someone must maintain and secure. In math, proof search got cheap (DeepMind's AlphaEvolve found new bounds this way), so the bottleneck moves to verification, and formal proof tools like Lean become core infrastructure. Adapt: be the person who can say "this is correct." Invest in tests, specs, formal checks and fast reading.

2. Building → choosing and distribution#

The old creed was "ideas are cheap, execution is everything." Now execution is the cheap part. Anything you can describe in a prompt can be cloned in a weekend. If your moat was "it was hard to build", you no longer have one. Adapt: build moats from what can't be put in a prompt: your customer list, your data, community, brand, relationships.

3. Public knowledge → unwritten knowledge#

Any public field can now be learned in weeks with an AI tutor. In Oct 2025, GPT-5 "solved" several Erdős problems. In fact it had found existing solutions in papers nobody had linked, so literature search got cheap overnight. But AI doesn't know what your customers complained about yesterday, or what gets said at dinner and never posted. Adapt: gather private knowledge on purpose. Note: written English is now cheap. Spoken English, for calls and relationships, is still expensive.

4. Content → clicks#

An article used to take hours, and Google sent free traffic for it. Now articles cost nothing and search answers the question itself. Pew (2025): users clicked a result on 8% of searches showing an AI summary, versus 15% without. They clicked the summary's own sources about 1% of the time. For anyone whose customers run on publish → search traffic → ads, the risk isn't rivals. It's fewer people starting content sites at all. Adapt: follow where value moves: newsletters, memberships, paywalls, communities, being cited by AI (GEO). Sell tools for owning an audience, not layouts for ad traffic.

5. Average → taste#

AI outputs the middle of what it learned from, so AI-made things start to look alike: same fonts, same gradients, same card layouts. "Pretty good" is no longer a job or a product. Covering every niche with dozens of demos (the old ThemeForest strategy) is now something AI does on demand. Adapt: be more opinionated, not less. A clear point of view beats coverage.

6. Buy → build#

Small tools, plugins and sites were bought because building cost more. Now customers ask AI for exactly what they need. For micro-SaaS and single-feature plugins, the real competitor is the customer plus an AI. Adapt: sell what people don't want to do themselves: updates, security, compatibility, hosting, support, someone responsible when it breaks. Sell the maintenance and the outcome, not the file.

7. Lock-in disappears#

Migrating code, changing stacks or moving a site used to be expensive, and that friction kept customers in place. AI migrates code and content cheaply now. Part of every platform's moat was switching cost. Adapt: keep customers because they want to stay, not because leaving hurts. And use cheap migration yourself: build importers that bring customers over from competitors.

8. Build time → feedback time#

Building takes a day, but people still take weeks to find a product, try it, pay and leave. That clock didn't speed up. Your build speed no longer limits how many experiments you can run; the number of real people you can reach does. Adapt: your own audience is your lab.

9. Headcount → your attention#

One person with agents can ship what took 10–20 people. The limit is how much you can direct, review and decide in a day. The ceiling for a one-person company moved way up, and thousands of other solo builders got the same leverage at the same time. Adapt: invest in what multiplies judgment: specs, reusable skills and agents, automated checks. This is also why the sleep and rest floors from Chapter 4 matter more. Decision energy is the bottleneck.

10. Looking credible → being trusted#

A polished site, good writing and a degree used to signal competence because they were costly. Now anyone can fake all of them in an hour, faces and voices included. A signal that costs nothing tells people nothing. Adapt: build signals that are expensive to fake: a real name and face over years, a public track record, in-person relationships, guarantees, a human who answers. Reputation is 10-year work, so start now.

By audience#

The next 10 years#

• Confident: generating things keeps getting cheaper. Trust, distribution, the physical world and in-person relationships stay expensive longer.
• Uncertain: whether checking gets automated too (probably first in math, through formal proofs). And whether AI assistants become the main shop, with agents picking the tools for users. If so, being the product the agent picks is the new SEO.

What this does to the success tree#

"Build product" and "landing page" lose weight. "Talk to customers", distribution and trust gain it. And several items from the cut list in Chapter 4 turn out to be exactly what's scarce: meeting people is shifts #3 and #10; painting is #5 (taste); travel is #3. The AI era raised the value of the zone you were about to cut.

Sources: Pew Research on AI summaries and clicks