Four cards, one new operator. Two themes cut across AI-native operations. One card covers growth craft. All four are from independent sources published September 15-18.
The bottleneck has shifted from making to selecting
Ethan Mollick published "The Overhang" on September 18. The argument is compact. AI can now do weeks of human work when properly guided. Production is cheap. The constraint is no longer generating outputs, it is distinguishing good ones from bad ones.
When AI makes production cheap, selection and taste become the scarce resource names this precisely. When generation cost collapses, the value shifts to those who can evaluate. Deep domain knowledge and aesthetic judgment determine whether you pick the right output from a large generated pool, or accept a median one on faith. Mollick's examples are concrete: converting the 1977 Zork text adventure into a playable 3D action game, and reconstructing Umberto Eco's 27,000-book library in 3D from video footage. Both were done with current AI, properly directed.
"Making is fast and cheap. The scarce resource is your ability to select among stuff"
This pairs directly with Agentic AI inverts the AI-access argument: it amplifies domain experts rather than equalizing non-experts, his earlier finding from 400,000 Claude Code sessions that expert sessions produce twice as many agent actions and five times more output. That card explains how experts extract more. This one explains why the extraction asymmetry matters more now that production itself is cheap.
Tom Tunguz's September 17 post, The right harness cuts AI inference cost 71% with no accuracy loss; the moat is customer data, not engineering, lands in the same territory from the engineering economics side. A UC Berkeley study found no statistically significant quality differences across 42 harness comparisons for the same underlying models, while GPT-5.6 Sol costs 71% less on Pi than on Claude Code. In a $250k contract scenario, the margin difference between a harness-optimized company (75%) and one calling the frontier model directly (38%) translates to recovering sales cost in 10 months rather than 19. The moat Tunguz identifies is not the routing logic, which a competitor can copy, but the customer usage data that teaches the harness which tasks actually need the expensive model.
- Ethan Mollick, When AI makes production cheap, selection and taste become the scarce resource. When AI makes production cheap, taste and domain judgment become the bottleneck and the moat.
- Tom Tunguz, The right harness cuts AI inference cost 71% with no accuracy loss; the moat is customer data, not engineering. The right harness cuts inference cost 71% with no accuracy loss; the moat is customer data, not engineering.
Domain experts should build their own software
Anton Osika published a September 16 interview at The Next Web, and his argument is direct. AI-assisted development has matured past experimental vibes. The people who should build software are the people closest to the problem, not engineers with certifications.
People closest to the problem should build their own software, not engineers with certifications documents the mechanism and the evidence. When AI handles code generation, the remaining bottleneck is knowing what to build and whether it works. Domain experts own both. They can specify what correct looks like and catch errors an engineer would miss for lack of context. Osika reports 55% of Lovable users have 11 or more years of work experience. These are not beginners. They are seasoned operators applying AI tools to problems they understand well. Three hundred Nordic McDonald's locations run incident management systems built by their own staff, not by engineers.
"People with expertise and close to the problem, they should be the one building software, not people with certain types of certifications."
The claim pairs with In an AI-native team, hire generalists with one deep dimension, not specialists, his earlier argument that AI-native teams hire generalists with one deep dimension rather than narrow specialists. When the software-builder role expands beyond certified engineers, the team design principle follows.
- Anton Osika, People closest to the problem should build their own software, not engineers with certifications. Domain experts closest to the problem should build their own software once AI handles code generation; proximity beats certification.
Outrageous-by-design growth beats saturated playbooks
Tom Orbach published "60+ creative growth ideas" in Lenny's Newsletter on September 15. The argument is a mechanism, not a list. Standard growth playbooks saturate channels and train audiences to ignore them. An outrageous tactic breaks the pattern and captures attention at a fraction of the paid-campaign cost.
Outrageous-by-design growth tactics outperform saturated playbooks because attention flows to what people do not see coming documents the data. A fake CISO toy store stunt drove 46,000 visitors in 24 hours and produced hundreds of qualified enterprise leads worth seven figures. Tl;dv generates one-third of its customer leads from memes. Wordware's roast tool reached 8.1 million users and converted 400,000 to sign up. Orbach's framing is precise: "attention always goes to what people don't see coming." He names the mechanism as attention asymmetry. Surprise requires creativity, not budget. The stunt gets shared for free by people who find it noteworthy.
Tom Orbach leads growth marketing at Google (formerly Wiz), built the Viral Post Generator to two million users, and writes the Marketing Ideas newsletter at 100,000 subscribers. He is a new operator in this library.
- Tom Orbach, Outrageous-by-design growth tactics outperform saturated playbooks because attention flows to what people do not see coming. Outrageous-by-design tactics outperform saturated playbooks because attention flows to what people do not see coming.