AI coding tactics, the death of SaaS, and the rise of in-person business

Executive overview

AI is making software cheaper and easier to build, eroding the competitive moat that SaaS companies once held. When any developer can ship an app quickly, network effects, proprietary data, and physical presence become the only durable advantages.

The video covers three areas: practical tactics for managing large logs in AI coding tools, a thesis on which business types will thrive as AI commoditises software, and a reading of the vibe coding debate through the lens of the Gartner hype cycle.

Physical businesses that adopt AI on the backend will outperform pure software companies within two to three years.

Managing large logs in Cursor

• Start a fresh conversation and ask the old one to summarise errors and nuances (no code) into a system prompt for a "new AI engineer stepping in."
• Pass that system prompt to a fresh session running a reasoning model (e.g. Claude 3.7 Thinking).
• Ask for root cause analysis only — no fixes, no code changes — before attempting a solution.
• Repeat root cause analysis one to three times; it usually surfaces the systematic connection between errors.
• For a TDD reinforcement loop: have the AI write a unit test first, then the function, then run the test against mock data; feed failures back in to rewrite the function, not the test.
• Add an integration test alongside the unit test to validate the full end-to-end flow.
• For massive log volumes, flatten the codebase to a single file and run it through Gemini's 2M-token context window for a one-shot fix.

The future of SaaS and sustainable competitive advantage

• SaaS multiples were high because software had low operating costs; that advantage shrinks as build costs collapse.
• AI lowers the barrier to entry so dramatically that software alone will have no durable moat within two to three years.
• Three remaining advantages: network effects (e.g. social platforms), private data (weakening as synthetic data improves), and in-person services.
• Physical services — HVAC, healthcare, hospitality, skilled trades — cannot be replicated digitally and will hold value longer.
• Businesses that pair in-person delivery with AI-run back-office operations (admin, marketing, branding) will have the strongest position.
• In-person social experiences (run clubs, silent reading clubs, flea markets) are already growing in dense cities and spreading globally.
• As AI saturates the internet with content, demand for verified human interaction and physical presence will increase.

Vibe coding and the hype cycle

• Engineers cycle through: fear → ridicule → resistance → acceptance — currently between fear and ridicule for vibe coding.
• Critics cite insecure code and demo-only limitations; most objections reflect skill gaps rather than fundamental limits.
• Vibe coding is a spectrum: heavy (no code review) to light (AI assists, human edits); neither extreme is universally right today.
• As models improve, the viable position on the spectrum shifts toward heavier reliance on AI.
• Coding abstraction has always progressed: assembly → C → Python → English; vibe coding is the next step, not an aberration.
• Expect trash-talk and pushback to intensify before acceptance sets in, mirroring past technology adoption cycles.