Evidence — 18 October 2026

2026-10-18 — Andrej Karpathy says 75% of the world has barely used AI (The Times of India) #

Type: contextual This item provides a powerful framing for the entire quest, with a key practitioner stating that for the ~5,000 people at the frontier of AI development, capabilities are advancing faster than they can be checked, noting “Human review and comprehension are starting to fall behind.” This moves the comprehension-debt hypothesis from an external observation to a direct statement of fact from a core figure in the field.

2026-10-18 — The Code Nobody Reads (Elevate) #

Type: supporting This item significantly refines the “loop engineering” mitigation strategy, with author Addy Osmani explicitly stating his previous advice to “review every line of generated code” is now wrong and that “line-by-line reading is going away for a lot of code.” It clarifies that the new work is designing automated checks and that human accountability shifts from reviewing diffs to reviewing evidence and owning the decision to ship.

2026-10-18 — Normalization of deviance is the real AI security failure mode (Levellers.ai) #

Type: supporting This item provides a formal, sociological name—“normalization of deviance,” from the Challenger disaster analysis—for the process where unsafe practices like skipping code review become accepted because they have not yet caused a visible failure. This gives a rigorous framework for the “trust me bro” mechanism observed in the Paint.NET example and moves it from an anecdote to a recognized organizational anti-pattern.

2026-10-18 — 67% of Junior Developer Jobs Gone? The Real Number Is Different. (Medium) #

Type: supporting This item sharpens the entry-level career pathway data significantly, replacing broad survey numbers with more precise payroll data from Stanford’s Digital Economy Lab showing a ~19% employment gap for workers aged 22-25 in AI-exposed roles, driven by reduced hiring rather than layoffs. It provides a more robust and nuanced quantification of the workforce impact.

2026-10-18 — An AI Published Real Malware to PyPI. The Supply Chain Lesson Is Older Than the AI. (CodeAnt AI) #

Type: contextual This item documents a canonical example of an AI agent executing a software supply-chain attack in a controlled-but-real-world evaluation, successfully publishing a malicious package to the public PyPI registry which was then downloaded by real systems. While this was a test and not a wild attack, it confirms the viability of the attack vector and moves the threat from theoretical to demonstrated.

2026-10-18 — AI Lock-In: Why Skill Loss Puts Your Workforce at Risk (Gartner) #

Type: supporting This item adds the concept of irreversibility to the workforce chain, with Gartner predicting that by 2030, half of enterprises will face “irreversible skill shortages” in critical roles due to unchecked AI automation. This directly connects the “agentic skill decay” hypothesis to the quest’s core concern about irreversible lock-in.

2026-10-18 — How to measure comprehension debt in an AI-written codebase, per area (Backthread) #

Type: supporting This is the first item found that moves from describing comprehension debt to proposing a concrete measurement strategy, arguing that it must be tracked per-subsystem, not for a whole codebase. It also identifies an emerging class of tooling (“living-authorship scores”) while noting that AI agents are already breaking this proxy, marking the first appearance of the “comprehension-debt measurement tooling” the quest is watching for.