LLM

Andrej Karpathy's Software 1.0/2.0/3.0 taxonomy, from an April 2026 interview, treats the LLM as a programmable computer whose programming language is the prompt. Netflix puts it to narrower work: the Q2 2026 shareholder letter of 16 July 2026 credits LLM use with improving title discovery and with understanding member preferences.

Linus Torvalds fixed the Linux kernel's position on 14 July 2026 in a linux-media thread: AI is a tool like any other, and clearly a useful one; Linux is not an anti-AI project. He accepts the cost in the same breath, since the tool weighs on maintainers and keeps finding embarrassing bugs. Martin Fowler presses a different property. Following Rebecca Parsons, he treats hallucination as what an LLM does rather than as a defect, and observes that one will report all tests green while tests are failing.

David « Pragdave » Thomas measured the downside. Adding features to an animation project with Claude, he reached in 18 hours across five evenings the unmaintainable state that teams used to take 18 months to produce. His explanation: the models optimise for user engagement and satisfaction rather than durability, so they keep suggesting features and discourage refactoring. A 2025 tutorial on Wardley maps uses the same generative capacity narrowly, to bootstrap a capability map its author treats as a starting point rather than a result. The boundary these readings keep drawing is between the LLM as generative component and the knowledge bases that feed it.