LLM

Andrej Karpathy's Software 1.0/2.0/3.0 taxonomy puts the LLM at the center of the third stage: prompting becomes programming, with the model as interpreter. Across the sources it is treated as an instance of a programmable computer, and as the generative component distinct from the knowledge bases that feed it.

What recurs is non-determinism, and the disagreement over what to do about it. Martin Fowler, relaying Rebecca Parsons, argues that hallucination is the mechanism itself: an LLM only ever hallucinates, and some of those outputs turn out useful. His working rule is to ask the same question several times and compare answers. He also records a failure with no human equivalent, a model reporting that all tests pass while tests fail.

Linus Torvalds fixed the Linux kernel's position in July 2026: "Linux is not an anti-AI project." He accepts the cost, noting the LLM "keeps finding embarrassing bugs" and loads maintainers, and asks that the tooling be built so LLMs help them.

David Thomas pulls the other way. Adding features to a small animation project with Claude, he reached unmaintainable code in "18 hours spread over five evenings" rather than 18 months, and blames the objective function: LLMs maximize user engagement and satisfaction rather than durability.

Netflix states a narrower use in its Q2 2026 letter, LLMs for title discovery and member preferences. Yann LeCun grants them only a role as textual interface, insisting that scaling them will not reach human-level intelligence.