Changelog

Edition 1.2.2 — released 1 August 2026. - New in Chapter 4 (§4.2.4): Boris Cherny, who built Claude Code, on having stopped prompting it at all and writing the loops that prompt it instead. - Traced that quotation to its primary source. It had been circulating in six different wordings, none from Cherny himself, and the version most often repeated turns out to be words he never said.

Edition 1.2.1 — released 1 August 2026. - Rewritten throughout in the author’s own voice, following an analysis of four of her own pieces: no em dashes, shorter and unevenly cut paragraphs, and section endings that turn rather than recap. No arguments or sources were changed. - New in Chapter 5 (§5.2.3): the second Trump AI executive order, EO 14409 of June 2026, which shifts the federal focus from AI safety to AI security, and the national-security memorandum three days later that promises civil-liberties protection but makes it unenforceable. - Corrected: a specification length, two misread research findings in §6.5, the number of fabricated cases in the Mata judgment, a Rog-O-Matic comparison, a misquoted Koch passage, two welded statistics, and a claim that Anthropic had downgraded binding safety commitments.

Edition 1.2 — released 24 July 2026. - New in Chapter 1 (§1.5): why AI capability is so uneven — recent leaps cluster where success can be checked by machine, and where the commercial value sits. - New in Chapter 2: Karpathy’s own account of the LLM wiki, including the scale at which it works (§2.5), and Boris Cherny on running agents unattended all day (§2.6). - New in Chapter 3 (§3.4): Karpathy’s original “vibe coding” post, quoted in full — the caveats he attached to it were there from the start and got lost later. - New in Chapter 4: a practitioner’s four kinds of agent loop, sorted by how much you hand over (§4.2.4), and a five-step map of AI adoption where the limit at every step is human (§4.3.2). - Clarified in Chapter 1: reinforcement learning is now explained where it first appears, and distinguished from learning that is tuned on human preference.

Edition 1.1 — released 24 July 2026. - New in Chapter 5: the July 2026 OpenAI–Hugging Face model-evaluation cyberattack (§5.1.1, §5.5.1), plus measured evidence that reinforcement learning grows a model’s reward-seeking. - New in Chapter 6 (§6.5): a fuller treatment of AI’s uneven benefits — who it lifts and who it leaves behind, Cory Doctorow’s reverse-centaur, and the risk of a displaced group turning against AI. - Updated Chapter 5 (§5.2): Australia’s National AI Plan and Office of AI, and the rise of Chinese open-weight frontier models with the risk of a US ban. - New across Chapters 4–6: current July 2026 research on independence of mind, where developers draw the line on AI autonomy, calibrated trust, the global spread of AI exposure, the erosion of the path from junior to senior, and what makes AI help or harm learning. - New in Chapter 4 (§4.2.1): Lilian Weng on harness engineering as an emerging practitioner view. - Site: improved search-engine and social-share metadata.

Edition 1.0 — released 8 July 2026.