Dario Amodei Said AI Would Write Almost All Code by Now. So, Did It?

A year ago, Dario Amodei said something that raised eyebrows even by the standards of Silicon Valley confident-speak. In a March 2025 interview, the Anthropic CEO predicted that within twelve months, "almost all code" would be written by AI. This week, that quote came back around.

@slow_developer posted the clip on Tuesday with minimal commentary — just the quote and a question mark. Within hours it had racked up tens of thousands of impressions and a comment section split sharply between "he nailed it" and "not even close."

So what does the evidence actually show?

There is no question AI coding tools have moved from novelty to infrastructure in the past twelve months. Claude Code, GitHub Copilot, and Cursor together claim hundreds of thousands of active developer users. "Vibe coding" — a term Andrej Karpathy popularised last year — has gone from a meme to a legitimate workflow at startups and, increasingly, at larger companies. GitHub has reported that Copilot now accounts for a significant and growing share of code committed across its platform.

But "almost all code" was never just about autocomplete. Amodei appeared to mean something closer to autonomous generation — AI writing complete features from a prompt, with humans reviewing rather than authoring. By that measure, the prediction has not landed. Gary Marcus, the cognitive scientist and longtime AI sceptic, noted this week that the definition of "AI-written code" is doing enormous work in these claims. "If a developer accepts 30 percent of a Copilot suggestion after editing it heavily, did AI write that code? Amodei's framing is slippery enough to be unfalsifiable," he wrote on Substack.

The gap between assisted and autonomous coding is where most of the argument lives. Large enterprise codebases, regulated industries, and safety-critical systems remain largely untouched by autonomous coding. The agents that have made the biggest headlines — Cognition's Devin, GitHub Copilot Workspace — still perform best on discrete, well-scoped tasks, not the sprawling and ambiguous problems that fill most engineers' actual backlogs.

Whether Amodei was directionally right or just optimistic on the timeline, the fact that a year-old interview clip can still generate this kind of heat says something about the stakes. CEO predictions from major AI labs do not just reflect expectations — they shape them, driving competitor roadmaps, startup pitch decks, and enterprise hiring plans. Getting the timeline wrong by twelve or twenty-four months in this market is not a footnote. It is a story.