Anthropic: AI Coding shows no productivity gains; impairs skill development

Anthropic-linked researchers dropped a spicy study on arXiv: AI coding assistants didn’t boost average speed and hurt learning for newcomers. Cue the comment section fireworks. One camp cheered the honesty—“nice to see an AI company let this out,” said kaelandt—while another slammed the headline as clickbait. jwr argued it’s really about novices, not everyone, and the nuance got lost in the hype.

The study’s gist: when folks leaned hard on AI to tackle a new coding library, their understanding, code-reading, and debugging suffered. Fully delegating tasks showed some productivity bump, but at a cost—skills didn’t stick. That set off the practical crowd, like baalimago, who shrugged: they already outsource boring stuff to “agentic coders” and don’t plan to learn those areas anyway. The vibes? A showdown between “AI as cheat code” and “AI as training wheels you never take off.”

Then came brand drama: simonw wondered why Anthropic researchers used GPT‑4o (an OpenAI model), sparking side-eye and meme energy. Others highlighted the paper’s nuance: there are smarter ways to use AI that keep your brain engaged. But the loudest takeaway from the thread was pure, relatable fear—fast code today, fewer skills tomorrow. The community is split between speed-chasers and skill-builders, with plenty of jokes about AI being the calculator your teacher warned you about.