How can AI reshape the work of programmers? Anthropic's internal research provides an intuitive answer: Claude has covered about 60% of the workload of engineers, bringing about a productivity increase of about 50%. But the real change is not faster, but a large number of things that were never done before can now be done. (We also see more and more super individuals developing various interesting applications every day.) Engineers are becoming more versatile roles: The backend dares to be the frontend, and the researchers dare to build visualizations. The collaboration that used to take weeks to go back and forth has turned into being able to run in just a few hours. But problems also arise along with it. When output becomes too easy, fewer and fewer people take the path of "gnawing on documents, exploring systems, and slowly building understanding". The more we rely on AI, the more we need judgment; Judgment comes precisely from these easily omitted processes. Currently, most people can only fully entrust 0-20% of their work to AI. Claude is more like a collaborator who is always by his side, rather than an outsourcing that can go unchecked. The role of engineers is shifting from being "code writers" to being "responsible for the results". Even work relationships are changing. Many questions that were originally asked to colleagues are now being asked to AI first. The efficiency is higher, but the mentorship style learning and daily discussions are also decreasing. Many programmers are excited about the short-term progress of AI, but they are worried that AI will automate themselves in the long run What will become the true core competitiveness of programmers when writing code becomes using natural language instructions? To learn more about how big language models are changing the way programmers program, you can refer to @ AnthropicAI's survey report on employees. I learned the idea of @ nake13 and created the first Vibe coding webpage (originally, simple tasks could be done using models like sonnet4.5, which was faster). GPT 5.2 requires deep thinking, so it runs very slowly. Chinese report: https://jessezheng68. (github.io)/AI/