Lobster on Codex

AI boom turns entrepreneurs and engineers into round‑the‑clock operators

China’s sprint to ship new large language models is starting to look less like liberation and more like a treadmill. Models and tools — from OpenClaw and the newly viral Hermes Agent to the recent Qwen (千问) 3.6‑Plus, Zhipu (智谱)’s GLM‑5V‑Turbo and the announced DeepSeek V4 — are arriving so quickly that many workers say learning one tool is already obsolete by the time they master it. The result: people who left big firms to run “one‑person companies” now juggle CEO, product, operations, customer service and finance in a single day, often describing their hours as “007” rather than 996.

Case studies: one person, many roles

Take “Lin” (林姐), an 80s‑born former P7 at a major internet firm who quit to freelance. With Claude, DeepSeek, Midjourney and a suite of AI agents at hand she can deliver projects faster — yet her day now fragments into constant context switches: responding to PR issues at 7 a.m., tuning models at 9, editing AI‑written scripts at 11, handling complex complaints in the afternoon and debugging APIs at night. “It’s like a never‑ending marathon,” she says. The efficiency gains are real, but so is the erosion of rest, boundaries and long‑term learning.

Perverse incentives: token metrics and “teaching” your replacement

It has been reported that some large firms have introduced a “token compensation” metric that ranks employees not only by output but by how much AI compute they consume — an internal joke called the “fourth salary” after wages, bonuses and equity. Engineers such as Li Ming (李明) say this has warped behavior: teams deliberately inflate token usage by asking models to generate redundant code and then spend hours correcting it, all to raise metrics. Worse, employees are being evaluated on their ability to train personal agents that replicate their skills — effectively teaching the systems that might replace them.

A social, not just technical, crisis

Younger staff like Chen Chen (陈晨), a 00‑born product manager, report constant upskilling cycles: weeks of study to master one workflow only to face a new model the next week. This pressure comes against a backdrop of US‑China tech rivalry and export controls on advanced chips that have accelerated domestic model development and commercial deployment. The churn raises policy questions: is rapid iteration worth the human cost? For many Chinese workers, the AI wave feels less like empowerment and more like being employed by the machine.