AI must stop answering and start finishing tasks, cites...

A new survey paper argues that AI won’t earn the label “coworker” until it stops answering questions and starts delivering finished work. Researchers from Tencent’s Youtu Lab, together with several Chinese universities, map the transition from “chatbot” to “digital colleague” along two axes: the cognitive core and tool‑assisted task execution. Their central claim flips the usual benchmark—rather than chasing better answers, the focus shifts to reliably turning intent into completed tasks within persistent work environments.

The authors trace the evolution from the early “fast answer” era, where models stored language patterns in parameters and generated text token by token without checking intermediate steps, to the emerging “thinking‑LLM” era. OpenAI’s o1 and Deepseek‑R1 exemplify this shift, pouring extra compute into answer generation, producing long chains of thought, and rewarding only verifiably correct solutions. By borrowing Daniel Kahneman’s System 1 versus System 2 framework, the paper frames the change as a move from intuitive, rapid responses to deliberate, self‑correcting reasoning. First‑generation agents could call APIs or browse the web, yet they remained fragile; the next step, the authors suggest, lies in reusable “skills” that let AI operate as a true digital teammate.