Tech Companies Shift AI Training to RL Workplaces as Scale Says Nearly Half of New Projects Use Them
Updated
Updated · Business Insider · Aug 12
Tech Companies Shift AI Training to RL Workplaces as Scale Says Nearly Half of New Projects Use Them
3 articles · Updated · Business Insider · Aug 12
Summary
Nearly half of Scale AI’s new training projects now use reinforcement-learning environments, underscoring a broader industry push to teach AI agents full workflows rather than just better answers.
Those simulated workplaces let models code, use enterprise software, make mistakes and recover over hours or days — a response to the limits of internet-text training and human feedback for complex automation.
Google is discussing a more than $1.5 billion investment in Mechanize, whose virtual work environments target software engineering first and are pitched toward automating a share of the $18 trillion U.S. wage economy.
Meta has sought employee keystrokes, clicks and screen activity to show AI how people actually use workplace software, while Uber’s “Agentic Pods” place AI engineers inside functions like finance and HR to map tasks.
The shift suggests the next AI race is moving from chatbots to digital replicas of real offices, where agents can be trained and evaluated on white-collar work.