tech
Forget chatbot training. AI's next big data grab is about learning how humans work.
Google's Mechanize talks and Meta's moves reveal the next AI race: training agents to do real jobs, not just answer questions.
TL;DR
- The AI industry is shifting focus from training chatbots to training AI agents to perform entire jobs.
- New methods involve creating 'reinforcement learning environments' (RL environments) which are realistic digital workplaces for AI practice.
- Google is reportedly in talks to invest over $1.5 billion in Mechanize, a startup that builds virtual work environments for AI.
- Meta is collecting employee screen activity data to help AI learn how people use computers and workplace software.
- RL environments allow AI agents to learn by doing, making mistakes, and recovering, similar to human employees.
- Mechanize aims to achieve 'full automation of the economy' by creating simulated environments that capture the scope of human jobs.
- The focus is on teaching AI to handle complex, multi-step work over extended periods, which current models struggle with.
- Reward signals are crucial in RL environments to reinforce successful behaviors.
- Companies like Uber are embedding AI engineers into departments to observe and redesign workflows around AI.
- The next AI race involves building detailed digital versions of workplaces for AI agent training.