Meta Tracks Employee Inputs for AI Training

🔑 Enhanced Key Takeaways

• •The initiative is part of Meta's 'Project Ghost' internal effort, which aims to automate repetitive administrative workflows like scheduling, data entry, and email triage to improve internal productivity.
• •Meta has implemented strict data anonymization protocols, claiming that keystroke logging is filtered to exclude passwords, personal communications, and sensitive financial information before being ingested into the training pipeline.
• •The program has faced significant pushback from internal privacy advocacy groups and labor unions, leading to the establishment of an 'opt-out' mechanism for employees in specific non-sensitive roles.

🛠️ Technical Deep Dive

• •The system utilizes a proprietary 'Action-Transformer' architecture designed to map raw input sequences (mouse coordinates, click events, key codes) to high-level task completion goals.
• •Data collection is handled by a lightweight kernel-level driver that captures event streams at 60Hz, which are then compressed and serialized into a custom format optimized for Large Action Model (LAM) training.
• •The training pipeline employs Reinforcement Learning from Human Feedback (RLHF) where human supervisors verify the 'intent' behind the recorded sequences to refine the agent's decision-making policy.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline