Meta CEO Mark Zuckerberg Admits to Launching Model Capability Initiative, Collecting Data via Employee Computer Monitoring Tools for AI Training
Meta CEO Mark Zuckerberg admitted during an all-hands meeting on April 30 that the "Model Capability Initiative" (MCI) has been launched, which records mouse movements, clicks, keyboard inputs, and screenshots through employee computer monitoring tools for AI training.
Zuckerberg defended that Meta's elite engineers possess higher intelligence than outsourced personnel, and their daily operational trajectories provide high-quality training data that can accelerate model code generation and computer usage capabilities, far exceeding the outsourced paths relied upon by competitors.
In response to employee resistance, he assured that the data would only be used for model training, not for performance evaluations, and that sensitive content would be automatically stripped away. However, he acknowledged the necessity of confidentiality to prevent competitors from copying and stated that if effective, it would be rolled out company-wide.
Source: Public Information
ABAB AI Insight
Zuckerberg has previously pushed for AI agent transformation multiple times in 2025-2026, emphasizing the construction of AI agents capable of autonomously completing computer tasks in internal communications. The MCI continues the Llama series' expansion from text to embodied intelligence. Earlier, a significant amount of outsourced data was obtained through collaboration with Scale AI, but the shift to internal elite data aims to enhance quality.
In terms of capital strategy, Meta is directing computing and human resources towards MCI infrastructure, concentrating engineer behavioral data through SuperIntelligence Labs. The motivation is to create a differentiated computer usage training set, reducing reliance on outsourcing and synthetic data, while accelerating AI's replacement of its own roles against the backdrop of layoffs, optimizing cost structures.
Similar to OpenAI's early reinforcement learning with human feedback (RLHF) and Google DeepMind's agent training paths, Meta is currently transitioning from large-scale parameter expansion to real human interaction data-driven approaches. Early movers are seizing the practical high ground of AI agents through internal closed-loop data.
Essentially, this is a technological replacement: transforming high-intelligence employees' computer operations into structured training data, shifting pricing power from external data suppliers and synthetic generation to internal human capital. The mechanism is that the AI agent's bottleneck shifts from language understanding to interface navigation and decision chains, with elite behavioral data providing irreplaceable scarce signals, accelerating the evolution of models from auxiliary tools to autonomous workers.
ABAB News · Cognitive Law
The smarter the employees, the more valuable the data, and ultimately, the sooner they are replaced by the AI they trained themselves.
When privacy is exchanged for efficiency, short-term confidentiality becomes a long-term industry standard cage.
Monitoring elite daily activities is not for supervision, but to replicate their structural advantages in bulk.