Meta has discovered AI self-improvement — Zuckerberg will limit public access to powerful models

Meta researchers report observing early signs of AI systems improving themselves without human help. Mark Zuckerberg — founder of Facebook, Instagram, Threads, and WhatsApp — calls this a first step toward Artificial Superintelligence (ASI) — AI able to exceed human cognitive abilities and enhance itself at an exponential pace.

Zuckerberg explained that AI development follows three levels: narrow AI, Artificial General Intelligence (AGI), and ASI. Current models achieve superhuman results only in specific tasks, while AGI would learn and adapt like a human. ASI would go far beyond, potentially triggering an “intelligence explosion.”

He referenced research into Gödel machines — theoretical AI able to rewrite its own code only after formally proving the improvement is beneficial. In 2024, scientists tested an AI agent based on this idea, showing better results in programming, science, math, and reasoning compared to human-designed agents.

Zuckerberg remains optimistic about ASI’s ability to accelerate innovation and empower individuals, but stressed caution. Meta will avoid open-sourcing its most powerful systems to prevent misuse, which is critical for global security.