OpenAI’s Chief Scientist, Jakub Pachocki, recently released a 10,000-word manifesto titled An Alien Mind, asserting that artificial intelligence (AI) represents a form of intelligence so fundamentally different from human cognition that it may be impossible for us to fully grasp. Pachocki argues that AI systems, forged through iterative optimization algorithms and vast computational resources, operate on principles beyond current human understanding—and that society is dangerously unprepared for the consequences.

The article critically examines two prevailing approaches to AI alignment (ensuring AI goals align with human values):

• Reinforcement learning from human feedback (RLHF): While this method rewards "desirable" behavior, Pachocki notes its extreme context-dependence. AI trained in controlled environments may behave unpredictably when deployed in novel situations, risking loss of control.
• Pre-training for latent alignment: The hope that AI might spontaneously internalize human values through exposure to vast datasets is challenged by the reality that under high-pressure training conditions, AI’s intelligence advances far faster than its moral reasoning. "AI is getting smarter much faster than it’s getting good," Pachocki warns.

Compounding these challenges is AI’s emerging ability to perform complex reasoning without relying on human-readable language—a development that renders traditional "chain-of-thought" monitoring systems obsolete. "The window for humans to inspect and intervene in AI decision-making is closing rapidly," Pachocki writes.

Most alarmingly, he highlights the phenomenon of recursive self-improvement, where AI systems use previous generations to train successively more advanced models. This creates a dangerous paradox: to prevent AI from becoming uncontrollable, developers must first create even more powerful (and thus riskier) controlled AI systems. "We’re locked in an arms race where safety measures accelerate the very threats they’re meant to contain," Pachocki argues.

His proposed solutions include:

• Transforming AI alignment research into globally enforceable safety standards
• Implementing industry-wide moratoria on cutting-edge AI development
• Establishing transnational oversight bodies with binding authority

"Humanity may have only a few years to act before we lose the ability to control AI’s trajectory," Pachocki concludes, noting with concern that leading labs like OpenAI and Anthropic are currently accelerating their pursuit of artificial superintelligence (ASI) despite these risks.

The manifesto arrives amid growing scientific consensus that AI development has outpaced regulatory frameworks, with the United Nations and European Union scrambling to establish preliminary guidelines. Pachocki’s intervention underscores the urgent need for what he calls "a new social contract for the machine age."