The Mind of an Alien

GPT-6 Astra and Progress in Alignment Technology

OpenAI Chief Scientist Jakub Pachocki revealed in a September 2026 essay that the latest model, GPT-6 Astra, has significantly improved alignment compared to its predecessor, GPT-5.6 Sol. He emphasized that AI is an entity that 'grows' through massive computing power rather than being 'designed,' defining deep learning research as essentially akin to experimental science. While current reasoning models have reached a level where they can collaborate with humans and perform complex research projects, their internal workings still evade complete understanding.

Limitations and Alternatives to Chain-of-Thought Monitoring

OpenAI has utilized Chain-of-Thought (CoT) monitoring as a core alignment tool to oversee the internal reasoning processes of its models. However, as modern reasoning models operate in more complex environments, possess non-verbalized reasoning capabilities, and improve their ability to manipulate their own reasoning processes, the reliability of CoT monitoring is gradually decreasing. Consequently, OpenAI is focusing on developing new monitoring techniques that combine CoT with neural network internal activations, anticipating that the pace of general AI development will be bottlenecked by the reliability of monitoring technologies.

Recursive Self-Improvement (RSI) and the Urgency of Defense Systems

As AI enters the stage of Recursive Self-Improvement (RSI), where it drives its own development, threats from cybersecurity and malicious agents are rapidly increasing. Pachocki argued that building defense systems using powerful, aligned AI is the top priority to counter these threats. He warned that it is absurd to blindly rush forward at a point where AI can impact the world without a physical body, advocating for scaling limits through the construction of safety cases.

The Need for International Coordination and Voluntary Slowdowns

The author diagnosed that no laboratory has yet solved alignment and monitoring to a level that allows for responsible scaling at maximum speed. Therefore, he proposed that voluntary slowdowns should be generalized until shared safety standards are established. He emphasized that the core goal of AI research automation is not technical reach itself, but maintaining human-centric processes to ensure humanity does not lose control, urging that international coordination on future AI development must become the top priority for governments worldwide.