OpenAI chief scientist Jakub Pachocki published an essay, An Alien Mind, on Sept. 6, 2026, reflecting on the recent progress of reasoning models and warning about the consequences of a sustained rapid rise in machine intelligence.
From RLSlow to scalable reasoning models
In the essay, Pachocki looked back to the RLSlow research project in mid-2023, which he said produced the first results showing that reasoning models could be trained in a scalable way. Three years later, he wrote, reasoning language models had become part of rapid economic growth and were beginning to push the boundaries of science.
He said those systems can operate computers and graphical interfaces, work with humans and other AI systems, and carry out research projects. At the same time, he said, they are reshaping computer security and introducing new dangers.
Larger capability jumps may still lie ahead
Based on internal results, Pachocki said he expects the current pace of progress may continue through recursive self-improvement. If AI continues along its current path, systems could see capability jumps of a similar or greater scale over the next few years and increasingly contribute to their own development, he wrote.
He also called for extreme caution, arguing that no one is prepared for the consequences of machine intelligence continuing to rise rapidly.
OpenAI’s stated response
Pachocki wrote that OpenAI will continue seeking technical solutions for alignment and monitoring, build defensive systems, and, if necessary, unilaterally stop further scaling. Even so, he said broader intervention will be required.
The essay also says machine intelligence is driven mainly by increases in compute, and that AI is growing more than it is being designed.
Alignment, monitoring and defense
On alignment, Pachocki distinguished between goal alignment and value alignment, and described generalization as the central challenge. He added that GPT-6 Astra shows materially stronger alignment than GPT-5.6 Sol, though more progress is still needed.
On monitoring, he said the main bet remains chain-of-thought monitoring, while evaluations show the capability relied on there is gradually weakening.
On scalable defense, the essay says models are becoming superhuman at breaking into computer systems. Pachocki wrote that the industry is now in a narrow window in which the best available models can still be used to significantly strengthen the security of critical systems.