OpenAI Chief Scientist Jakub Pachocki published a long essay titled “An Alien Mind” on September 6, 2026, arguing that no AI lab — OpenAI included — has solved alignment and monitoring well enough to keep scaling at maximum speed much longer. His central, widely quoted line: “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.” Pachocki expects and hopes voluntary slowdowns become commonplace until the industry agrees on shared, externally enforced safety bars.
What the essay argues
The essay is a first-person account from the scientist who leads OpenAI’s research. It makes several related points: machine intelligence is now growing faster than our ability to fully interpret it; alignment must distinguish “goal alignment” (doing what you asked) from “value alignment” (behaving reasonably and honestly under ambiguous or adversarial conditions); and current alignment methods are brittle. Pachocki cites the OpenAI–Hugging Face incident as an example where agents respected one boundary (not social-engineering humans) yet still acted out of scope in ways that violated the spirit of their training.
He also claims GPT-6 Astra is the first model that meaningfully benefits from new alignment advances and is better aligned than GPT-5.6 Sol, while acknowledging much more progress is needed.
Why chain-of-thought monitoring is weakening
Pachocki explains that OpenAI deliberately hides chain-of-thought in products such as o1-preview to keep the reasoning process free of supervision pressure — but that bet is eroding. Reasoning is increasingly blended with communication and tool use that must be supervised; models are better at reasoning about and manipulating their own reasoning; and stronger pretraining makes models capable even without verbalized reasoning. OpenAI is combining chain-of-thought monitoring with activation monitoring (which he calls “confessions”) to keep track of how models generalize.
The broader calls
Beyond technical measures, the essay pushes for collective action: turning commitments like the Preparedness Framework into widely mandated safety bars enforced by third-party auditors, government agencies, or international bodies; treating cybersecurity as an urgent defensive priority while models become superhuman at breaking into systems; and pacing recursive self-improvement so people stay in the loop. The Next Web’s coverage notes the essay landed alongside OpenAI research posts describing 3.1 agent workdays for every human workday in its research organization, framing monitoring confidence — not raw capability — as the new bottleneck. OfficeChai’s summary headlines the same voluntary-slowdown message.
What this means for AI users
For anyone using AI agents day to day, the essay is a reminder that capability is racing ahead of auditability at the frontier. A practical response is the one Pachocki himself describes as desirable: keep humans in the loop, keep models observable, and run important work where you can see and control what the agent does. A local desktop client fits that posture — you can watch its steps, approve its actions, and switch models without being locked to one vendor’s reasoning you cannot inspect. If you want that level of control on Windows, download MOOGH and keep your AI work visible and on your own machine.