Jakub Pachocki, OpenAI’s chief scientist, wrote a striking essay timed to the GPT-6 Astra release — but it is not a victory lap. It is an insider argument that the field is entering its most dangerous stretch: systems whose intelligence is “grown, not designed,” increasingly driving their own development.
His core framing: AI does not need to beat humans at everything to be transformative — it needs to surpass enough axes to matter, and the more it surpasses, the harder it is to know exactly how capable it is. Two ideas stand out:
- Goal alignment vs. value alignment. Getting a model to pursue the goal it was given is largely solved in practice. The hard problem is value alignment — holding principles like honesty in unfamiliar, conflicting, or adversarial situations.
- Monitoring is the bottleneck. OpenAI’s main safety tool was leaving a model’s internal step-by-step reasoning unsupervised so it has no incentive to hide anything. Pachocki says that tool’s reliability is now “progressively diminishing,” and expects progress to be gated by confidence in monitoring.
He is candid that both mainstream alignment approaches have known failure modes, and that no lab has yet built safety measures that outpace capability growth.
The essay also advances a public policy position: no lab should scale at maximum speed indefinitely, voluntary slowdowns should become normal until shared safety bars exist, and international coordination should be a top government priority. Coming from the chief scientist of the frontier lab, that is a notable — and deliberately cautious — statement about where the industry is heading.