A job posting that reads more like an unintentional thought experiment than a corporate listing is circulating among AI circles. The company that created ChatGPT, OpenAI, is seeking a safety researcher to help them get ready for the day when AI will be able to write better code, create more intelligent architectures, and perform tasks that are currently performed by human researchers. The pay range is between $295,000 and $445,000. What’s required? Become “tasteful and strategic.”
It’s difficult to avoid thinking about that phrase for a while. Delicious. The word seems almost charming in a field where the stakes are, according to OpenAI’s own framing, potentially civilization-scale. However, if you carefully read the posting, you’ll see a strange picture of a company that hires someone to figure out exactly what it’s preparing for before it happens.
The position is part of OpenAI’s Preparedness team, which focuses on what the company refers to as extremely serious threats to international security. These are not optics-focused theoretical exercises. The group attempts to foresee misuse scenarios before they arise, tracks actual capability curves, and develops mitigation strategies. Their current focus is on recursive self-improvement, which is the notion that an AI system could iteratively improve its own capabilities without significant human involvement.
Here, context is important. Both OpenAI and Anthropic’s coding tools have seen exceptionally rapid increases in capability over the past six months. Individuals in these labs became aware. The CEO of Google DeepMind, Demis Hassabis, also stated that humanity is currently at the “foothills of the singularity.” It’s still unclear if that framing is accurate or theatrical, but it’s telling that those closest to this technology are using such language.
The work described in the job posting is really challenging. Tracking the rate at which AI is automating technical research tasks, creating defenses against data poisoning attacks, creating experiments to find out if models are subtly misaligned, and creating tools to interpret model reasoning directly from neural activations are some possible areas of focus. On its own, each of those tasks would be a significant research endeavor. The job description calls for one person to work seamlessly between all of them, creating rudimentary prototypes and quickly iterating. Although it’s possible that no single hire can actually accomplish all of this, the posting’s urgency suggests OpenAI believes it cannot afford to wait for the ideal applicant.
There’s a feeling that this listing is also a signal, an effort to let regulators, the public, and the larger research community know that OpenAI is seriously considering the risks associated with recursive self-improvement before the technology is fully developed. Depending on who you ask, that may be careful reputation management or sincere preparation. Perhaps the most honest thing that can be said about this industry at the moment is that when you watch this dynamic unfold from the outside, it seems like both things are true at the same time.
Beneath all of this is a subtle but important irony. Building artificial general intelligence is OpenAI’s stated objective. At the same time, it is employing people expressly to guard against the fallout from accomplishing that objective. The precaution and the mission are funded by the same balance sheet and operate concurrently within the same building. No job posting, no matter how well compensated, can yet provide an answer to the question of whether those two efforts can truly keep up with one another.
