tasteful and strategic enough to reason about risks that do not yet exist
In a moment that marks the distance between today's tools and tomorrow's unknowns, OpenAI is paying up to $445,000 a year for someone to stand watch over a capability that does not yet fully exist — artificial intelligence that improves itself without human guidance. The hire reflects a broader industry reckoning: the pace of AI advancement has outrun even its architects' expectations, and the window for deliberate preparation may be narrowing faster than anyone anticipated. It is, in the oldest sense, the act of building the levee before the flood arrives — or trying to.
- AI coding tools are advancing so rapidly that the complexity of tasks frontier models can handle doubles roughly every seven months, compressing years of anticipated progress into months.
- Sam Altman has publicly committed to deploying a fully automated AI researcher by March 2028, turning what was once theoretical speculation into a corporate deadline with a public audience.
- Anthropic's co-founder estimates a 60% chance that AI R&D will proceed entirely without human involvement by end-2028, while a leading AI safety evaluator warns that any responsible civilization would be moving far more cautiously.
- OpenAI's new safety researcher would work to defend against data poisoning, interpret how models reason internally, and measure how deeply AI tools are already replacing human technical labor within the company itself.
- The gap between what the industry is building and what safety researchers believe is prudent appears to be widening — and the race to hire people who can think carefully about risks that don't yet exist is itself a signal of that tension.
OpenAI has posted a job listing for a safety researcher at up to $445,000 a year — not to address a crisis already unfolding, but to prepare for one that may be approaching. The role sits within the company's Preparedness team and centers on recursive self-improvement: the point at which AI systems can autonomously research, design, and train better versions of themselves, with little or no human involvement.
Six months ago, this was largely theoretical. It no longer feels that way. Researchers studying AI capabilities have found that the complexity of tasks frontier models can complete doubles roughly every seven months — a pace that surprised even the people building these systems. OpenAI is already monetizing AI coding through its Codex product, and Sam Altman has announced plans to deploy an automated AI research intern by September 2026 and a fully autonomous AI researcher by March 2028.
The person hired for this role would work across multiple fronts: defending models against data poisoning, building tools to interpret internal model reasoning, running experiments on self-improving systems, and tracking how extensively AI tools are already replacing human technical work inside OpenAI itself. The broader Preparedness team also addresses cybersecurity vulnerabilities, biological and chemical risks, and threats from agentic AI.
OpenAI is not alone in confronting these questions. Anthropic has published early research on using AI to oversee more powerful AI, with results described as promising but limited. The company's co-founder estimates a 60% probability that AI research and development will proceed entirely without human involvement by end-2028. The head of METR, a laboratory that evaluates AI capabilities, offered a starker assessment: a reasonable civilization would be moving much more slowly. The distance between what is being built and what safety researchers believe is wise appears, by most accounts, to be growing.
OpenAI is spending up to $445,000 a year to hire someone whose job is to worry about a problem that may not exist yet. The company posted a listing for a safety researcher to join its Preparedness team, tasked with preparing for the moment when artificial intelligence systems become capable of improving themselves without human intervention. The salary range—$295,000 to $445,000—is substantial, but what stands out is the unusual qualification the company is seeking: someone "tasteful and strategic" enough to reason about risks that exist only in theory.
The concept at the center of this hire is called recursive self-improvement. It describes an AI system that can research, design, and train better versions of itself autonomously, with minimal or no human involvement. Six months ago, this was largely academic speculation. Today, it has become an active priority across the industry. The shift happened because coding tools from OpenAI and Anthropic have advanced faster than even their creators anticipated. Researchers at METR, a laboratory studying AI capabilities, found that the complexity of tasks frontier AI models can complete doubles roughly every seven months. The practical implication is stark: within months or a few years, AI systems may handle software engineering work that currently takes human programmers days or weeks.
OpenAI's leadership has been explicit about the timeline. Sam Altman announced in October that the company aims to deploy an "automated AI research intern" running on hundreds of thousands of chips by September 2026, followed by a "true automated AI researcher" by March 2028. He acknowledged the risk of failure but framed transparency about these ambitions as a matter of public interest. The company, preparing for a public offering this year, is already monetizing AI coding through its Codex product. Automating its own research operations represents the next frontier.
The researcher hired for this role would work across several fronts. They might defend OpenAI's models against data poisoning—attacks that corrupt AI systems through their training datasets. They could build tools to interpret how models reason internally. They would run experiments to understand what happens when self-improving systems are let loose. They would also track how extensively AI coding tools are being used within OpenAI itself, measuring progress toward automating the work of technical staff. The broader Preparedness team, which this role sits within, focuses on preventing severe harms from AI. Other open positions on the team address automated red-teaming for cybersecurity vulnerabilities, biological and chemical risks, and threats from agentic AI systems.
OpenAI is not alone in this race. Anthropic published research in April on using AI models to oversee more powerful AI models, though results were described as promising but limited. Jack Clark, Anthropic's co-founder and policy head, estimated in May that there is roughly a 60 percent chance of seeing AI research and development conducted entirely without human involvement by the end of 2028. Elizabeth Barnes, chief executive of METR, offered a sharper observation: any reasonable civilization would be moving much more slowly and carefully with AI development. The gap between what companies are building and what safety researchers believe is prudent appears to be widening.
Citas Notables
This work relies on reasoning about problems that might exist in the future, but might not exist now. So it's especially important that people in this role are tasteful and strategic.— OpenAI job listing
Any reasonable civilization would clearly be taking things much more slowly and carefully with AI.— Elizabeth Barnes, METR chief executive
La Conversación del Hearth Otra perspectiva de la historia
Why would OpenAI need to hire someone to think about a problem that doesn't exist yet? Isn't that premature?
Because the problem is arriving faster than anyone expected. Six months ago, recursive self-improvement was theoretical. Now the capabilities are advancing so quickly that researchers think it could happen within two years. You don't wait for the fire to start before you think about how to contain it.
But what does it actually mean for an AI to improve itself? How is that different from just getting better at its job?
The difference is autonomy. Right now, when an AI gets better, humans are involved—we collect data, design the next version, train it, test it. Recursive self-improvement means the AI does all of that itself. It identifies its own weaknesses, designs solutions, trains itself, and repeats. Without asking permission.
And OpenAI thinks this will happen by 2028? That seems aggressive.
Sam Altman said it publicly. They're aiming for an automated AI researcher by March 2028. He acknowledged they might fail, but the fact that they're being transparent about the timeline suggests they take it seriously internally. And Anthropic's Jack Clark thinks there's a 60 percent chance it happens by the end of 2028.
What's the actual job, then? What does this researcher do with all that money?
They're essentially building the guardrails before the train leaves the station. Testing what happens when self-improving systems run. Figuring out how to defend against attacks on those systems. Measuring how fast the capability is actually advancing. It's preparation work—trying to understand the problem well enough to contain it if it arrives.
Does anyone think this is moving too fast?
Yes. The head of METR said any reasonable civilization would be moving much more slowly and carefully. But the companies building this see the potential and the competitive pressure. So they're hiring safety researchers to keep up.