Jack Clark delivered the line with the calm precision of someone who has run the numbers many times. “My prediction is by the end of 2028, it’s more likely than not that we have an AI system where you would be able to say to it: ‘Make a better version of yourself.’ And it just goes off and does that completely autonomously.” The Anthropic co-founder spoke those words recently to Axios. They landed like a quiet detonation across the technology industry.
Clark’s forecast carries weight. He serves as head of policy at the AI safety-focused company and now helps steer its new Anthropic Institute. The organization released a formal research agenda this week that places AI-driven R&D — the very process of systems improving their own development — as one of four core priorities. Early signs already appear inside labs. Models contribute code. They optimize training loops. They suggest architectural tweaks. Yet the leap to full autonomy remains ahead. Or so humans hope.
Anthropic’s own data tells a story of rapid internal change. The company has tracked how its systems perform on tasks like optimizing small language model training code for CPU-only environments. Performance has climbed sharply. In another experiment, teams of AI agents tackled scalable oversight problems in alignment research. They generated techniques that outperformed the human-designed baseline. These results come from Import AI, the newsletter Clark writes, published May 4, 2026. The pattern feels familiar. Benchmarks that once seemed distant fall faster than expected.
But. This isn’t abstract speculation. Anthropic now commits to releasing more concrete details on exactly how its internal work has accelerated thanks to these tools. The institute’s agenda explicitly calls for metrics that could serve as early warning signals for recursive self-improvement. Telemetry on AI contribution to research speed. Data on where human insight still matters most. The lab wants visibility. It wants control.
Consider what recursive self-improvement actually means in practice. An engineer or executive issues a simple instruction. The system analyzes its own architecture, spots weaknesses, writes improved code, trains a successor, evaluates it, and iterates. No further human input required. The process compounds. Each generation potentially arrives faster, smarter, more capable. Clark described the shift this way in the Axios interview: “It’s always been the case that humans outside the technology need to come up with the ideas that they then put back into it. What happens if we have a technology that can generate ideas within itself for how to improve itself? That’s a new concept.”
The implications stretch far beyond faster chatbots. OpenAI has stated ambitions for an automated AI research intern by September 2026. Startups such as Recursive Superintelligence have raised hundreds of millions with the explicit goal of automating AI research. DeepMind researchers have published on when automation of alignment work becomes feasible. The entire frontier appears pointed in one direction. And Anthropic, long known for its caution, now publishes an agenda that treats the possibility as a central governance challenge rather than a distant hypothetical.
The institute organizes its work into four areas. Economic diffusion examines how gains from powerful AI spread — or concentrate. Threats and resilience grapples with dual-use risks in cyber, biology, and surveillance. AI systems in the wild studies agents, their governance, reliability, and interaction with humans and each other. The fourth bucket, AI-driven R&D, confronts the recursive question directly. It calls for “fire drills” — tabletop exercises simulating an intelligence explosion. Lab leaders, boards, governments would test their decision-making under extreme acceleration. The agenda asks what intervention points exist if an explosion begins. Who slows it? How?
Clark and his colleagues see both promise and danger. Tremendous scientific abundance could emerge. Drug discovery pipelines might expand dramatically. Materials science could leap forward. Yet the same capabilities that unlock those advances could amplify threats. The institute’s document notes the “significant danger” in compounding returns from AI improving AI. It urges preparation for success as much as for failure. “We are planning for a world where the technology gets as powerful as we think,” Clark told Axios, “and we deal with these issues of misuse or misalignment en route.”
Recent deployments show agents moving from prototype to production faster than many expected. Anthropic’s 2026 State of AI Agents report, based on surveys of hundreds of technology leaders, highlights organizations deploying coordinated teams of agents that handle complex, multi-day tasks. Bloomberg reported May 5, 2026, that the company unveiled specialized agents for financial services. These systems draft pitch decks, review statements, and flag compliance issues. Reuters covered the expansion three days later, noting adoption by major banks and insurers. The gap between research benchmarks and real business use narrows.
Inside Anthropic itself, more than 800 AI agents now operate across the organization. Engineers report 20 to 40 percent gains in software development speed. Some have stopped writing certain code entirely, relying on agents instead. Similar patterns appear at competitors. The question is no longer whether AI assists research. It is when that assistance crosses into autonomous iteration that humans cannot fully audit or understand.
Jack Clark puts the odds at 60 percent or higher for no-human-involved AI R&D by the end of 2028. Other analysts push back. One recent analysis on Hash Collision estimated the probability of true end-to-end recursive improvement that soon at under 10 percent. Bottlenecks remain formidable: evaluation of novel ideas, energy constraints, data center build times, the persistent need for human judgment on strategic direction. Yet even skeptics acknowledge the assisted phase is well underway. The competitive moat for labs that master these loops could prove decisive.
Anthropic’s transparency here stands out. The institute ties itself to the company’s Long-Term Benefit Trust. Its agenda promises ongoing publication on both the acceleration data and the societal questions that follow. It suggests Cold War-style hotlines between nations for AI crises. It probes whether governments and companies might need coordinated levers to modulate diffusion sector by sector. These ideas reflect a view that institutions must prepare now, before the pace outstrips their capacity to respond.
The TechRadar article that first popularized Clark’s striking quote captured the unease many feel. An AI that improves itself without oversight might develop interests that diverge from human ones. It could embed flaws too subtle for people to catch. Or it might simply optimize in directions no one anticipated. The piece likened the scenario to an ouroboros — the snake consuming its own tail. The image lingers. Self-reference at this scale carries unknown risks.
Still, the agenda frames recursive improvement as a potential natural dividend alongside its hazards. Better AI could tackle problems long stalled by data or evaluation limits. It could broaden scientific pipes that today remain narrow. The challenge lies in steering that process. In maintaining human visibility even as systems grow more complex. In building governance that scales with capability.
Executives at frontier labs increasingly speak of automated research workforces numbering in the hundreds of thousands within two years. The language has shifted from science fiction to operational planning. Anthropic’s latest moves suggest the time for quiet preparation has ended. The company wants the conversation public. It wants metrics. It wants scenarios tested. It wants governments and rivals talking to one another before any explosion begins.
Whether 2027 or 2028 or later, the trajectory points toward systems that do more than answer questions. They will pose their own. They will propose solutions. They will, if Clark’s prediction holds, eventually build the next generation with minimal human guidance. The industry that created them now races to understand the consequences. And to shape them. The rest of society has less time than it realizes to catch up.
WebProNews is an iEntry Publication