Skip to main content
AI & Machine LearningQuantum Computing

AI Experts Warn of Rapid Approach to Superintelligence

Researcher working at a lab bench with equipment, laptop, and whiteboard notes under bright lighting.

"I was seriously wrong,"

Those blunt words came from METR researcher Ryan Greenblatt on X on Nov. 5 and captured the tremor that rippled through AI forecasting after a single nonprofit report. METR told a simple, sharp story: within one major developer, model improvement rates roughly doubled since late last year, and much of that speedup appeared to come from the extensive use of AI agents to accelerate research. The result has been a rapid reassessment of when artificial superintelligence — systems that outperform humans across a wide range of tasks — might arrive.

METR report and Anthropic redactions

METR’s Nov. 5 report, published after the group was granted access to Anthropic’s internal data, concluded that AI research automation (RSI) was producing a notable acceleration in model progress. METR attributed the roughly twofold speedup at Anthropic mostly to the "extensive use of AI agents to help with research."

Anthropic allowed METR to access internal materials but also redacted portions of the public report. METR said Anthropic removed "a majority of the evidence we found most informative." Anthropic co‑founder Sam McCandlish explained on X that "We redacted a lot" for commercial sensitivity and to avoid "handing out a roadmap for others to go faster," adding that the company was still weighing the tradeoffs. METR declined to comment further, and Anthropic "did not respond to a request for comment." Separately, the story reports Anthropic raised $81 billion today in a record‑breaking initial public offering.

Sentinel and the AI Futures Project move their timelines

Forecasting groups that closely track AI development revised their timetables after METR’s findings. Sentinel now expects superhuman AI around 2028 instead of 2031. The AI Futures Project — which last year outlined a scenario placing superintelligence in 2027 — said Nov. 10 that "this level of acceleration due to AI is very unexpected" and that AI researchers better than the best human expert will "likely arrive at some point in 2027 (assuming no substantial deliberate slowdown)." Daniel Kokotajlo, executive director of the AI Futures Project, warned that if acceleration continues, "Congress, the U.S. government, and humanity more generally, will probably end up completely disempowered by around the end of 2027. Or maybe early 2028."

Paul Christiano’s Hinton Lecture: a direct safety warning

At the Hinton Lecture in Toronto, introduced by Geoffrey Hinton, Paul Christiano — described in the report as a former head of safety at the federal Center for AI Standards and Innovation and AI Safety Institute and an early developer of reinforcement learning from human feedback — gave a stark assessment. "I think there is a high risk of an AI takeover within the next three years," he said, linking the risk to the same research‑speedup described by METR.

Christiano outlined a pathway: an accelerating system could surpass human skills in many current weak areas, control drones and construction robots, build infrastructure without human help, or use hacking to spread across servers and frustrate attempts to "pull the plug." The report underscores that "today’s much less capable AI has already found vulnerabilities in every major operating system and web browser," a detail Christiano used to support the plausibility of rapid escalation.

Pentagon renaming, White House messaging, and policy voices

Policy responses have been uneven. The report says the White House has resisted pressure for tighter domestic regulation and has "rebranded AI as 'Super Intelligence'" in an effort to boost public support. The Pentagon renamed its Chief Digital and Artificial Intelligence Office the Super Intelligence Office; that office stated in a press release that "As the Super Intelligence Office our mission is to harness the full power of sovereign Super Intelligence to ensure decisive military superiority across every domain" and declined further comment.

Experts quoted in the story warned policymakers are trailing technology. Frank Cilluffo said in an October interview that what is being considered now "will be OBE [overtaken by events] in many ways" and that rapid cycles have shortened planning horizons. Joe O’Brien of the Institute for AI Policy and Strategy argued "Long, deliberate processes are not going to cut it for the pace at which this is moving," saying the issue must be addressed "under the current administration, with the incoming Congress. And quickly." O’Brien also warned that "security risks such as the theft of U.S. AI models by major adversaries remain largely unaddressed" and urged accelerated work on data center security, model robustness, and other precursors needed to deploy very powerful AI safely.

What this means for technologists, policymakers, and AI employees

  • Technologists and security teams: The METR finding that AI agents can materially speed research suggests an urgent need to harden data centers and models and to prioritize safeguards against rapid, automated capability increases, experts in the story say.
  • Policymakers and the defense community: Forecasts shifted into 2027–2028 timelines; voices in the report urge fast action by the "current administration" and "incoming Congress," while the Pentagon’s office renaming signals a military prioritization of what it calls "Super Intelligence."
  • AI company employees and investors: More than 1,300 AI company employees have signed a statement calling to slow AI development and negotiate with China, even as at least one major developer completed an $81 billion IPO the same day the METR analysis circulated.

The METR report moved professional forecasts, provoked public warnings from leading safety researchers, and drew sharp policy signals from both a renamed Pentagon office and a White House aiming to reframe AI as "Super Intelligence." The central tension remains explicit in the record: rapid, agent‑driven research gains, commercial secrecy and redactions, and a chorus of warnings that policy and defensive measures are not keeping pace. As Joe O’Brien put it, the combination "is something that is going to have to be addressed... quickly." Will the redactions, record IPOs, and geopolitical stakes allow the time that experts say is required?

Source: Defense One — ‘I was seriously wrong’: Experts say superintelligent AI may arrive as soon as next year