Speaking before economists, central bankers, and international finance officials at the annual meetings of the International Monetary Fund (IMF) and the World Bank in Bangkok, Anthropic co-founder Jack Clark delivered an uncompromising assessment: the technological singularity has officially begun.

Clark, who directs policy and societal impacts at the frontier research laboratory behind Claude, argued that artificial intelligence is already beginning to compress scientific workflows at speeds exceeding human oversight. In formal remarks accompanying the address, Clark projected a near-future landscape powered by the equivalent of "a billion virtual scientists," urging sovereign governments to invest immediately in physical infrastructure—such as automated wet labs, robotics, and high-performance computing—to absorb the coming influx of machine-driven scientific output.

Dramatic Jumps Across Frontier Benchmarks

To substantiate his claim that machine cognition is crossing critical thresholds, Clark pointed to steep performance gains on rigorous academic evaluations over short time horizons.

Chief among these metrics is Humanity’s Last Exam (HLE), an evaluation set assembled across dozens of academic disciplines to test advanced expert reasoning. Benchmark scores that stood at just 8% across frontier models in early 2025 have surged to 67.7%, recorded by Anthropic’s Claude Opus 5.5 operating with tool integration. Similarly, on FrontierMath—a benchmark testing complex, novel mathematical proofs—success rates vaulted from 2% in late 2024 to 94% within months.

While Anthropic has noted that evaluation score margins are becoming more nuanced indicators of real-world operational utility, the rate of increase across non-saturated benchmarks underscores a fundamental transformation in automated reasoning.

Beyond Simulation: Tangible Math and Physics Breakthroughs

Clark’s address emphasized that frontier artificial intelligence is no longer restricted to text synthesis and code generation; it is actively resolving specialized theoretical problems:

* Mathematical Verification: Anthropic recently documented an advance concerning the Riemann zeta function, in which an internal research model improved a established lower bound regarding the distribution of non-trivial zeros from 41.6% to 67.2%. The computational proof was independently verified via the Lean formal verification language. * Theoretical Physics: Independent evaluations using the TheoPhysEval benchmark revealed that advanced frontier systems correctly solved 49 out of 50 unpublished theoretical physics problems while equipped with specialized tools, performing the work at an estimated compute cost of under $3 per question.

These advances illustrate how AI systems are transitioning from passive assistants into active research collaborators capable of offloading dense computational phases, allowing human researchers to focus on hypothesis selection and validation.

Macroeconomic Repercussions and Defensive Security

The acceleration poses immediate challenges for macroeconomic modeling. Anthropic’s economic research division has projected aggressive growth scenarios under rapid autonomy adoption, modeling a potential U.S. gross domestic product of $44.4 trillion by 2030 (measured in 2025 dollars)—a 32.4% elevation compared to standard baseline projections without transformative AI.

However, advanced problem-solving capabilities also introduce asymmetric defensive challenges. Anthropic noted that specialized preview models have successfully pinpointed potential cryptographic vulnerabilities in experimental signature schemes and weakened encryption primitives. While existing commercial protocols remain unaffected, the findings demonstrate that institutional security teams must adopt equally capable AI models to protect digital infrastructure.

As Clark made clear to international policymakers, the boundary separating theoretical artificial general intelligence from global economic reality has dissolved. Sovereign institutions now face the pressing task of building legal, economic, and physical frameworks capable of keeping pace with autonomous discovery.