Tech • AI • Robotics • Game
Anthropic Just Shocked the World... AI Singularity Is Here!
TL;DR
Anthropic co-founder Jack Clark told an IMF audience that frontier AI is already delivering near-expert or expert-level performance across multiple research domains, while a new economic model suggests it could lift US GDP to $44.4 trillion by 2030 in an extreme adoption scenario, with major gains paired with severe disruption for knowledge workers.
KEY POINTS
Clark’s singularity claim
At the Per Jacobson Lecture in Bangkok, Clark argued that the “singularity” is no longer a distant prospect but a present reality. He pointed to rapid capability gains inside Anthropic, including a claim that employees now generate nearly 10 times more code than before, and said AI systems have evolved from assistive tools into what he called “virtual polymaths.”
Benchmark leaps in under two years
Clark highlighted a sharp jump on Humanity’s Last Exam, a broad test spanning chemistry, computer science, mathematics, literature and ecology. The strongest AI reportedly scored 8% when the benchmark appeared in January 2025, while Claude Opus 5.5 has now reached 67.7%. On Frontier Math, performance was said to have climbed from 2% at launch in late 2024 to 94%.
Physics test stirred the strongest reaction
A new theoretical-physics evaluation known as TheoPhys used 50 unpublished research questions written by leading physicists, limiting the possibility of memorized answers. GPT-6, Astra and Opus 5.5 initially scored above 80%, but later review pushed one model to 98%, or 49 out of 50, after physicists found small errors in some of their own reference solutions that the AI had flagged correctly.
A vision of mass digital research labor
Clark framed the implications in labor terms: there are only about 10 million full-time researchers worldwide, yet modern civilization depends heavily on their output. In the scenario he described, compute could eventually support not just thousands but potentially billions of tireless, cross-disciplinary digital researchers, compressing drug development timelines from years into weeks and accelerating materials discovery in fields such as superconductors and batteries.
Recursive self-improvement is moving from theory to concern
One of the most sensitive claims was that AI is increasingly helping build better AI. That prospect, often described as recursive self-improvement, has long been treated as a theoretical threshold in AI safety debates. It is also a geopolitical concern because it raises the possibility that early leaders could compound their advantage quickly.
How the $44.4 trillion figure was built
The widely cited $44.4 trillion estimate comes from an Anthropic Institute report, Economic Scenarios for Transformative AI, by Anton Korinek, Chad Jones, Simon Sakaue, Tess Carter and Peter McCrory. The model treats the economy as a bundle of tasks drawn from labor data. It asks what share of those tasks AI can do, how broadly it is adopted, how much productivity improves, how much work is automated versus augmented, and how many new human tasks appear.
Three scenarios, from internet-like gains to economic shock
Without AI, the model places US GDP at about $33.5 trillion in 2030. In a modest case, GDP is 1.6% higher at $34.1 trillion. In a substantial case, GDP is 8.3% higher at $36.3 trillion, with annual growth rising to 5.4%. In the extreme case, GDP is 32.4% above baseline, reaching $44.4 trillion, while annual growth around 2030 hits 15.4%.
The biggest gains do not come from scientific discovery in the model
One striking feature of the report is that accelerated science contributes little to the headline result by 2030. Even in the extreme case, the stock of ideas is only 0.6% above the no-AI path. That means the $44.4 trillion outcome mainly reflects AI doing existing work far more efficiently, not a full-blown explosion of new discoveries.
Knowledge workers absorb most of the pain
The model divides workers into knowledge occupations and everyone else. In the extreme case, knowledge-worker wages end up 11.5% below their no-AI path, knowledge employment falls 21.5%, and unemployment for that group rises to 17.9%. Overall unemployment reaches 11.9%. If wages are rigid, knowledge-worker unemployment climbs to 24%; if wages adjust fully, unemployment stays lower but pay falls by more than 42%.
Capital captures a larger share of the upside
Today, labor receives roughly 60% of national income and capital about 40%. In the extreme scenario, that flips to 45% for labor and 55% for capital. Total labor income in 2030 is almost flat, up just 0.5%, while capital income jumps 81%. The report estimates that transfers equal to about 9% of GDP could offset losses for displaced knowledge workers, but acknowledges that compensation on that scale would be historically unprecedented.
CONCLUSION
The analysis presented to central bankers combined extraordinary claims about AI capability with a sobering distributional warning. Even if transformative AI delivers historic growth, the decisive policy challenge may be how to manage who wins, who loses and how quickly the shock arrives.
Ask a question