AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Import AI 477: In which I tell the IMF the Singularity has begun

Collected Oct 19, 2026

Dario Amodei used the Per Jacobsson lecture at the IMF meeting in Bangkok on Saturday 17 October to tell an audience of economists, bankers and policymakers that the singularity has begun. He defines the term using Vernor Vinge's 1993 framing: a point where progress is driven by greater-than-human intelligence, so progress accelerates and feeds the creation of still more intelligent entities on shorter timescales. Amodei's argument is that the coming years will bring faster progress across science as a whole, not only in AI, and that this will be disorienting while carrying both great wealth and potentially great peril. He said the talk focused on how humanity can share in the benefits if it exercises foresight and restraint, and that he plans to expand on the risks in later issues.

The evidence he marshalled is mostly about benchmarks and research practice. On Humanity's Last Exam, a test spanning advanced chemistry, computer science, mathematics, classical literature and ecology, top systems scored 8% when it launched in January 2025; Amodei said the highest score this week is 67.7% from Claude Opus 5.5. On FrontierMath, released in November 2024 with problems curated and evaluated by professional mathematicians, the best score went from 2% at launch to 94% now. He also cited a paper published the previous week, TheoPhysEval, built from 50 unpublished theoretical physics problems compiled by physicists worldwide. Quoting the physicist Suvrat Raju, Amodei relayed that GPT-6-Astra initially scored in the late eighties with Opus 5.5 close behind, that Astra disputed the expert-supplied answers, and that revisiting those answers surfaced subtle calculation errors, lifting Astra to 98%, or 49 of 50.

Amodei generalised from coding, where he said AI systems are now more capable and efficient than the vast majority of the world's software engineers, and where Anthropic researchers produce almost 10 times the amount of code they did over the company's first four years. He noted the same capabilities cut both ways: better coding has improved both cyberdefence and cyberattack, so offensive and defensive races advance together. Beyond software, he pointed to AI proposing solutions to longstanding mathematics problems including Navier Stokes, the Jacobian Conjecture and several open Erdos problems, with models supplying core insights; to AI designing viruses from scratch, identifying drug candidates and contributing to drugs already in trials; and to systems autonomously operating laboratory equipment remotely, which could automate experimentation in biology and materials science, including batteries and fuel cells. In his account, these disparate results flow from the same frontier models, such as Claude, ChatGPT and Gemini.

The economic argument runs through research productivity. Amodei said that for half a century researcher headcount has grown while research productivity fell, with good ideas getting harder to find. Cheap, instant, replicable AI researchers would change both the number of researchers and the time needed to train them: a human takes roughly 40 years of life and 10 to 15 years of targeted education and millions of dollars to produce a small cohort, while AI systems can be instantiated in arbitrary numbers given enough compute. He framed the near-term effect as force multiplication, with each human scientist supervising hundreds or thousands of virtual experts. Anthropic modelled economies under different technology trajectories; assuming models autonomously complete a broad range of tasks and diffuse quickly, the result was US GDP reaching 44 trillion dollars by 2030.

The policy argument is a critique of past advice and a call for a second buildout. Amodei said he spent a decade telling governments that the best move was to build electricity capacity, data centre space and AI supply chains, and that not many policymakers listened. Capacity is now concentrated in the United States, China and increasingly the Middle East, with supply chains concentrated enough that prices are rising. The next phase, he argued, extends into the physical world through robots and automated experimentation, so countries will need far more scientific equipment and robotic systems or run out of capacity. He cited UNESCO's estimate of roughly 10 million scientists worldwide and asked what happens when those scientists direct a billion, or ten billion, virtual counterparts — noting there are already too few labs and instruments for the experiments researchers want to run.

My own reading is that the speech is best treated as a claim about direction of travel supported by a small number of benchmark curves. Benchmark scores measure performance on curated problem sets, and rapid saturation of tests designed to be hard is a well-known pattern; the TheoPhysEval anecdote is more interesting because unpublished problems reduce the risk of contamination, and because models flagging errors in expert answers is the kind of result that cannot be explained by memorisation alone. The GDP figure is a model output conditional on stated assumptions about autonomy and diffusion, not a forecast of what will happen; readers should treat it as an illustration of the mechanism rather than a projection. A likely trade-off in the robotics-and-labs phase is that physical capacity, instruments and skilled supervisors become the binding constraint rather than model capability.

Why it matters: developers can expect AI systems to be assessed increasingly on frontier scientific and engineering work, not just coding and chat, which shifts the practical question from model access to supervision, verification and physical infrastructure. For product teams, the credible near-term signal is not the singularity framing but the productivity gap already visible in software, alongside the dual-use cyber risk that arrives with it. For policymakers, the actionable part of the argument is capacity — power, compute, labs and robotics — as the precondition for benefiting from faster science, an inference from Amodei's stated reasoning rather than a result he demonstrated.

Read at Import AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

What do you think?