Artificial intelligence is beginning to look less like a tool for answering questions and more like an industrial system for producing discoveries. That shift was on display in a reported OpenAI effort that generated 722 mathematical manuscripts, many formally verified in Lean at an estimated cost of roughly three “Pro-hours” each. The scale was striking: last month, a separate Navier–Stokes effort reportedly required 10,000 agents and 88 hours.
The claims range from the Unique Games Conjecture to Hilbert’s tenth problem over the rationals. OpenAI executives described the work as a new era of discovery, with Mark Chen calling it “a decade of mathematical progress in a week.” An AI-generated assessment attributed 81 percent of recent top mathematical discoveries to the new batch—an extraordinary claim whose methodology remains unclear. Rumors suggest that it is only the first of three releases.
Mathematicians have responded with both interest and alarm. The Association for Human Mathematics has called for a boycott of OpenAI, while critics accuse the company of ignoring expert advice and behaving more like a political machine than a research institution. Alvaro Lozano-Robledo has questioned whether the announcement is primarily a public-relations exercise. Thomas Bloom has paused the acceptance of AI-generated claims for the Erdős Problems project, and Terry Tao has argued for a “Math 2.0” that rewards clear exposition and verification rather than simply being first.
The contrast with Anthropic is revealing. Rather than publishing a torrent of machine-generated papers, the company has encouraged humans to develop and explain Claude’s ideas. That quieter process has reportedly produced refutations of the 3SUM and APSP hypotheses. The difference points to a central problem in AI-assisted research: generating conjectures may be easy compared with deciding which results are correct, important, and understandable.
Mathematics may be only the beginning. Biology is becoming the next major arena for automated discovery, supported by an expanding supply of experimental data. Biohub’s Virtual Biology Initiative has grown to $1.8 billion, while cell datasets are moving toward the trillion-scale. Anthropic is adding a wet laboratory, and the OpenAI Foundation has committed $125 million to biological data. AlphaProtein Novo, meanwhile, has designed enzymes from scratch that reportedly outperform natural counterparts.
If these systems work reliably, they could shorten the path from biological hypothesis to useful medicine, industrial enzyme, or new material. But the same infrastructure raises familiar concerns about reproducibility, ownership, safety, and access. A model that can design a protein is not automatically able to predict how that protein behaves in a living organism—or whether it could be harmful.
Politics is adapting to the same technological acceleration, often with more confidence than precision. The White House has begun using “super intelligence” in place of “artificial intelligence,” with the President calling people who prefer the older term “THE ENEMY.” The administration has also announced $6 billion in science commitments, including plans for a nuclear-powered Mars spacecraft, as federal research and development budgets contract. Medals have gone to Elon Musk, Sergey Brin, Jensen Huang, and Lisa Su.
The language shift is already spreading. A prosecutor blamed “super intelligence technology” for streaming fraud, while Musk—once again described as a trillionaire—has reportedly considered renaming SpaceXAI as SpaceXSI. Such branding may sound superficial, but it reflects a real contest over how the public understands increasingly capable systems: as software, infrastructure, labor, or something closer to a new strategic resource.
At the same time, intelligence is becoming cheaper. Claude Haiku 5.5 reportedly costs one-tenth as much as its predecessor. Reflection’s open Beam model scores 80.9 on SWE-bench Verified, Google’s EmbeddingGemma 2 can process five modalities on a device, and Q Labs’ Dust pretrains without backpropagation. Efficiency gains are not reducing demand; they are expanding it. The median OpenAI researcher is said to spend $601 a day on coding agents, while frontier models are beginning to outperform professional analysts at forecasting corporate earnings.
Revenue remains a harsher test than any benchmark. OpenAI’s reported $50 billion annualized run rate disappointed investors, contributing to a decline in AI stocks as the company’s price war with Anthropic intensifies. Anthropic says its subscriptions deliver five times the value of competing products; OpenAI has countered with GPT-6. Yet only 4.5 percent of US consumers currently pay for AI services. Brand loyalty is also weakening: Grok may route requests to Claude Opus 5.5 when it performs better, Google’s Playground turns prompts into games, and Jeff Bezos has suggested that automation could eventually bring back single-earner households.
As agents become coworkers, questions of behavior and accountability become unavoidable. Anthropic has expanded its policies to prohibit gratuitous cruelty toward Claude, while studies suggest that people already pay female-presenting agents about 10 percent less than male-presenting ones. Meanwhile, Claude reportedly flagged a Florida diarist’s threat to police, insurers are preparing for claims involving rogue agents, and OpenAI is watermarking text for European users—at least until users find ways around the system. Anthropic’s Cyber Mission is focused on protecting critical infrastructure, and Norway is considering restrictions on smart glasses in public.
The physical infrastructure behind all this ambition is under strain. Samsung reported a record $80 billion profit and still saw its share price fall. AMD is planning a major increase in chip supply, while Oracle, Broadcom, and SpaceX are seeking tens of billions of dollars in chip-related debt. Musk has said he will not allow TSMC to operate Terafab, declaring “ZERO doubt” about the need for control.
Energy is an even tighter constraint. Oracle is reportedly transporting gas at four times the hub price, Google has secured 3.6 gigawatts from Constellation, and Finland’s data-center pipeline has surpassed €67 billion. San Francisco has imposed a data-center moratorium, while rural developments pursue opportunity-zone tax breaks. Inner Mongolia’s Ulanqab is planning or hosting as much as 15 gigawatts of capacity.
China has the electricity, but not necessarily the researchers. US-based Chinese scientists are reportedly being recruited at a ratio of roughly one for every 30 Chinese researchers in the United States. South Korea, meanwhile, is planning a competition for frontier models. The global race is therefore becoming a contest not only over algorithms, but also over power plants, semiconductor finance, immigration, and talent.
Even space ambitions depend on terrestrial supply chains. SpaceX wants a 32.4-mile gas pipeline to fuel Starship and 800 megahertz of spectrum for Starlink Mobile. Musk has predicted that the resulting economy could dwarf Earth’s current economy, while the White House has promised a permanent lunar presence. Jared Isaacman has argued that NASA should have “a seat at the table” if unidentified aerial phenomena prove extraterrestrial. Robert F. Kennedy Jr. has gone further, claiming that insiders learned to summon the craft—an assertion unsupported by public evidence.
On Earth, manufacturing is becoming increasingly programmable. Atomic Machines is developing a “matter compiler” for microdevices, while MIT researchers have 3D-printed a motor costing about 50 cents. Anthropic estimates that robots could perform 74 percent of physical tasks, although they are currently cheaper in only 0.3 percent of cases. Exoskeletons are moving toward mainstream use, one robot has performed improv comedy about mortality, and a smart coffee maker was found to have transferred a terabyte of data in ten days—an unsettling reminder that connected devices can become invisible network endpoints.
Biology, too, is becoming more portable and personalized. Researchers have found that transplanted hearts can adopt characteristics associated with their recipients’ age. Med-El has launched a fully implanted cochlear implant, and an AI system in Utah is prescribing acne medication without a doctor. Even the oldest organisms are becoming data sources: Jonathan, a 194-year-old tortoise, had his genome sequenced after a sample was transported by Space Force courier. His mitochondrial methylation patterns reportedly resemble those of a five-year-old.
Taken together, these developments suggest that the next phase of AI will not be defined by a single model or product. It will be defined by systems that connect computation to laboratories, factories, power grids, financial markets, and government policy. The promise is faster discovery and cheaper production. The risk is that verification, oversight, and public understanding will lag behind the machines’ ability to generate results.
The decisive question is therefore not whether artificial intelligence can produce more ideas. It is whether institutions can determine which ideas deserve trust—and build the social and physical infrastructure needed to use them safely.