An open quantum computing conjecture had been awaiting proof for 16 years. An AI system proved it on its third day of operation and then passed Lean verification after two corrections. This shows that AI-assisted research has moved beyond retrieving information to producing new, testable results.
What This Is
Physical Superintelligence (PSI) used an undisclosed Anthropic model to prove a conjecture proposed by Scott Aaronson and Andris Ambainis in 2008. Lean is a proof assistant that checks whether a logical derivation is valid step by step. The crucial point is not merely that the system “gave an answer,” but that it produced a complete, machine-verifiable proof.
Another development comes from Meta Reality Labs. Its Tactus tactile glove uses fabric-integrated sensors to record details such as force and posture when fingers touch objects. It aims to address a problem cameras face when contact is obscured and cannot be clearly observed. Such devices are both data-collection tools and potential execution interfaces for robots entering homes and factories.
Industry View
We believe the key development is the redistribution of work in research: AI proposes and tests large numbers of hypotheses, while researchers select problems, interpret their significance, and bear responsibility. Improvements in model capability need to appear not only in the quality of conversations, but also in whether they can produce results society is willing to adopt.
The risks are equally clear. A single case does not prove that AI has broadly surpassed experts. The conjecture’s long public availability also provided mature conditions for verification. More realistic questions include whether different teams can independently review the proof, how research findings can be reproduced, and whether the cost of running models will cancel out the benefits.
Meta’s approach also exposes a hardware bottleneck. AI can reason within a screen, but acting requires sensors, robots, and feedback from real environments. Whether data can scale and costs can fall matters more than staging a polished demonstration.
Impact on Regular People
For enterprise IT: Companies should prioritize AI use cases that are verifiable and can integrate into existing workflows, such as materials research, code verification, and quality inspection, rather than purchasing a generic “research agent” outright.
For careers: Knowledge work will increasingly adopt a model in which machines produce first drafts and humans review the evidence. Defining problems, judging accountability, and explaining results will become more important.
For the consumer market: In the short term, tactile gloves are more likely to support robot training and industrial data collection. As execution devices mature, home-service products may gain more reliable fine-motor capabilities.