AI news, 8 November: AI system proves 18-year-old quantum computing conjecture

AI proves long-open quantum conjecture. Physical Superintelligence (PSI) announced that an internal research system running a non-public Anthropic model produced a complete proof of the Aaronson-Ambainis conjecture, open since 2008. The conjecture addresses when quantum computers can achieve exponential speedups over classical ones on unstructured problems. PSI said the system found a full proof on the third day of a run; the Lean proof assistant accepted it, and the system later strengthened the result twice. Stanford researchers who recently settled a special case called it “a huge result” with major implications for quantum speedups, though independent peer review is still underway. The finding highlights AI’s growing role in advanced mathematics and theoretical computer science.

METR finds AI doubling Anthropic’s progress pace. Safety group METR reported that Anthropic’s model capabilities on well-specified tasks are advancing roughly twice as fast as a year earlier, packing about two years of prior progress into one. METR estimated that a majority—perhaps around 80%—of the current acceleration stems from AI assisting the company’s own research and development, not just more computing power. Staff self-reported large productivity gains with internal models. The assessment, based on high access to internal data, raises questions about compounding speedups and whether Anthropic has crossed internal safety thresholds; Anthropic redacted some evidence.

Thinking Machines ships Inkling 2 Small. Mira Murati’s Thinking Machines Lab released Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters (17 billion active), image understanding, and a long context window. It scores strongly on coding and agent benchmarks and is available now via the company’s Tinker platform. Full weights are promised under an Apache 2.0 license next week after external safety testing. The launch positions the lab as a notable open-weights contender outside major Chinese labs.

Anthropic presses ahead with IPO amid lower valuations. Anthropic is moving forward with plans for a pre-Thanksgiving Nasdaq listing even as prospective investors dial back valuation expectations. Talk has shifted from earlier figures near $2 trillion toward ranges around $1.1–1.5 trillion, according to reports from Bloomberg and others, following market turbulence. The company filed its S-1 earlier and aims to raise a large sum; timing and final pricing remain fluid.

Claude gains read-only bank access; DeepMind probes old RSA. Anthropic launched a Personal Finance beta for U.S. Pro, Max and Ultra users, letting Claude connect to bank and card accounts via Plaid for spending insights and charts. Access is strictly read-only—it cannot move or invest money—and related chats stay out of training. Separately, Google DeepMind said human-guided AI agents found two weaknesses in legacy 1024-bit RSA (signature padding limits and cheaper batch factoring); modern 2048-bit keys are unaffected, according to the company.