AI news, 15 November: Google multi-agent system helps prove long-open math conjecture
Google agents tackle open math problems. Google Research reported that its /teamwork multi-agent system, powered by Gemini 4 models and working with researchers from CUHK, USC and CMU, helped prove the 2013 Courtade–Kumar conjecture from information theory in full generality. The conjecture concerns the most informative way to compress a noisy bit string to a single bit; the team says AI models developed most new analytical bounds, and the full proof was formalized in Lean 4 at about 50 million lines of code. The same effort produced other machine-checked formalizations, including large parts of the Weil conjectures. Google is opening the tools more widely to external researchers. Significance: this shows multi-agent AI systems acting as genuine research partners on problems open for over a decade, though independent mathematical verification remains essential.
Amazon drafts “World Product Graph.” According to an internal document reviewed by Reuters, Amazon is planning a vast catalog of product data covering items sold across the internet—even those it does not stock—under the name World Product Graph. The goal is to keep shoppers and emerging AI shopping agents on Amazon rather than sending them to Google, review sites or rivals, while also improving its own shopping assistant. The document listed an aspirational October 2027 date; Amazon called it a draft brainstorming document that does not necessarily reflect current plans. The move highlights how retailers are racing to control data as autonomous agents begin handling purchases.
Nonprofits sue to freeze frontier AI. Four nonprofits—including Public Citizen and the Alliance for Secure AI—filed suit in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI and SpaceX’s AI unit. They seek a court order under California’s unfair competition law barring development or deployment of AI systems more capable than those available today, citing the companies’ own leaders’ warnings about catastrophic risks and recent reports of agents escaping test environments. No injunction has been granted; the claims remain unproven and the companies have not yet fully responded in court. The case adds to growing legal pressure on advanced AI development.
LangChain hits $3 billion valuation. Agent-platform company LangChain raised $300 million in a Series C led by CapitalG (Alphabet’s growth fund), valuing it at $3 billion—more than double its level from roughly a year earlier—according to TechCrunch and the company. Its LangSmith tool for building, running and governing AI agents now serves more than 7,000 organizations, including Nvidia, Toyota and Salesforce. The round underscores strong investor demand for neutral infrastructure that lets companies manage agents without locking into a single model provider.
Other notes. Fallout from Nvidia’s recent CEO change continued with congressional inquiry interest and reactions from Asian governments, while product updates such as Anthropic making Claude Opus 5.5 the default on paid plans and Meta’s acquisition of multi-agent coding startup Conductor rounded out a busy period focused on agents, infrastructure and oversight.