Relevance 10/10Importance 10/10
OpenAI announced Monday that a model it began training on August 28 has resolved more than 100 long-standing open problems in mathematics in under a month, including a claimed solution to Navier-Stokes existence and smoothness, one of the seven Millennium Prize Problems. After backlash from hundreds of mathematicians, the company stood up an independent Advisory Group on Mathematics and AI at the Institute for Advanced Study in Princeton, with members including Timothy Gowers, Martin Hairer, Edward Witten and Ravi Vakil. The catch: OpenAI has not released the list of problems, and the group explicitly will not advise on pacing.
Relevance 10/10Importance 9/10
Xiaomi published MiMo-V2.6-Pro, a 1.02-trillion-parameter sparse MoE, and the cheaper 309B Flash variant to Hugging Face under an MIT license, along with a 9B distill, the technical report, the RL training framework, and over 7,000 RL environments. Pro scores 46 on the Artificial Analysis Intelligence Index — the highest of any open-weight model, matching Grok 4.7 — with roughly $2.6M to $3.5M in RL compute. Shadowing the launch: Anthropic's report naming Xiaomi among seven Chinese labs it accuses of industrial-scale Claude distillation, with 400,000+ replayed exchanges observed in March and April.
Relevance 10/10Importance 8/10
Grok 4.7 landed Monday at $2 per million input and $6 per million output tokens — unchanged from 4.6 — and went live same-day in Cursor, GitHub Copilot, the xAI API and the major routers. It posts 71% on DeepSWE v1.1, edging past Claude Fable 5.1 Max, and 64% on EEBench. But on Terminal-Bench 4.0 it manages just 26% against 60% for GPT-6 Astra, and its overall index score of 46 sits well behind the 53 posted by GPT-6 and Claude Fable 5.1.
Relevance 10/10Importance 8/10
Anthropic's head of life sciences confirmed that "real lab work" is happening today at a quietly built biology facility in the San Francisco Bay Area, where Claude controls microscopes, liquid handlers and robotic arms via a new Model Hardware Standard protocol. It pairs with Claude Science, a research workbench wired into 60-plus scientific databases. The company says the lab is not drug-discovery-specific and that humans stay in the loop — a notable line from the lab that has been loudest about AI bioweapon risk.
Relevance 9/10Importance 8/10
Researchers at AIR Security disclosed the first true supply-chain vulnerability in the AI agent ecosystem, affecting Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI. The flaw defeats SHA pinning — an attacker names a git branch after the pinned 40-hex commit hash and the agent checks it out silently, with no user interaction. Anthropic patched in Claude Code 2.1.179 and OpenAI in Codex 0.146.0; reporting puts roughly 134,000 agents and 925 active skills in the blast radius, with Copilot still exposed.
Relevance 9/10Importance 8/10
Amazon began blocking Meta's Muse assistant from its retail site Sunday night after Meta declined a request to pull the bot. Amazon says Muse doesn't identify itself while browsing and appears to capture and store customer credentials, and that it never authorized the access. This follows Amazon's suit against Perplexity over Comet and moves against Google and OpenAI shopping agents — awkward, given Meta signed a multibillion-dollar deal in April to run agentic workloads on Amazon's Graviton chips.
Relevance 9/10Importance 8/10
The UN's Independent International Scientific Panel on AI issued a thematic brief warning that safeguards for increasingly capable agents are not advancing as fast as the agents themselves. The panel cites the combination of misaligned goals, rising capability, and permissive deployment environments, and invokes the precautionary principle — urging governments to act before agent risks are fully characterized.
Relevance 8/10Importance 9/10
SoftBank is marketing over $11 billion across dollar and euro tranches, pricing September 24 and settling September 29 — which would be the largest non-financial corporate bond deal ever out of Asia Pacific. Proceeds partly fund the third tranche of its OpenAI investment, taking cumulative exposure to roughly $64.6 billion for about 13% of the company. SoftBank has issued nearly $15 billion in notes this year, more than any other speculative-grade borrower globally.
Relevance 9/10Importance 7/10
DAMO Academy released weights, code and training framework for a vision-language model trained on 420,000+ contrast-enhanced abdominal CT exams and 15 million anatomy-focused image-text pairs. Across roughly 40,000 real-world exams it averaged 0.913 AUC over 146 clinical findings and outperformed 23 of 26 expert radiologists in a head-to-head study published in Science. With AI prompts, physician sensitivity rose about 10% and reading time dropped more than 30%.
Relevance 8/10Importance 8/10
The two labels filed a 45-page complaint in federal court in Massachusetts alleging Suno's v6 still infringes 60,202 recordings. Their theory: v6 was trained on user interactions with Suno's earlier allegedly-infringing models, making it "the fruit of the same poisoned tree." Warner, BMG and Believe licensed content for v6; Sony and UMG did not, and are seeking statutory damages and fees. Suno calls the claims "fundamentally flawed on both the facts and the law."
