OpenAI's latest model just solved more than 100 unsolved math problems
OpenAI says an internal model resolved over 100 long-standing open mathematical problems, some in as little as 24 days, prompting the company to form an independent Advisory Group on Mathematics and AI to vet the results. The announcement adds to mounting evidence that frontier models are closing in on research-level mathematical reasoning. OpenAI has not yet detailed which specific problems were solved or opened the work to full peer review. | Read at TechCrunch →
AMD nears $1 trillion market cap as Meta's AI news lifts chipmakers
AMD and Intel shares surged after news tied to Meta's Muse AI agent reignited investor enthusiasm for AI compute demand, pushing AMD's market value to the brink of $1 trillion. The rally reflects how closely chipmaker valuations now track any signal of accelerating AI adoption, even from consumer-facing products rather than data center deals. Analysts note the swings underscore how sensitive the sector remains to sentiment. | Read at Bloomberg →
Alibaba plans a 10-trillion-parameter model and unveils its own AI chip
Alibaba announced plans to build an AI model with 5 to 10 trillion parameters, roughly on par with the largest frontier systems, while also unveiling the Zhenwu V900, a new in-house AI accelerator chip. The moves signal Alibaba's push toward self-sufficiency in both model scale and the silicon needed to train and run it. The company also outlined a substantial cloud capacity expansion to support the buildout. | Read at US News →
Amazon locks out Meta's Muse AI shopper while Shopify rolls out the welcome mat
Amazon cut off Meta's Muse AI agent from shopping on its site, saying Meta never disclosed or sought permission for the agent's access and that Muse's credential handling posed security risks. Shopify took the opposite approach, partnering with Meta to let Muse complete purchases through Shop Pay across its merchant stores. Mark Zuckerberg said on X that "more partnerships like this" are coming, highlighting a widening split over how retailers want to handle AI shopping agents. | Read at GeekWire →
OpenAI pushes for US-led global rules on frontier AI and self-improvement
OpenAI called for the United States to lead development of international technical standards governing frontier AI systems, including models capable of recursive self-improvement. The proposal comes as automated AI research accelerates and follows scrutiny of a recent OpenAI-linked agent cyberattack incident involving Hugging Face. The company argues coordinated standards are needed before more autonomous systems are deployed at scale. | Read at CNBC →
Security researcher finds a Meta Muse flaw that lets malware hijack the AI
Researcher Patrick Wardle of Objective-See uncovered a zero-day vulnerability in Meta's Muse AI assistant on Mac that lets local malware hijack the agent, intercept dictated prompts, and steal authentication credentials by exploiting an undocumented, unprivileged setting. Wardle published a proof-of-concept exploit showing the compromised agent could inherit access to files, email, and calendars. Meta had not publicly addressed the specific findings at time of publication, though it offers bounties up to $300,000 for Muse security flaws. | Read at Cybersecurity News →
Xiaomi's new open-weight model ties Grok 4.7 atop the leaderboard, for $3M
Xiaomi released MiMo-V2.6-Pro, an open-weight model that now ranks as the top-performing open-source system in the world, matching xAI's Grok 4.7 on benchmarks despite reportedly costing only about $3 million to train. The company also released a cheaper V2.6-Flash variant alongside a large reinforcement-learning environment stack. The launch underscores how quickly Chinese labs are closing the performance gap with far larger training budgets from US rivals. | Read at VentureBeat →
Salesforce and ServiceNow both cross major AI revenue milestones this quarter
Salesforce reported $3.4 billion in combined annualized recurring revenue from its Agentforce and Data 360 products, while ServiceNow surpassed $1 billion in AI annual contract value, with agentic deployments up ninefold in nine months. The two are chasing enterprise AI dollars differently: Salesforce embeds agents into existing CRM workflows, while ServiceNow sells an "AI Control Tower" governance layer across models and workflows. Both companies still face pressure to prove deployed agents translate into durable, renewing revenue. | Read at Superpower Daily →
Nvidia's Jensen Huang says there's "0% chance" AI ends the world by 2030
Nvidia CEO Jensen Huang dismissed AI extinction warnings as "doomsday narratives," arguing no one currently has the ability to build a superintelligent system capable of threatening humanity. He said prominent doomsayers "must be doing it for ulterior reasons" and that the rhetoric has "done a lot of damage" to public understanding of AI. The comments put Nvidia's leadership at odds with safety-focused voices at rival labs like Anthropic. | Read at Bloomberg →
Anthropic says Claude now leads a quarter of the work building its successor
Anthropic disclosed that Claude led 26% of the company's model research and development as of August, up from zero in February, often completing tasks "end-to-end from a high-level prompt" under human supervision. About 90% of Anthropic's R&D now involves Claude in some capacity, monitored by roughly 30,000 tracking agents watching for misbehavior. Anthropic said it wants to "minimize the gap between what frontier labs know and what the public knows" about the pace of self-improving AI. | Read at Fast Company →
OpenAI and Anthropic reportedly discussed a pact to stress-test each other's models
OpenAI and Anthropic were in talks over a legal agreement to red-team and stress-test each other's AI systems for vulnerabilities, according to a report, discussions that gained urgency after a cyberattack incident tied to OpenAI agents and Hugging Face. The proposed arrangement would formalize a practice the labs have done informally before. Neither company has confirmed a final deal, and terms of any agreement remain unclear. | Read at Seeking Alpha →
Google brings AI agent development to Kotlin, no Python required
Google released Agent Development Kit for Kotlin 1.0, a production-ready framework giving Kotlin, Android, and JVM developers the same orchestration, memory, and human-in-the-loop tools previously available only in Google's Python and Java versions. The kit supports both on-device inference through LiteRT-LM and cloud inference via Firebase AI Logic, letting apps switch between them. The release targets enterprise concerns about deploying AI agents safely in production mobile environments. | Read at InfoQ →
Know someone who'd want this? Forward this email their way.
Got a story we should've covered? Reply and let us know.