OpenAI's AI solves 100+ open math problems

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🧮 OpenAI's AI solves 100+ open math problems

🎵 Apple unveils its first live music venue

🏆 Xiaomi's new AI tops all open models

🔓 Meta patches Muse exploit that let attackers control the AI agent

🤖 SpaceXAI unveils Grok 4.7

🇨🇳 Alibaba introduces China's most powerful AI chip

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

🧮 OpenAI's AI solves 100+ open math problems LINK

OpenAI says the internal model behind its disputed Navier-Stokes work has also produced solutions to more than 100 other previously unsolved math problems across many fields, prompting debate among mathematicians over how such claims should be judged.

To address that scrutiny, OpenAI is setting up an independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study in Princeton, with nine mathematicians advising on how results are assessed and announced.

The unpaid panel cannot steer OpenAI's research pace or halt projects; it only offers advice, and one founding member, Camillo De Lellis, recently signed a letter warning that competition between AI companies pressures the math community.

🎵 Apple unveils its first live music venue LINK

Apple has opened Apple Music Hall, a live music venue holding 600 people inside London's Battersea Power Station, which also houses the company's UK headquarters and its Apple Battersea store.

The space pairs a 38-foot configurable stage with a 48-speaker spatial sound system, two recording studios, and at least 16 cameras, letting a single show produce a broadcast-ready mix, Spatial Audio recording, and multicamera video.

The stage can face the audience or be set up in the round, and Apple says the venue continues a strategy that began with the iTunes Festival in 2007, later renamed Apple Music Festival, and Apple Music Live.

🏆 Xiaomi's new AI tops all open models LINK

Xiaomi's new MiMo-V2.6-Pro has become the strongest openly available AI model, scoring 46 on Artificial Analysis's Intelligence Index while charging a fraction of what similarly capable rivals like Kimi K3 and Qwen cost per task.

The model runs on a mixture-of-experts design with 1.02 trillion parameters, only 42 billion active per request, and costs about $0.44 per million input tokens and $0.87 per million output, or roughly $0.13 per test task.

Xiaomi credits the jump to expanded reinforcement learning that took under six days and cost about $2.62 million, though Anthropic accused Xiaomi of feeding user chats into Claude to pull out training data.

🔓 Meta patches Muse exploit that let attackers control the AI agent LINK

Meta has fixed a zero-day flaw in its Muse app for macOS that let attackers hijack the AI agent's account, rushing out a patch within hours of the bug being reported publicly.

Security researcher Patrick Wardle found the bug used a hidden Muse setting that let local attackers reroute cloud dictation to their own server, then abuse the agent's access to take photos and write harmful files, often without warning the user.

Meta's David Singleton downplayed the danger, calling it a local attack that needed malicious code already running on the person's machine, so the real risk was low, though Wardle criticized the company for not building in security from the start.

🤖 SpaceXAI unveils Grok 4.7 LINK

SpaceXAI has released Grok 4.7, which it calls its strongest model for coding and knowledge work, now available through Cursor, Grok Build, the Grok API and outside platforms.

Pricing starts at $2 per million input tokens and $6 per million output tokens, with a version running twice as fast for double the cost, matching Grok 4.6 on price while using a bigger base model.

Trained on harder, hours-long tasks, Grok 4.7 verifies its own work and, on safety testing, scored 62.4% on LatchBio's biosafety benchmark and let through 3.3% of risky dual-use prompts on HackerBench.

🇨🇳 Alibaba introduces China's most powerful AI chip LINK

Alibaba has revealed the Zhenwu V900 accelerator, a chip it calls China's most powerful for AI, said to be three times faster than the last version and able to run in clusters of up to 500,000 units.

CEO Eddie Wu announced the chip on Tuesday, saying the V900 is built for training frontier models, and the company plans to build a system with 5-10 trillion parameters, far bigger than current open-weight rivals.

Alibaba's chip unit T-Head, which it plans to list publicly, has shipped 560,000 units of the Zhenwu family to over 400 outside customers, part of a wider AI push backed by more than $53 billion over three years.

Other news you might like

• DoorDash agrees to pay $131.5M to settle New York City probe into delivery worker payLINK

• OpenAI, Anthropic held talks to ‘stress-test’ each other’s AI models: reportLINK

• China's Z.ai disables AI coding assistant features after security issueLINK

• X now tells users when local laws have reduced their posts' visibilityLINK

• EU to require data centres to disclose energy and water efficiencyLINK

• OpenAI faces lawsuit from British Columbia over Tumbler Ridge shootingLINK

• AMD just joined the $1 trillion club, beating Intel to the milestoneLINK

• More Xbox Layoffs Hit as Massive Reorg Continues; Activision to Make Next ‘Halo’ GameLINK

• Internet Explorer is finally, truly dying, and Microsoft wants you ready by 2029LINK

• The Business of Building GodLINK

• Engineer says Claude Code has made his job "soul-sucking" as workers spend 12-hour days pressing enterLINK

• California joins US states clamping down on data center gold rushLINK

• DeepSeek to brief UN Security Council on AI risks this week, Reuters reportsLINK

🧰 Trending tools

Pulsetic RUM: monitors websites, APIs, and services for uptime, performance, and SSL, with multi-channel alerts, status pages, and incident management to keep teams informed.LINK

Fide Island: transforms your MacBook notch into a command surface for media, calendar, files, clipboard, and quick calculations without breaking focus.LINK

Embedful: build and embed secure customer-facing analytics dashboards into your SaaS product, connecting existing data without maintaining your own dashboard infrastructure.LINK

Project Feed: a collaborative workspace where game teams share updates, review media, manage tasks, and collect client feedback, free for up to three editors.LINK

PeakHour 6: monitors your Mac's network in real time, using on-device AI to grade connection quality with Wi-Fi diagnostics, speed tests, and outage alerts.LINK

Brev: connects goals, meetings, and metrics across your existing tools so leaders can track execution, spot problems early, and coordinate through Slack or Teams.LINK

📚 Trending papers & reports

Chatbot safety defense makes a model forget the specific harmful knowledge attackers try to unlock, sharply cutting the success rate of jailbreak attempts without dumbing down the model on normal tasks.LINK

Fathom-Vaidya is a medical AI trained to both diagnose conditions and handle real patient conversations, with its 30-billion-parameter version scoring ~50% on a hard clinical test, beating GPT-5.LINK

Author-identification tools only reliably spot who wrote a text when their setup separates a person's true writing style from the topic and format, meaning off-the-shelf detectors often fail the moment the subject changes.LINK

UK legal search gets a new testing ground that judges whether tools can pinpoint the exact relevant paragraphs in past rulings, not whole documents, and today's best systems still fall short.LINK

Pathology slide reading gets an AI assistant that trims gigapixel tissue scans to their most useful regions before answering, hitting ~80% accuracy on cancer-image questions while cutting memory and speeding up responses.LINK

See you tomorrow for a new dose of ☕️ Techpresso!