🔬 OpenAI Releases Hundreds of New AI Math Results

OpenAI has published 722 math papers from an AI model the public still cannot use. It’s the same model behind last month’s reported solution to a problem that had been unresolved for roughly 90 years, now producing results across several branches of mathematics. Researchers still need to establish how much holds up, but the potential is hard to ignore for anyone who has spent years trying to answer one difficult question. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: OpenAI has shared hundreds of mathematical results from a model it has not yet released, including proposed solutions to longstanding unsolved problems.

Key Points:

OpenAI’s new results cover several branches of mathematics, including number theory, geometry and mathematical physics. They come from the model behind its reported September solution to the Navier–Stokes Millennium Prize Problem, a question about fluid motion that had been unresolved for roughly 90 years. The collection includes a claimed advance on how closely fractions can approximate pi, along with 10 summaries explaining how the model approached selected problems. OpenAI will fund workshops to help mathematicians understand the work, though it acknowledges that some results without formal verification could contain errors.

OpenAI’s latest work suggests AI could produce new mathematics faster than people can check and understand it. If the output keeps growing, there may not be enough mathematicians with the right expertise to keep up. Computers can check whether a formal proof follows the rules, but researchers still need to work out what it adds and how others can use it. Universities and labs could end up with more promising findings than their experts have time to examine, leaving potentially useful ideas waiting for someone who can make sense of them.

Enterprise buyers won’t put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.

Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.

Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.

No compliance hire. No consultant. Your auditor signs off.

Learn more and get $1,000 off.

Evolving AI: Claude for Google Workspace can now edit your files directly without making you switch between apps.

Anthropic’s Google Workspace add-on puts Claude in a sidebar next to your file so you can select the content you want it to work on. You can connect other apps to bring in customer records and call notes when preparing something like a review deck. Teams can also save their instructions and preferred format as a skill to reuse for the next deck, with the beta available on all paid Claude plans.

Claude’s move into Google Workspace builds on its Microsoft 365 integrations and brings Anthropic into both major office suites where people already spend their working day. Google already lets developers choose rival models in Antigravity and OpenAI’s Codex supports alternative providers, giving users a similar choice within their coding tools. These arrangements could help companies keep people using their software even when a rival offers a more capable or cheaper model. Anthropic gets to compete for that work without asking businesses to replace the apps their teams depend on. For you that means more choice over which assistant helps with the job without having to move your files or persuade colleagues to use another tool.

Evolving AI: Figure founder Brett Adcock has launched Hark Pro as a personal AI assistant that remembers your preferences and handles everyday tasks.

Hark wants the next generation of computers to understand your routine well enough to handle tasks before you think to ask. Its launch examples include booking movie tickets and then offering to reserve nearby parking, while Action Buttons let you approve tasks such as paying a bill with one tap. Hark asks before spending money or sending messages unless you grant standing permission. The software also forms the foundation for its first hardware product, which is planned for 2027.

Hark joins Meta’s Muse and OpenAI’s Dots in a race to make personal AI part of daily life. An assistant that remembers your preferences and reliably handles follow-ups could give you a reason to stay with one platform, but that convenience requires access to personal information. Meta emphasized privacy when launching Muse, yet a user reported that it shared his home address with a buyer after he misunderstood an “Allow Always” setting. Hark and its competitors will need to prove they can reduce your workload while keeping you in control of what their agents share or do in your name.

QUICK HITS

🇫🇷 Mistral launched Mistral Large 4, a 1T-parameter open-weight model with 49B active parameters and a public API preview.

🔐 Anthropic expanded its cyber access program, giving vetted security teams Claude models with fewer cybersecurity restrictions.

💰 DeepSeek is reportedly raising nearly $12B at a roughly $74B valuation as it prepares for a possible Chinese IPO.

🛡️ CrowdStrike, AWS and Nvidia opened a global AI-security accelerator for startups building cybersecurity products around autonomous agents.

 *partner link