OpenAI says an unreleased model helped crack the Navier-Stokes problem, one of maths' hardest open questions, but academics are already disputing who deserves the credit. Meta shipped a personal AI agent that runs errands in the background, and a wave of AI safety resignations put existential risk back in the headlines. There is fresh movement in image generation, music and European funding too.

Summary: OpenAI says an internal model more advanced than its latest release used 10,000 coordinated agents running for 88 hours to solve a variant of the Navier-Stokes equations, a century-old open problem in fluid dynamics. The agents produced millions of messages and 130 billion tokens, and the result was formally verified in Lean. But NYU's Tristan Buckmaster and Anthropic's Levent Alpoge, who worked on the same problem for a year and published partial results a day earlier, say OpenAI may have drawn on their unpublished drafts stored in Codex logs. OpenAI denies viewing their work directly while admitting user data may have influenced training.

Why it matters: It signals AI can now attack deep unsolved scientific problems at scale, but also raises hard questions about labs quietly absorbing private user research.

Source: OpenAI

Summary: Meta introduced Muse, an always-on personal agent that connects to email, calendars, shopping and home devices to handle tasks like booking reservations and sending emails. It integrates with apps such as Gmail, Spotify, OpenTable and Ticketmaster, can write custom integrations on the fly, and runs in a secure cloud environment with Stripe payments. Purchases still require user approval. The US rollout offers a free tier of roughly 100 million tokens a week, with paid plans at 20 dollars and 100 dollars a month. Meta promises a confidential virtual machine to keep user data private.

Why it matters: This pushes consumer AI from answering questions to taking real actions, but it hands a company with a mixed privacy record deep access to personal accounts.

Source: Meta

Summary: Evan Hubinger, who leads alignment science at Anthropic, said there is more than a 10 percent chance advanced AI could kill all humans within a decade if left unchecked. His comments followed the resignation of researcher Jacob Coxon, who worked on pretraining at both OpenAI and Anthropic and warned the labs are gambling with human lives by racing toward self-improving models. Both point to the danger of recursive self-improvement, where models accelerate their own development. Anthropic's own August safety report admits it lacks a clear plan to align superintelligent AI. UK officials have called for an international treaty.

Why it matters: Senior insiders publicly warning about catastrophic risk adds weight to calls for enforced safety standards and could shape regulation.

Summary: A developer built a site over a long weekend that lets Claude users connect by voice or video with other people who are also waiting for responses. Users can vent about slow response times, share what they are building, or make new friends. It captures how people are now building social experiences around AI's idle moments.

Why it's interesting: It shows how deeply AI tools have woven into daily habits, enough that waiting for them becomes its own social scene.

Source: GitHub

Summary: Alongside a faster image model, OpenAI added a Sketch feature that lets users draw directly inside ChatGPT and turn rough doodles into polished pictures using an @ command. It shifts image generation from typing careful prompts to hands-on visual input, with templates and shareable prompts to speed up editing.

Why it's interesting: It moves AI creation away from wordy prompting toward drawing, which is more intuitive for many people.

Summary: Suno released new v6 music models trained on licensed data through partnerships with Warner Music Group, BMG and Believe, sidestepping earlier legal disputes. The lineup includes two paid models and a free mini version, plus fan remixes, and can take images, video or audio as input. Suno is betting monetisable fan-created content can convince major artists to opt in, even as other lawsuits continue.

Why it's interesting: It marks a shift from AI music built on scraped audio toward licensed deals with the industry it once fought.

Context: For finance, operations or budgeting teams that rely on spreadsheets and need to catch costly errors before decisions are made.

Prompt: Act as a specialised spreadsheet auditor for finance and operations. Review the attached Excel or Google Sheet and detect broken formulas, unexpected hard-coded values, inconsistent references and data formatting issues. Return a table with columns for sheet and cell, the issue, why it matters, severity, and a recommended fix. Confirm that all assumptions and references are consistent and that totals reconcile correctly.

Example use case: A finance lead runs a quarterly forecast through it before a board meeting and finds a hard-coded value that had quietly broken a revenue total.

Summary: DeepMind launched AlphaGenome Atlas, an AI resource that predicts the effect of every possible single-letter change in human DNA, some nine billion variants. It is free for non-commercial research and includes a score to flag potentially significant variants. DeepMind stresses the predictions must be confirmed by lab testing.

Why it is important: It could speed up research into rare diseases and poorly understood regions of the genome.

Summary: French startup Mistral raised 3 billion euros in a Series D at a valuation above 21 billion euros, the largest equity round in European tech history. Its model centres on AI tools enterprises can run themselves for data control, with plans for a one-gigawatt data centre network across Europe by 2030. It already counts Airbus and HSBC among more than 125 enterprise clients.

Why it is important: It signals strong demand for AI that avoids cloud lock-in and keeps data on home turf.

Summary: At its biggest event under new leadership, Apple pushed AI into its whole product range. Siri AI enters beta on iOS 27, and the iPhone 18 Pro and Pro Max run a new 2nm A20 Pro chip with 32 neural cores for on-device tasks. A foldable iPhone arriving late October reportedly uses AI in hinge manufacturing and adds image provenance features to verify authenticity.

Why it is important: Apple embedding AI at every layer brings the technology to a vast mainstream user base.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas