7 min read · 13 sources

OpenAI's Model Just Solved a Millennium Problem in 88 Hours

OpenAI says its newest AI technology has solved the Navier - Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, in just 88 hours. The problem asks whether solutions to the Navier - Stokes equations - which govern fluid flow - always exist and remain smooth, or whether they can break down under certain conditions. If the proof holds, it’s a landmark result in mathematical physics, but the bigger story is the speed: what took mathematicians centuries to formulate, this model resolved in under four days.

Engineers should treat this with cautious optimism. The result hasn’t been peer-reviewed yet, and OpenAI’s track record with “breakthrough” claims warrants skepticism. But if verified, it signals that AI can now tackle open problems in pure mathematics, not just pattern recognition. For teams building on fluid dynamics simulations - weather modeling, aerodynamics, even financial risk - this could eventually mean more accurate and efficient solvers, though practical applications are likely years away.

Apple's Foldable iPhone Is Finally Here

Apple is set to unveil its first foldable phone today, after over a decade of secret development. CEO Tim Cook pushed the project forward after seeing Samsung and Huawei’s early foldable devices in Asia around 2020, and the result is reportedly wider than most Android foldables, with pricing discussions as high as $2,000+. This marks Apple’s first major form-factor change since the iPhone X, and it’s a big bet that the foldable market is ready for a premium entry.

For engineers, the interesting part is the hardware. Foldables have historically struggled with hinge durability and screen creasing, and Apple’s reputation for polish means they’ve likely solved these issues - but at what cost? The $2,000+ price point puts it in luxury territory, and the wider display suggests a focus on multitasking. If Apple gets this right, it could legitimize foldables for the mainstream; if not, it’s an expensive lesson. Expect teardowns and stress tests within days of release.

Meta's Muse: An AI Agent That Does Your Clicking for You

Meta launched Muse, an AI agent that acts as a personal digital assistant, autonomously using software apps and websites on your behalf. You can instruct it to send emails, book travel reservations, make purchases, and more, through an app or WhatsApp. Muse can connect to Meta’s own apps to learn about you, and it also links to third-party services like Spotify, Ticketmaster, Shopify, Gmail, and OpenTable.

This is a meaningful step beyond chatbots. Muse isn’t just generating text; it’s taking actions in the real world, which means it needs permissions, API access, and the ability to handle edge cases like payment confirmations or canceled bookings. For engineers, the key question is reliability: an agent that books the wrong flight or sends an email to the wrong person is worse than no agent. Meta’s approach of starting with a limited set of partners suggests they’re aware of this, but the real test will be how it handles unexpected inputs. If Muse works as advertised, it could set the standard for agentic AI - and put pressure on Apple and Google to ship their own versions.

Google DeepMind's AlphaGenome Atlas: All 9 Billion DNA Variants, One Click Away

Google DeepMind released the AlphaGenome Atlas on September 8, a public repository with precomputed predictions for all 9 billion possible single-letter DNA variants in the human genome. The tool, based on the AlphaGenome model announced in January, provides a single-number impact score for each variant, showing at a glance whether it’s likely to be harmful or benign. This removes the need for researchers to run the computationally demanding model themselves.

For bioinformatics and genomics teams, this is a massive time-saver. Previously, getting variant impact predictions meant setting up GPU clusters and running AlphaGenome locally - a barrier for many labs. Now it’s a simple lookup. The impact score isn’t a clinical diagnosis, but it’s a powerful filter for prioritizing which variants to investigate further. Expect this to accelerate research into rare diseases and cancer genomics, and to become a standard reference in the field. The fact that it’s open access is a win for scientific reproducibility.

The CPU Shortage Is Real, and It's Coming for Your Infrastructure

The tech industry is facing a CPU shortage, with server order lead times stretching to ~6 months, up from the usual 1-2 weeks, and prices up 10-20% since March. The culprit: AI agent workloads. These systems consume CPU for orchestration, sandboxing, and tool calls - not just GPU compute. Intel noted that AI datacenters now use 1 CPU per 4 GPUs, a ratio that could rise to 1:1 as agents become more complex.

Most software teams have never planned for CPU scarcity - it’s always been the cheap, abundant resource. That’s changing. If you’re running agent-based systems, you need to audit your CPU-to-GPU ratio now, because lead times of six months mean you can’t react quickly. Options include optimizing orchestration code, consolidating workloads, or reserving capacity early. The price increase is modest so far, but if demand keeps climbing, it could squeeze budgets across the board. Treat CPU as a constrained resource in your capacity planning, not an afterthought.

The Math Problem Is That We Don't Know Which Problems Matter

Terence Tao, one of the world’s leading mathematicians, published a post arguing that the “collection of good, fruitful open problems is now being mined in a non-renewable fashion.” His point: working out whether a question is worth working on is a lengthy, deliberate, and subjective process. While there are infinite possible problems in mathematics, the vast majority aren’t worth solving. The indiscriminate use of powerful solution-extraction tools - like the AI that just solved Navier - Stokes - achieves the short-term goal of solving problems but risks exhausting the supply of meaningful ones.

This is a warning for anyone using AI in research. The tools are great at finding answers, but they’re terrible at knowing which questions matter. If you let AI drive your research agenda, you’ll get a pile of solved trivia, not breakthroughs. The bottleneck is shifting from computation to curation: deciding what’s worth working on. For engineers, this means your judgment about what to build is more valuable than ever, even as AI automates the execution. Don’t outsource your problem selection to the model.

Key App Developers Are Sitting Out Apple's Siri AI

Apple’s new Siri AI - which lets the assistant perform actions inside third-party apps - is facing adoption problems. Several staple apps are still not working with the early version available to developers, and it’s unclear whether they will. The stakes are high: these apps could become more useful on Apple devices by integrating with Siri, but enabling it risks sidelining their own assistants and giving Apple more control over the user relationship.

If Siri AI rolls out this year with only support from some apps and not others, it could create a confusing experience for users - one where Siri works in Uber but not Lyft, or in Spotify but not Apple Music. That inconsistency could turn users off from the assistant entirely. For developers, this is a classic platform dependency dilemma: integrate early and risk losing control, or wait and risk being left behind. Apple’s history suggests they’ll push hard, but the outcome depends on whether they can convince developers that Siri AI drives enough new usage to justify the trade-off.

Quick Links

Source: macrumors.com ↗

• Apple acquires Sonera, a startup working on non-invasive brain and body magnetic field sensing tech - The acquisition signals Apple’s investment in novel biosensing hardware, which could lead to new health and accessibility features in future devices.

• A fly’s brain map trained an AI to play Mario 64 - Google’s fly connectome, with over 166,000 neurons and 125 million connections, was used to train an AI that can play the game, demonstrating the feasibility of using full neural maps for real-time control tasks.

• Mechanistic interpretability research has hit a wall - Early success in reverse-engineering simple AI models hasn’t scaled to real language models; different teams’ maps disagree, and strategies that worked on toy AIs don’t transfer. Don’t rely on interpretability for safety or debugging yet.

• OpenAI’s ChatGPT Images 2.5 is out - Image generation latency is down by up to 50% compared with Images 2.0, making the tool more practical for iterative workflows.