Claude ports Bun: AI coding, mobile Codex & new research

Today is a good day for everyone who sees AI not just as a chatbot, but as a tool for real work: migrating code, making models more robust, mobilizing workflows. On top of that, there are a few research results that show where the AI world is still probing its own limits – elegantly phrased, but with plenty of practical relevance.

In short: between “AI rewrites code” and “AI still doesn’t quite understand the world” lies today’s real news. And that news is interesting enough to take a closer look.

🧩 Claude ports Bun from JS/TS to Rust

Probably the most concrete headline of the day: Claude Code has migrated large parts of the Bun codebase from JavaScript/TypeScript to Rust – and apparently not as a lab demo, but as real engineering work completed within just a few weeks. Source: heise

Why does this matter? Because code migration to Rust is usually not a side task. Anyone who has ever moved a larger codebase between language ecosystems knows: types, ownership, concurrency, tests, build tooling – none of that gets done by simply hoping for the best. The fact that an AI tool was used productively here is a strong signal for AI-assisted engineering workflows. It’s no longer just about snippets or small refactorings, but about complex, long-term restructurings. That also means teams will increasingly think about which parts of their legacy or performance-critical software can be modernized faster with AI. And yes, it’s also a small reality check for anyone who still thinks “AI can only do toy stuff” is a valid position.

📱 OpenAI brings Codex to the ChatGPT mobile app

OpenAI is bringing Codex to smartphones: users will soon be able to access the coding tool directly in the ChatGPT app. That means code help, app control, and presumably a much more mobile workflow than before. Source: The Verge and TechCrunch

At first glance, that sounds like a simple product announcement, but strategically it’s quite important. OpenAI is visibly responding to pressure from Claude Code and other coding assistants that currently have a lot of momentum in the developer community. Mobile is more than just “convenient on the go”: if AI coding lands directly on the phone, it changes the way you initiate, review, and delegate tasks. Not everything has to stay tied to the desktop anymore. For teams, that means shorter feedback loops, faster approvals, and more control on the move. For you as a user: the next bug fix might be born with a thumb on the way to the train. Whether that boosts productivity or just makes waiting in meetings more efficient remains to be seen.

📊 New scaling laws for pretraining under data scarcity

A new arXiv paper examines how mixture pretraining behaves when high-quality data is scarce. The core message: if you have little target data – for example, for a low-resource language or a specialized domain – you need to balance the mix of generic and domain-specific data very carefully. Source: arXiv

This is more relevant in practice than the dry title might suggest. Many teams build models for exactly these kinds of niches: medicine, law, industry, internal knowledge bases. In those settings, “more data” is not an option, so data mixing itself becomes model architecture. The study therefore delivers a more important message than just a new curve in a paper: data strategy is model strategy. And because generic data doesn’t automatically help when it dilutes the target signal, the old AI wisdom still holds: quality beats quantity – except when the quantity happens to be made up of carefully curated quality data. Then, of course, the metric you can sell best wins.

🧠 MPINeuralODE: better learning of dynamic systems

With MPINeuralODE, another paper proposes combining Neural ODEs with physics-informed residuals and a multiple-initial-condition curriculum. The goal is more robust generalization to new initial conditions and longer time horizons. Source: arXiv

Why is this exciting? Because many neural differential equation models look good during training but fail in reality outside their known trajectories. That is exactly what this paper addresses. The approach is relevant for anyone modeling time series, simulations, or complex dynamic systems – from robotics to the natural sciences. The real message is: not only the model, but also the learning setup must understand the structure of the problem. Or put differently: if you take the world seriously as a dynamical system, you shouldn’t let the model rely on random data. Papers like this are less “product headline,” but often the building blocks for the next generation of robust AI systems.

🔐 BSI warns about risks in the public charging network

The Federal Office for Information Security has examined the public charging network and identified security problems. In the worst case, even network stability could be affected. Source: heise

This isn’t a classic AI story, but it’s a good example of how digitized infrastructure is becoming increasingly vulnerable. Why include it here? Because thinking about such issues is important for AI systems as well: as energy, mobility, and IT infrastructure converge, security questions quickly become system-critical. For companies, that means AI and IoT deployments must not be considered in isolation. Anyone building smart systems is also building attack surfaces. The market loves efficiency, and so do attackers. The BSI is reminding us that “connected” does not automatically mean “advanced” – it often just means “complex” first.

🚀 Startup Battlefield 200: application deadline is approaching

TechCrunch reminds us that applications for the Startup Battlefield 200 will soon close. The deadline is May 27, with VC access, visibility, and $100,000 in equity-free funding on offer. Source: TechCrunch

For AI startups, this is more than a PR note. In a market where models, tools, and expectations seem to shift every two weeks, a pitch slot like this can determine your first major connections. If you’re building infrastructure, agents, data pipelines, or vertical AI products, this gives you a stage. And yes, visibility in the AI market is almost a second currency alongside capital. If you’re founding right now, it’s worth a look – even if you end up realizing your product is still three iterations away from being demo-ready. That’s still better than no feedback at all.

🛠️ Tool tip of the day

If you’re currently working on code migration, refactoring, or AI-assisted development, it’s worth taking a look at modern coding assistants for large repos and multi-file changes. Especially for Rust migrations or complex production codebases, they can save you a lot of manual grunt work. #

Don’t want to miss any news? Subscribe to the newsletter