Published on May 03, 2026 at 18:00 CEST (UTC+2)
Mercedes-Benz commits to bringing back physical buttons (105 points by teleforce)
Mercedes-Benz commits to bringing back physical buttons
Mercedes-Benz has announced it will revert to using physical buttons for frequently used controls in its upcoming vehicle interiors, acknowledging that touch-sensitive controls were a mistake. This decision follows customer feedback and aligns with a broader industry trend of manufacturers reversing the shift to all-touch interfaces. The article highlights the importance of user experience and ergonomics over minimalistic design trends.
Alert-Driven Monitoring (25 points by khazit)
Alert-Driven Monitoring
The article argues that infrastructure monitoring should prioritize alerts over dashboards, as dashboards often become "cool office art" that nobody watches. It advocates starting from first principles—identifying what behaviors indicate service failure—rather than from existing metrics. This approach aims to build trustworthy alerting systems that reduce noise and improve operational reliability.
Embedded Rust or C Firmware? Lessons from an Industrial Microcontroller Use Case (106 points by mrtz)
Embedded Rust or C Firmware? Lessons from an Industrial Microcontroller Use Case
This academic paper compares two teams concurrently developing the same IoT functionality—one in C and one in Rust—over several months. The findings show no significant advantage for C over Rust in memory footprint or execution speed on microcontroller hardware. The Rust-based Ariel OS demonstrated an efficient runtime with a smaller footprint than the traditional bare-metal C stack, suggesting Rust is ready for embedded systems.
Automating Hermitage to see how transactions differ in MySQL and MariaDB (13 points by zdw)
Automating Hermitage to see how transactions differ in MySQL and MariaDB
The author explores transaction isolation levels and anomalies like Dirty Writes and Dirty Reads, noting ambiguities in the SQL standard. They introduce a new open-source tool called Monastery that automates tests across multiple databases and isolation levels. The article uses a bowling alley thought experiment to illustrate transactional integrity issues.
Show HN: Apple's Sharp Running in the Browser via ONNX Runtime Web (110 points by bring-shrubbery)
Show HN: Apple's Sharp Running in the Browser via ONNX Runtime Web
This GitHub project enables users to upload a single image and generate Gaussian splats entirely in the browser using Apple's ml-sharp model via ONNX Runtime Web. The tool allows previewing results and downloading a .ply file, demonstrating the feasibility of running complex 3D machine learning models directly in the browser without server-side computation.
What Is Z-Angle Memory and Why Is Intel Developing It? (10 points by rbanffy)
What Is Z-Angle Memory and Why Is Intel Developing It?
While the full article is unavailable, the title indicates Intel is developing a new memory technology called Z-Angle Memory. This likely targets improvements in memory bandwidth, latency, or density, which could have implications for AI/ML workloads that are memory-intensive.
Security Through Obscurity Is Not Bad (9 points by mobeigi)
Security Through Obscurity Is Not Bad
The author challenges the common mantra that "security through obscurity is bad," arguing that obscurity is beneficial as an additional layer of defense. They differentiate between relying solely on obscurity (weak) and using it alongside other security measures. The article also refutes the claim that AI makes all obscurity trivial to bypass.
Group averages obscure how an individual's brain controls behavior: study (75 points by hhs)
Group averages obscure how an individual's brain controls behavior: study
Stanford Medicine researchers found that averaging brain scans across groups hides individual differences, especially in children struggling with goal-oriented tasks. Analyzing data per-person reveals distinct brain activity patterns that group averages miss, with implications for understanding conditions like ADHD. The study emphasizes the need for personalized neuroscience approaches.
A couple million lines of Haskell: Production engineering at Mercury (344 points by unignorant)
A couple million lines of Haskell: Production engineering at Mercury
This blog post describes the experience of building and maintaining a large-scale Haskell codebase at the fintech company Mercury. The author reflects on the language's promise of eliminating null pointer exceptions and how that holds up at scale. It covers practical production engineering challenges and the benefits of Haskell's type system in a real-world financial application.
Porsche will contest Laguna Seca in historic colors of the Apple Computer livery (8 points by Amorymeltzer)
Porsche will contest Laguna Seca in historic colors of the Apple Computer livery
Porsche Penske Motorsport will race two Porsche 963s at Laguna Seca wearing a livery inspired by the 1980 Porsche 935 that originally carried the Apple logo. The design celebrates the 75th anniversary of Porsche Motorsport and the 50th anniversary of Apple. The article highlights the longstanding relationship between the two companies.
Cross-platform AI inference goes mainstream
The project running Apple's SHARP model in the browser via ONNX Runtime Web shows that sophisticated 3D machine learning tasks (Gaussian splatting) can now be executed client-side. This trend democratizes AI, reduces server costs, and enables privacy-preserving applications. Developers should invest in ONNX and WebGPU-based inference pipelines for edge deployment.
Functional programming languages gain traction in AI production systems
Haskell's use at Mercury (a fintech company) for a 2-million-line codebase demonstrates that strongly-typed, functional languages are viable for critical AI-adjacent infrastructure. Type safety and immutability reduce bugs in data-heavy applications. AI/ML teams should consider languages like Haskell or Rust for backend systems where correctness and reliability are paramount.
Personalized AI models over population-averaged approaches
The Stanford study on individual brain scans warns that group averages obscure meaningful patterns. This parallels a growing trend in AI: moving from one-size-fits-all models to personalized or fine-tuned models (e.g., for healthcare, recommendation systems). Researchers and practitioners should prioritize per-subject or per-user modeling to capture individual variability.
Security through obscurity as a viable AI defense layer
The debate around obscurity in security is relevant to AI/ML adversarial robustness. While pure obscurity (e.g., hiding model architectures) is fragile, combining it with traditional defenses (e.g., adversarial training) can raise the cost for attackers. AI developers should consider techniques like model obfuscation or input randomization as part of a layered security strategy.
Rust emerges as a strong alternative to C for embedded AI/ML firmware
The comparative study shows Rust matches C in performance and memory footprint on microcontrollers, while offering memory safety. This is critical for AI at the edge, where firmware bugs can be costly. Rust’s ecosystem, including Ariel OS, is mature enough for industrial IoT use cases. Teams building embedded ML accelerators or sensor nodes should evaluate Rust for new projects.
Efficient alert-driven monitoring for AI/ML systems
The article's "start with failure" philosophy is directly applicable to monitoring machine learning models in production. Instead of building dashboards of metrics like CPU or latency first, teams should define what constitutes model failure (e.g., drift, data quality issues) and design alerts around those signals. This reduces alert fatigue and improves MLOps reliability.
Hardware innovation (Z-Angle Memory) will fuel next-gen AI accelerators
Intel's development of Z-Angle Memory hints at novel memory architectures that could break the von Neumann bottleneck. For AI/ML, memory bandwidth and capacity are often the limiting factors for training and inference. If Z-Angle Memory delivers higher bandwidth or lower latency, it could enable larger models and faster training cycles. AI hardware designers should watch this technology closely.
Analysis generated by deepseek-reasoner