Rust borrow checker flaw allows lifetime bypass via covariant field assignment
A bug in Rust's borrow checker allows struct lifetime requirements to be bypassed through covariant field assignment. The issue, reported on current nightly and stable versions back to 1.85.0, compiles code that should be rejected. When run with AddressSanitizer, the program crashes with a stack-use-after-scope error. The Rust team classifies this as a P-high issue, noting that while it is an 11-year-old flaw, exploiting it requires an intentionally malicious Drop implementation. The team emphasizes that fixing type system unsoundnesses is a major priority to ensure memory safety.
github.com
·
HN discussion
Anthropic replaces git with in-house system Scion to handle AI-driven merge volume
Anthropic has fully migrated its monorepo from git to an in-house version control system called Scion. The company states that daily merges grew from 2,500 in May to over 30,000 by September, making source control a primary bottleneck. Scion uses sparse and virtualized checkouts to reduce local data requirements. The migration took five months and included building a code review interface and identity management system. Anthropic notes that traditional git mirrors strained capacity as the repository size exceeded disk and memory limits.
anthropic.com
·
HN discussion
OpenAI engineer arrested for hiding audio recorder in executive conference room
Federal prosecutors arrested an OpenAI engineer on suspicion of hiding an audio recorder in an executive conference room at the company’s San Francisco headquarters. Abdulrahman Alfozan, a Saudi national, was charged with attempting to intercept oral communications. The device was concealed in toilet paper behind a television and contained encrypted data. Prosecutors allege Alfozan used productivity tools to research meeting schedules. A chatbot account tied to his spouse submitted prompts about building and encrypting a recording device. Surveillance footage shows him near the room on multiple dates, despite lacking entry authorization. An unattended bag later revealed a second similar device.
sfstandard.com
·
HN discussion
EmbeddingGemma 2 converted into a speech-to-text model
Developers converted Google's EmbeddingGemma 2 into a speech-to-text model. The tool processes an hour of audio in about 13 seconds on a base MacBook Air, running entirely on device. It requires macOS 14 or later and Apple Silicon, with models downloading on first launch. The approach leverages the model's multimodal embedding space to map audio slices to text. A small reader moves each short audio slice into the model's text space, where the closest of roughly 229,000 English words wins. The base model was never trained for speech-to-text; only the reader is trained. A tiny extra model adds punctuation and capitals. The audio encoder stays exactly as it is.
eg2.haku.sh
·
HN discussion
Ben Recht calls OpenAI’s faster Fourier transform a mathematical artifact
Ben Recht argues that OpenAI’s claimed faster Fourier transform is a mathematical artifact rather than a practical breakthrough. The paper, part of OpenAI’s recent mathematics release, offers a theoretically smaller logarithmic term than the standard FFT. However, Recht notes the hidden constant is so large that the algorithm only outperforms the standard method at impossibly large input sizes. He further criticizes the work for relying on a computation model that assumes arbitrarily large complex number multiplications occur in a single step.
argmin.net
·
HN discussion
Minions aggregates idle Apple Silicon GPUs for distributed inference
Researchers present Minions, a distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to handle large-scale machine learning inference. The system addresses limitations in centralized GPU infrastructures, such as high costs and limited elasticity, by leveraging existing hardware without requiring additional procurement. Minions provides flexible task execution, fault tolerance, and efficient resource utilization. The authors report that this approach transforms idle devices into a practical inference platform for environments with constrained centralized resources.
machinelearning.apple.com
·
HN discussion