HackerNews Digest

OpenAI announced GPT‑2, a language model trained on text from 8 million web pages that can predict the next word and generate coherent, style‑adaptable prose from prompts. Citing potential misuse—such as automated disinformation, impersonation, and spam—the organization released only a reduced‑size version and withheld the full model, training data, and code. Media coverage exaggerated the threat, prompting debate among AI researchers. Technically, GPT‑2 improves on earlier generators by handling multiple word senses, producing longer passages, and adapting to diverse writing modes, though it still exhibits repetition and occasional incoherence. Experts argue that withholding the model may be a limited safeguard, as well‑resourced actors could replicate it using cloud compute, and that the decision could impede academic research. The episode underscores the nascent field’s lack of consensus on ethical release practices and mirrors historic tensions over technology dissemination, such as the 1990s encryption debate.

The United States and Iran announced a provisional two‑week cease‑fire, mediated by Pakistan’s prime minister, contingent on Iran’s “complete, immediate and safe” reopening of the Strait of Hormuz. President Trump posted that, subject to this condition, he would suspend planned bombings of Iranian power plants and bridges. Iran’s foreign minister confirmed the agreement, and Israel reportedly consented once the strait’s blockade ends. The cease‑fire is intended to allow negotiations on Tehran’s 10‑point proposal for a longer‑term armistice. Prior to the announcement, Israeli forces struck several Iranian infrastructure sites, including rail bridges and a power substation, while the U.S. hit roughly 50 military targets on Iran’s Kharg Island oil terminal. In response, Iran’s Revolutionary Guards attacked Saudi Arabia’s Jubail petrochemical complex. UN representatives labeled Trump’s threats as potential war‑crime incitement, noting that attacks on civilian infrastructure violate international law.

Anthropic’s Project Glasswing introduces Claude Mythos Preview, an AI model that autonomously discovers and exploits zero‑day vulnerabilities in major operating systems, browsers, and other critical software. In recent tests the model uncovered a 27‑year‑old remote‑crash flaw in OpenBSD, a 16‑year‑old code error in FFmpeg missed by five million automated tests, and a chain of kernel bugs in Linux allowing privilege escalation; all have been patched. Benchmarking (CyberGym, CTI‑REALM) shows Mythos Preview outperforming Claude Opus 4.6 and prior models. Anthropic has partnered with firms including Cisco, AWS, Microsoft, CrowdStrike, Google, JPMorgan Chase, and others, granting them access to the model for vulnerability detection, binary black‑box testing, endpoint hardening, and penetration testing. The initiative commits $100 M in model‑usage credits, $2.5 M to Alpha‑Omega/OpenSSF, and $1.5 M to the Apache Software Foundation, with post‑preview pricing of $25–$125 per M tokens via Claude API, Bedrock, Vertex AI, and Microsoft Foundry. Project Glasswing will publish disclosed vulnerabilities, issue best‑practice recommendations (disclosure, patching, supply‑chain security, SDLC), and engage with governments to shape safeguards for AI‑driven cyber capabilities.

NASA’s Artemis II crew performed a seven‑hour lunar far‑side flyby on 6 April 2026, capturing the first human‑viewed images of that region and a prolonged in‑space solar eclipse. The photo set includes an Earth‑set view through the Orion window showing a muted blue Earth with bright clouds over Australia/Oceania, and the crater Ohm with terraced walls and central peaks. A halo surrounding the dark lunar disk is under investigation for possible corona or zodiacal‑light origins; the eclipse lasted about 54 minutes, far longer than terrestrial totalities. Planetary bodies visible in the frame are Venus (glint on the left), Saturn (below the Moon’s right edge), and Mars (right edge). Additional imagery shows Mission Control screens, crew interactions, and a back‑lit Moon with Earth‑reflected illumination. The release, dated 7 April 2026, emphasizes the technical significance of these unprecedented visual observations.

None

None

The blog post recaps a Thotcon presentation on binary obfuscation techniques that maintain compatibility with Link‑Time Optimization (LTO). It outlines the problem that conventional obfuscation often disrupts LTO, reducing the benefits of compiler‑level optimizations. The author describes approaches for applying obfuscation after LTO or integrating it into the compilation pipeline so that the optimizer’s analysis and transformations remain intact. Key points include preserving symbol information, handling indirect branches, and ensuring that the resulting binary remains functional while being harder to reverse‑engineer. The post also references practical considerations such as toolchain support and performance impact, providing a concise summary of methods that achieve obfuscation without sacrificing LTO‑derived performance gains.

The article promotes Justin Sandercoe’s guitar‑learning method: develop skill by listening and transcribing instead of relying on pre‑written tabs. The process is: * Choose an easy song with simple, single‑note riffs (e.g., “Breaking the Law,” “The Distance,” “The Ghost of Tom Joad”). * Use blank tab paper; play the track, pause at each note, locate it on the neck, and write it down. * Continue until the entire song is notated, then compare your transcription with online tabs or video performances to correct errors and notice techniques such as capos or unconventional picking. Repeated transcription builds ear training, note‑finding, and muscle memory, enabling near‑tempo play after a single pass. The author compiles completed songs into a practice playlist, emphasizing full‑song learning (intro, verses, chorus, bridge) rather than isolated riffs. A sample playlist includes tracks from The Strokes, Yeah Yeah Yeahs, The Cure, Rage Against the Machine, and Smashing Pumpkins. The method aims to improve overall musicianship and adaptability across styles.

Cambodia inaugurated a stone statue of Magawa, the award‑winning land‑mine detection rat trained by the United Nations Mine Action Service. The sculpture depicts a large rat wearing a black harness and a medal on a blue collar. During its service, Magawa cleared more than 141,000 m² (approximately 20 football pitches) of contaminated land and could survey a tennis‑court‑size area in about 20 minutes. The unveiling highlights the rat’s contribution to de‑mining efforts and serves as a public tribute to the animal’s role in improving safety in former conflict zones.

The post describes how data‑movement friction—copying between local file systems and Amazon S3—impedes workflows in genomics, machine‑learning, media, and other domains. Early attempts to run containerized analyses (“bunnies”) on S3 highlighted the need for a storage layer that supports both object and file‑system semantics. Amazon’s response evolved through three primitives: S3 Tables (managed Iceberg tables with automatic compaction and replication), S3 Vectors (S3‑native vector‑index type for similarity search), and now S3 Files, which integrates Amazon Elastic File System (EFS) into S3. S3 Files lets any S3 bucket or prefix be mounted on EC2, containers, or Lambda, exposing objects as a POSIX‑compatible file system and propagating file‑level changes back to S3. The design required extensive trade‑offs between file‑system and object‑store semantics, leading to a solution that prioritizes simplicity for applications while preserving S3 durability, cost, and performance characteristics.