Published on July 29, 2026 at 06:00 CEST (UTC+2)
Codex Security (392 points by bakigul)
Codex Security – OpenAI’s open-source CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities in code. It integrates into CI workflows, scans repositories, and tracks findings over time. The tool uses OpenAI’s Codex (likely an AI model) to assist with security analysis, requiring Node.js 22+, Python 3.10+, and an OpenAI API key.
LearnVector – Andrew Ng's AI company building one‑to‑one learning experiences (71 points by ajhai)
LearnVector – Andrew Ng’s new AI company focused on creating one-to-one personalized learning experiences. The mission is to overcome the historical scarcity of great teaching by using AI tutors, rather than replacing humans. It challenges the chatbot model, arguing that simple Q&A leads to cognitive offloading and unreliable answers, and instead aims for trustworthy, adaptive instruction.
Show HN: I was tired of opening 2 tabs for every HN link, so I made a userscript (181 points by twalichiewicz)
HNewhere – A lightweight userscript that detects Hacker News discussions for any article you’re reading and loads them into a resizable sidebar. It eliminates the need to open a separate tab for each HN link, making it easier to browse comments alongside content. The script is open source and works with Tampermonkey, Violentmonkey, or Userscripts.
Half-Life ported to Mac OS 9 (182 points by freediver)
Half-Life ported to Mac OS 9 – A community port of the classic first-person shooter to PowerPC-based Macs running Mac OS 9, 28 years after its original release. The port uses a fork of the Xash3D engine and supports the full game, multiplayer, and expansion packs. Performance varies based on GPU and VRAM, but it’s a significant achievement for retro Mac gaming.
Substack writers, you need a website (449 points by speckx)
Substack writers, you need a website – An opinion piece arguing that writers should treat Substack as a distribution tool, not their primary digital home. The author warns against relying solely on a subdomain (e.g., xx.substack.com) because it puts content under Substack’s control. Owning your own domain and using a proper CMS gives long-term stability and SEO control.
Hubble: Open-source notetaking app for you and your agents (61 points by handfuloflight)
Hubble – An open-source notetaking app designed for both humans and AI agents. It stores notes in Markdown and HTML, supports tagging, and allows building custom views like maps or bookshelves. The app emphasizes agent-friendly organization, with the tagline “the best notepad for you and your agents.”
Multiple Mouse Cursors in Wayland (36 points by marvinborner)
Multiple Mouse Cursors in Wayland – An exploration of multi-seat (multiple mice, multiple cursors) support on Linux with Wayland. The author created tools and patches to enable shared desktop environments where two people can interact simultaneously, useful for pair programming or collaborative games. The post notes that core Wayland support is excellent but many applications still lack multi-cursor awareness.
Teach yourself programming in ten years (1998) (73 points by vinhnx)
Teach Yourself Programming in Ten Years – Peter Norvig’s classic essay debunking “learn X in 24 hours” books. He argues that true programming mastery requires deliberate practice over a decade, citing expertise research. The piece emphasizes writing real programs, collaborating with peers, and studying classic works rather than rushing through shallow tutorials.
Truth is not a direction: a Tarski attack on LLM probes (37 points by abelaer)
Truth is not a direction: a Tarski attack on LLM probes – A technical critique of the linear representation hypothesis for truth in LLMs. The author uses a diagonalization argument (inspired by Tarski) to show that no linear probe on an LLM’s embedding space can reliably detect truth. The paper warns against overconfidence in safety-relevant probes, suggesting that truth is not a simple geometric direction.
ReFrame – The EPaper Camera (55 points by phil294)
ReFrame – The EPaper Camera – An open-source experimental camera that uses a color ePaper display to show one photo at a time. The slow, process-driven display (15 seconds per photo) creates a deliberate, print-like experience. The camera is designed for reflection rather than rapid sharing, and its hardware and software are fully open source.
AI-assisted security scanning is maturing into production-grade tooling – OpenAI’s Codex Security (article 1) represents a shift from experimental code generation to practical, CI-integrated security auditing. This trend shows AI is being embedded into developer workflows to catch vulnerabilities automatically, moving beyond simple autocomplete. Implication: Security teams should evaluate AI-powered scanning tools, but also be aware of false positives and model biases in security contexts.
Personalized AI tutors challenge the “chatbot as teacher” paradigm – LearnVector (article 2) explicitly rejects unfiltered chatbots for learning, citing cognitive offloading and untrustworthy answers. The emphasis is on structured one-to-one AI instruction with guardrails, signaling that the next wave of edtech AI will focus on pedagogy, not just conversational fluency. Implication: AI education products must prioritize trustworthiness and learning science over raw interactivity.
Cracks are appearing in the linear representation hypothesis for truth – The Tarski attack (article 9) provides a theoretical limitation on using linear probes to detect truthfulness in LLMs. This challenges a popular safety research approach and suggests that truth in LLMs may be more complex and context-dependent than a simple geometric direction. Implication: AI safety researchers should diversify beyond linear probes and consider more robust methods (e.g., consistency checks, mechanistic interpretability) to assess honesty.
Open-source AI tools are becoming agent-friendly – Hubble (article 6) explicitly markets itself as a notepad for both humans and AI agents, with agent-ready views and Markdown/HTML export. This reflects a growing trend where applications design data structures and APIs to be consumed by autonomous agents, not just human users. Implication: Developers building agent ecosystems should adopt open, structured formats (Markdown, JSON, etc.) to ensure interoperability.
Classic programming wisdom remains relevant in the age of AI hype – The resurgence of Peter Norvig’s “Teach Yourself Programming in Ten Years” (article 8) on Hacker News suggests a counter-reaction to “learn AI in 24 hours” courses. The piece warns that deep expertise requires years of deliberate practice, not quick tutorials. Implication: As AI lowers barriers to entry, the value of foundational programming skills and long-term learning will increase, not decrease.
AI probes and interpretability face fundamental theoretical limits – Article 9 connects Tarski’s undefinability theorem to LLM probing, echoing earlier work on the impossibility of perfectly detecting truth in formal systems. This insight raises the bar for AI alignment: if truth probes are provably incomplete, we need complementary approaches like behavioral audits or training-time interventions. Implication: The AI safety community should invest in multiple, non-reducible methods for evaluating model honesty.
Human-centric AI design is pushing back against replacement narratives – Andrew Ng’s LearnVector (article 2) contrasts with the “AI will replace people” trope by focusing on human development. Similarly, the ReFrame camera (article 10) embraces slowness and deliberation over AI-assisted instant gratification. Together, these indicate a trend where AI is designed to augment human agency rather than automate it away. Implication: Product designers should consider “AI as accelerator” rather than “AI as replacement” to build trust and long-term value.
Analysis generated by deepseek-reasoner