Published on May 21, 2026 at 06:00 CEST (UTC+2)
An OpenAI model has disproved a central conjecture in discrete geometry (835 points by tedsanders)
An OpenAI model has disproved a central conjecture in discrete geometry
OpenAI's AI model successfully disproved a long-standing conjecture in discrete geometry, marking a milestone in AI-driven mathematical discovery. The model identified a counterexample to a problem that had resisted human attempts for years. This demonstrates AI's growing ability to not just solve but also challenge established mathematical theorems. The result signals a shift where AI can actively contribute to foundational research in mathematics.
GitHub confirms breach of 3,800 repos via malicious VSCode extension (606 points by Timofeibu)
GitHub confirms breach of 3,800 repos via malicious VSCode extension
GitHub suffered a security breach after an employee installed a poisoned VS Code extension, leading to the exfiltration of about 3,800 internal repositories. The company quickly removed the malicious extension and secured the compromised device. While no customer data outside those repos appears affected, a hacker group called TeamPCP claimed responsibility and is demanding $50,000 for the stolen code. The incident highlights the growing threat of supply-chain attacks through developer tooling extensions.
Show HN: I reverse engineered Apple's video wallpapers (125 points by kageroumado)
Show HN: I reverse engineered Apple's video wallpapers
A developer created "Phosphene," a macOS video wallpaper engine that leverages Apple's private WallpaperExtensionKit framework. The app lets users play custom video files as desktop and lock-screen wallpapers, integrating natively with System Settings. It reverse-engineers the same private framework used by Apple's own Aerials wallpapers, ensuring stable out-of-process playback that survives app quits. This project demonstrates deep reverse-engineering of macOS internals for user customization.
DOS Zone (136 points by rglover)
DOS Zone
DOS Zone is a web-based platform offering over 2,000 classic DOS games playable directly in the browser. It supports mobile devices, offline play, and includes multiplayer hubs, with a curated collection of Russian-language games. The site organizes games by genre and popularity, featuring titles like Grand Theft Auto: Vice City, DOOM, and Quake III Arena. It serves as a comprehensive archive and emulation portal for retro gaming enthusiasts.
The Letter S, by Donald Knuth (1980) [pdf] (72 points by bambax)
The Letter S, by Donald Knuth (1980) [pdf]
This PDF is a classic 1980 paper by computer scientist Donald Knuth, dedicated to the typographic design of the letter "S." Knuth explores the mathematical and aesthetic principles behind letterform construction, using his METAFONT system. The document is a blend of technical rigor and artistic sensitivity, illustrating how computer science can intersect with typography. It remains a seminal reference in digital type design.
Colorado Amended SB051 (Age Verification Bill) to Exclude Open Source Projects (161 points by ki4jgt)
Colorado Amended SB051 (Age Verification Bill) to Exclude Open Source Projects
Colorado's SB051, an age verification bill originally aimed at online platforms, was amended to explicitly exclude open source software projects from its requirements. The change came after advocacy from the open source community, which warned that broad age verification laws could cripple decentralized and collaborative development. The amendment protects open source contributors from liability, though the bill still targets commercial platforms. This highlights ongoing regulatory battles between online safety laws and open source ecosystems.
Flipper One Tech Specs (296 points by gregsadetsky)
Flipper One Tech Specs
The Flipper One is an advanced multi-tool device for hardware hacking, with specs including a 256×144 pixel monochrome LCD, a Rockchip RK3576 CPU, and various ports (USB-C, Ethernet, GPIO). It runs on Flipper OS with both Linux (CPU) and MCU firmware, supporting modular expansion via an M.2 slot. Designed for pentesting and research, it features a heat sink, anodized aluminum body, and Wi-Fi/Bluetooth connectivity. The device is under active development with final weight pending.
Haskell Foundation 2026 Update (19 points by azhenley)
Haskell Foundation 2026 Update
The Haskell Foundation announced that its executive director José is stepping down after several years of service. The board is using this transition to restructure the organization, shifting focus toward funding technical work and giving members more direct influence over projects. The goal is to improve the Haskell ecosystem by creating clearer links between contributions and tangible improvements. The update reflects a strategic pivot toward community-driven, member-owned development.
Your Most Improbable Life (57 points by jger15)
Your Most Improbable Life
Kevin Kelly argues that the goal of life should be to become the most improbable version of oneself, leveraging uniqueness to avoid competition and AI replacement. He connects this to physics concepts of entropy and exotropy, where life accelerates entropy to create local pockets of order and unpredictability. The essay encourages embracing unlikely paths and creative risk-taking. It frames human distinctiveness as a strategic advantage in an AI-driven world.
How fast is N tokens per second really? (329 points by hexagr)
How fast is N tokens per second really?
This interactive tool lets users visualize and compare LLM token generation speeds from 5 to 800 tokens per second across four modes: code, text, think (reasoning model), and agent (tool calls). It shows that raw token rates feel very different depending on the content type—code looks faster than prose, while thinking pauses create a different rhythm. The tool helps internalize benchmarks like "47 tok/s on M3" or "800 tok/s on Cerebras" by showing real-time streaming. It also clarifies tokenization approximations and highlights that eyeballs become the bottleneck at very high speeds.
AI as a mathematical discovery engine
The OpenAI model disproving a discrete geometry conjecture marks a shift from AI as pattern recognizer to AI as creative disprover of human assumptions. This trend suggests that LLMs and specialized models will increasingly be used to generate counterexamples and explore mathematical frontiers. For AI/ML development, it implies that investing in formal reasoning and theorem-proving capabilities can unlock entirely new research directions, potentially accelerating progress in pure mathematics and theoretical computer science.
Supply-chain attacks via AI/ML developer tooling
The GitHub breach through a malicious VS Code extension underscores the expanding attack surface created by the proliferation of AI-powered code assistants, plugins, and extensions. As developers rely more on AI tooling (e.g., Copilot, code generators), the risk of compromised extensions grows. This trend demands stronger sandboxing for developer tools, automated scanning of extension behavior, and zero-trust policies for internal code repositories. The incident is a wake-up call for the AI tool ecosystem to prioritize security alongside productivity.
Regulatory pushback from open source and AI communities
Colorado's amendment to exclude open source from age verification reflects a broader tension between AI/tech regulation and open source development. As governments introduce laws targeting AI transparency, safety, and content moderation, open source projects often face disproportionate compliance burdens. The trend points to a growing need for carve-outs and legal frameworks that protect decentralized innovation while still addressing harmful uses. For AI/ML, this could shape how open-source models are distributed and governed.
Human uniqueness as a strategic AI hedge
Kevin Kelly's essay on becoming "improbable" taps into a cultural narrative: as AI automates predictable tasks, the most valuable human contributions will be those that defy statistical likelihood. This insight resonates with the AI/ML community's focus on "creativity," "serendipity," and "out-of-distribution" thinking. For developers and organizations, it suggests investing in cross-disciplinary skills, radical exploration, and non-standard problem-solving that machines cannot easily replicate. The trend may influence career advice and product design in AI-driven markets.
Token speed perception affects LLM user experience
The token speed tool reveals that raw throughput numbers (e.g., 60 tok/s) feel very different depending on whether the output is code, text, or reasoning chains. This matters for AI/ML product designers and researchers: optimizing for speed alone is insufficient if the user experience breaks down due to mode-switching pauses or variable content types. The trend suggests that next-generation inference engines should prioritize “responsive streaming” that adapts to content semantics, and that benchmarks should include qualitative user tests alongside raw token-per-second metrics.
Reverse engineering of private frameworks for AI-adjacent applications
The Phosphene project reverse-engineers Apple’s private wallpaper framework, a trend that parallels how many AI/ML tools exploit undocumented or semi-public APIs. This highlights a growing practice of “gray-area” engineering to extend capabilities beyond official SDKs. For the AI/ML community, it raises both opportunities (quickly prototyping novel interfaces) and risks (API breakage, legal challenges). The trend encourages better documentation of system-level AI integrations by platform vendors.
Open source AI infrastructure and community governance shifts
The Haskell Foundation restructuring—moving from a donor model to a member-owned technical focus—exemplifies a broader movement in open-source AI/ML communities. As foundations like Hugging Face, PyTorch, and LangChain grow, they face similar governance challenges: balancing corporate sponsorship against community ownership. The trend indicates that sustainable AI/ML ecosystems require transparent decision-making, direct funding of concrete technical work, and mechanisms for members to prioritize projects. The Haskell Foundation’s pivot could serve as a model for other AI-related open-source organizations.
Analysis generated by deepseek-reasoner