Published on May 18, 2026 at 06:00 CEST (UTC+2)
GenCAD (163 points by dagenix)
GenCAD – This paper from MIT presents GenCAD, an AI model that generates parameterized CAD command sequences from image inputs. Unlike methods that output meshes or point clouds, GenCAD produces editable CAD programs convertible to 3D solid models via geometry kernels. The architecture combines an autoregressive transformer encoder, contrastive learning, and diffusion priors to learn representations across multiple computational engineering modalities. This work aims to bridge the gap between AI generation and practical engineering design by preserving accuracy and modifiability.
I turned a $80 RK3562 Android tablet into a Debian Linux workstation (277 points by tech4bot)
I turned a $80 RK3562 Android tablet into a Debian Linux workstation – The project, rk3562deb, details how to run full Debian 12 on a Doogee U10 tablet (RK3562 SoC) without unlocking the bootloader, booting from an SD card. The entire reverse engineering process was done from scratch using AI assistants – Claude, Codex, and Google Gemini – leveraging Firefly RK3562 open-source repositories as a starting point. This demonstrates how AI tools can significantly accelerate low-level hardware hacking and custom Linux porting.
Ask an Astronaut: 333 hours of Q&A footage with astronauts (63 points by gaws)
Ask an Astronaut: 333 hours of Q&A footage with astronauts – The website “Ask an Astronaut” curates over 333 hours of real-time question-and-answer sessions recorded aboard the International Space Station. It provides direct access to astronauts answering public questions about life in space, science experiments, and mission experiences. The project serves as an educational and historical archive, making rare astronaut insights freely available.
Prolog Coding Horror (71 points by RohanAdwankar)
Prolog Coding Horror – This article discusses common pitfalls in Prolog programming, focusing on two major defect types: reporting wrong answers and failing to report intended solutions. It warns against using impure, non-monotonic constructs like cuts, if-then-else, and var/1, advocating for declarative alternatives such as dif/2 and meta-predicates. The author argues that following a small set of rules can prevent costly defects and preserve logical purity.
A Good Lemma Is Worth a Thousand Theorems (2007) (21 points by susam)
A Good Lemma Is Worth a Thousand Theorems (2007) – Mathematician Doron Zeilberger argues that lemmas – seemingly minor but powerful results – often contribute more to mathematical progress than major theorems. He cites Szemerédi’s Regularity Lemma as a prime example, leading to multiple Fields medals and enabling breakthroughs like Green–Tao’s theorem on primes. The essay celebrates lemmas as the true workhorses of mathematics, while theorems get the credit.
Two EA-18 fighter jets collide at Mountain Home airshow, pilots ejected safely (133 points by ChrisArchitect)
Two EA-18 fighter jets collide at Mountain Home airshow, pilots ejected safely – During the Fighter Skies airshow at Mountain Home Air Force Base, two U.S. Navy EA-18G Growlers from Electronic Attack Squadron 129 collided mid-air. All four crew members successfully ejected and were evaluated for injuries, with no fatalities reported. The incident is under investigation by naval authorities.
Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep (207 points by Bibabomas)
Semble – Code search for agents that uses 98% fewer tokens than grep – Semble is a code search library designed for AI agents, returning exact code snippets using ~98% fewer tokens than traditional grep+read approaches. It indexes an entire codebase in under a second and offers latency reductions of ~10x compared to alternatives. The tool supports MCP servers, CLI, Python API, and bash integration, optimizing LLM-driven code retrieval for agent workflows.
Cannibalistic attacks between gray seals leave telltale “corkscrew” injuries (46 points by gmays)
Cannibalistic attacks between gray seals leave telltale “corkscrew” injuries – This Science article identifies that the mysterious “corkscrew” injuries found on gray seal carcasses are caused by cannibalistic attacks from other gray seals. The injuries involve a helical pattern of cuts, previously attributed to unknown predators or human activity. Researchers confirmed the cause through necropsy and behavioral observations.
Magical Realism: “Northern Exposure” 25 Years Later (2015) (84 points by walterbell)
Magical Realism: “Northern Exposure” 25 Years Later (2015) – Roger Ebert’s retrospective examines the TV series “Northern Exposure,” highlighting its blend of magical realism, quirky characters, and philosophical themes set in a fictional Alaskan town. The show, which aired from 1990 to 1995, is praised for its unique storytelling and cultural impact. The article reflects on its legacy and why it remains beloved decades later.
WriteUp: 16 Bytes of x86 that turn Matrix rain into sound (20 points by HellMood)
WriteUp: 16 Bytes of x86 that turn Matrix rain into sound – This demoscene write-up explains a 16-byte x86 real-mode DOS program that simultaneously draws a Sierpinski fractal on screen and generates audio output. The code uses video memory as a calculation space, employing a short loop of x86 instructions to produce algorithmic density. It exemplifies extreme constraint programming, combining graphics and sound in minimal machine code.
Generative AI for Engineering Design (CAD) – GenCAD illustrates a growing trend of using transformer-based models and diffusion priors to generate parametric design solutions directly from images. This moves beyond 3D shape generation (meshes/voxels) toward editable, physics-aware CAD programs, which are crucial for manufacturing and simulation. Implication: AI is increasingly becoming a tool for practical engineering workflows, reducing time from concept to production-ready designs.
AI-Assisted Reverse Engineering and System Porting – The RK3562 tablet project explicitly credits AI assistants (Claude, Codex, Gemini) for helping reverse engineer hardware and build a Debian port without vendor documentation. This indicates that LLMs are now capable of accelerating low-level tasks like device driver development and bootloader configuration. Takeaway: AI agents are evolving from code generators into collaborative problem-solvers for embedded systems and Linux porting.
Token-Efficient Code Search for AI Agents – Semble’s claim of using 98% fewer tokens than grep+read addresses a critical bottleneck in LLM-powered software engineering: context window limits and token cost. By returning only relevant code snippets, it enables agents to process large codebases without exhausting token budgets. This trend points to specialized infrastructure tools designed explicitly for agent workflows, not just human developers.
AI in Demoscene and Algorithmic Density – The 16-byte x86 demo showcases extreme algorithmic density – a concept that resonates with efficient AI model design. While not directly AI, the idea of maximizing output per compute unit parallels efforts to create small, efficient neural architectures (e.g., tinyML, pruning). Insight: The demoscene’s constraint-driven creativity can inspire AI researchers working on low-footprint models for edge devices.
Prolog and the Return of Logic Programming in AI – The Prolog horror article highlights pitfalls of non-monotonic constructs, which is relevant as AI systems increasingly rely on symbolic reasoning and neuro-symbolic hybrids. While deep learning dominates, logic programming’s declarative style is seeing renewed interest for explainability and rule-based verification. Implication: Understanding purity and monotonicity in Prolog can inform better design of LLM-based reasoning pipelines that rely on consistent knowledge bases.
Multi-Modal Learning Across Engineering Domains – GenCAD’s use of contrastive learning to align image and CAD command representations exemplifies the trend of multimodal foundation models. This mirrors developments in areas like text-to-3D and image-to-code. Companies investing in multi-modal AI will likely expand into technical domains (e.g., architecture, mechanical engineering), not just natural language and vision.
AI-Assisted Hardware Hacking as a New Normal – The RK3562 project’s reliance on AI for reverse engineering suggests that even niche, undocumented hardware can be tamed with LLM guidance. This democratizes access to hardware customization, potentially accelerating IoT and embedded development. However, it also raises concerns about security – AI could be used to find vulnerabilities as easily as to build systems.
Analysis generated by deepseek-reasoner