HackerNews Digest

GenCAD is an image‑conditional generative model that produces fully parameterized CAD programs, not just 3D solids. It addresses the difficulty of training AI on complex CAD data structures (e.g., B‑rep) by learning joint representations of CAD command sequences and reference images. The architecture comprises four stages: (1) an autoregressive transformer encoder for latent encoding of CAD command sequences; (2) a contrastive learning module that aligns latent spaces of CAD commands and CAD images; (3) a latent diffusion model that generates CAD‑latent vectors conditioned on images; and (4) a decoder that translates these latents into executable parametric CAD commands. By outputting the complete CAD command history, GenCAD enables accurate, modifiable models suitable for engineering, manufacturing, and design space exploration, surpassing mesh, voxel, or point‑cloud approximations.

The repository provides **rkdebian**, a build system that creates a bootable Debian 12 (Bookworm) image for the Doogee U10 tablet (Rockchip RK3562). The image runs from an SD card, leaving the internal eMMC untouched; removing the card restores Android. It includes U‑Boot, a 6.1‑kernel (or firefly‑5.10), and a root filesystem built via debootstrap. Build dependencies are installed on an x86‑64 host, and the `build.sh` script supports targets such as `all`, `image`, `updatepkg`, and various configuration flags (UI session, GPU stack, display server, CPU governor, rootfs size, etc.). The system supports local LLM inference on the RK3562 NPU using Rockchip’s RKLLM stack; a sample conversion for Qwen 3‑0.6B (W8A8 quantization, 1 NPU core, optimization 0) shows ~1.8 s latency versus ~4.8 s for Qwen 2.5‑1.5B. Additional features include camera drivers (s5k5e8, s5k4h5yb), flashlight control, power‑profile mapping, and OTA update handling via `rk-apply-update`. Third‑party components (Mali GPU, Rockchip MPP, Seekwave Wi‑Fi/BLE) are included under their respective licenses.

The provided excerpt consists solely of the title “Ask an Astronaut” and contains no additional content to summarize.

The page critiques common Prolog antipatterns and demonstrates how impurity and outdated constructs hinder declarative programming. Key issues listed are: loss of solutions (incorrect or missing answers), reliance on global state, impure I/O (e.g., `format/2` within predicates), and use of low‑level arithmetic (`is/2`) that forces procedural evaluation. An example, `horror_factorial/2`, uses `is/2` and a cut, causing instantiation errors and limiting query generality. A pure alternative replaces `is/2` with constraint‑logic programming operators (`#>`, `#=`) and omits cuts, yielding a relational factorial that correctly enumerates all solutions (`N=0,1,2,…`). The author concludes that adopting pure, constraint‑based definitions eliminates these “horrors” and restores declarative debugging capabilities.

Doron Zeilberger argues that a good lemma outweighs many theorems in mathematical impact. He cites well‑known lemmas—Schur’s Lemma, Lovász’s Local Lemma, and especially Szemerédi’s Regularity Lemma—as examples of tools that have generated major advances, including two Fields Medals and the Green‑Tao theorem on arithmetic progressions in primes via a hypergraph extension. Zeilberger emphasizes that lemmas “do the work” while theorems receive credit, and that a valuable lemma must (1) apply broadly across diverse problems, (2) become instantly obvious once seen, and (3) possess an aesthetically pleasing proof. He references the “Lattice Paths and Determinants” chapter in Aigner & Ziegler’s *Proofs from THE BOOK* and notes that observations can be even more significant than lemmas.

Two U.S. Navy EA‑18G Growler jets from Electronic Attack Squadron 129 collided during an aerial demonstration at the “Gunfighter Skies” airshow on Sunday at Mountain Home Air Force Base, about 80 km south of Boise. All four crew members ejected safely and were examined by medical personnel; no one on the base was injured. The base locked down and Idaho State Highway 167 was closed for a multi‑day investigation. Spectator video shows the jets contact, spin together, and explode on impact. Weather at the time was clear with wind gusts up to 29 mph. The show, the first at the base since 2018, also featured the U.S. Air Force Thunderbirds. Historical references note a 2003 Thunderbirds crash, a 2022 fatal collision at a Dallas airshow, and a long‑term decline in airshow fatalities—none since 1952. Officials, including Col. David Gunter and the Silver Wings of Idaho board, emphasized crew safety and the effectiveness of emergency response.

Semble is a CPU‑only code‑search library designed for LLM agents. It indexes a repository in ~250 ms and answers queries in ~1.5 ms, offering ~200× faster indexing and ~10× faster retrieval than a code‑specialized transformer while attaining NDCG@10 = 0.854 (≈99 % of a 137 M‑parameter model). The system splits files into code‑aware chunks (via Chonkie) and scores queries with two static retrievers: Model2Vec embeddings (potion‑code‑16M) for semantic similarity and BM25 for lexical matching. Scores are fused by Reciprocal Rank Fusion and re‑ranked using adaptive weighting, definition boosts, identifier‑stem matching, file‑coherence, and noise penalties (e.g., test files, .d.ts). Results are returned as minimal chunks, reducing token usage by ~98 % compared with grep‑plus‑read workflows. Semble can run as an MCP server for agents such as Claude Code, Cursor, Codex, OpenCode, or as a standalone CLI (`semble search`, `semble find‑related`). Installation is via `pip install semble` or `uv`, and token savings are tracked in `~/.semble/savings.jsonl`. The project is MIT‑licensed and citable via Zenodo.

Released at Outline Demoparty May 2026, “wake up! 16b” is a 16‑byte x86 real‑mode DOS program that uses the VGA text buffer (DS = 0xB800) as a calculation canvas to generate an infinite Sierpinski fractal while simultaneously driving the PC speaker. After BIOS int 10h sets 40×25 text mode (character 0x20, attribute 0x07), the loop performs `lodsb`, `sub si,57` (net –56 bytes per iteration), `xor [si],al`, and `out 61h,al`. Modeling the algorithm with 16‑byte steps shows a running prefix sum whose values follow 2·C(k+p, p‑1) (mod 256); replacing addition with XOR isolates bit 1, yielding Rule 60 cellular automaton behavior that maps directly to the Sierpinski triangle. The `out 61h,al` instruction toggles speaker bit 1, producing square‑wave audio whose rhythm mirrors the fractal’s geometry. The –56‑byte stride visits only addresses multiples of 8, requiring 8 192 iterations (seven wraps) and halving the audio frequency (one octave lower). Spatially, –56 ≡ 24 (mod 80) bytes shifts the pattern 12 columns right per step, resulting in ten evenly spaced vertical columns. Output varies with initial memory contents; a full clear would improve determinism but exceeds the 16‑byte constraint.

None