HackerNews Digest

The scraped text consists solely of a page header and a user‑interface prompt. The header reads “ChatGPT – Jacobian Conjecture Counterexample,” suggesting the page was intended to discuss a purported counterexample to the Jacobian Conjecture, but no substantive content follows. The only additional text is a generic call‑to‑action: “Get responses tailored to you. Log in to get answers based on saved chats, plus create images and upload files.” No abstract, definitions, mathematical statements, proofs, references, or any explanatory material related to the Jacobian Conjecture or a counterexample is present. Consequently, the page provides no technical information beyond its title and a login prompt.

None

Gigatoken is a Rust‑based tokenizer offering up to ~1000× speedup over HuggingFace and tiktoken tokenizers while providing drop‑in compatibility. It can be used via its native API or in compatibility mode (`as_hf()`, `as_tiktoken()`) with minimal code changes. Benchmarks on Apple M4 Max (16 cores) and AMD EPYC 9565 (144 cores) show throughput of 8 GB/s (≈1900 Mtok/s) versus 6 MB/s (≈1.4 Mtok/s) for HF, confirming consistent gains across modern x86 and ARM CPUs. Performance stems from SIMD‑accelerated pre‑tokenization, reduced branching, extensive caching of pre‑token mappings, and minimized Python‑Rust interaction. Gigatoken supports most BPE tokenizers, including Llama 3, Qwen 2/3, DeepSeek, GLM 4/5, Nemotron 3, Kimi K2, Phi‑4, TinyLlama/Phi‑3, Gemma 3/4, and related fine‑tunes; SentencePiece and WordPiece are less optimized or unsupported. Limitations include slower ABI3 Python calls, missing file‑sink API, incomplete Windows testing, and pending SIMD porting. Users can validate support with the provided CLI benchmark command.

SIMD (single‑instruction multiple‑data) lets a CPU apply one instruction to several values simultaneously, turning byte‑wise loops into vector‑width chunks (e.g., 4, 8, 16 lanes). The typical SIMD pattern consists of five steps: (1) broadcast constants and create vector accumulators; (2) iterate over the input in full‑vector strides; (3) perform the parallel operation (comparison, arithmetic, etc.); (4) reduce or store the vector result (e.g., using @reduce, @bitCast, @ctz to locate the first failing lane); and (5) handle the remaining elements with a scalar tail. A concrete example scans decoded Unicode codepoints for the first value ≤ 0xF, achieving up to 4× speed on ARM NEON, 8× on AVX2, and 16× on AVX‑512 (≈5× real‑world gain on an AVX2 desktop). Compilers can auto‑vectorize simple loops but often miss opportunities, so explicit SIMD code provides predictable performance. Understanding this pattern enables any developer to identify and safely rewrite hot loops that process large contiguous data without needing assembly expertise.

The article evaluates whether frontier AI labs “pelican‑max” their models by over‑optimizing for the informal “pelican on a bicycle” benchmark. The author generated 1,008 SVG images using eight animal × six vehicle prompts across seven models (GPT‑5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen 3.7‑Max, GLM‑5.2, DeepSeek V4 Pro), three samples per prompt at temperature 1.0. SVGs were rendered to PNG, judged by GPT‑5.6 Luna on animal, vehicle, and action coherence (1‑5), and annotated by Gemini 3.1 Flash‑Lite for recognized elements. Results show: • Pelicans rank 6th of 8 animals; bicycles rank 2nd‑last of 6 vehicles, indicating no special advantage. • Fixed‑effects regression finds no significant per‑lab boost for pelicans, a modest but non‑significant boost for bicycles (only Gemini 3.5 Flash at p = 0.022, not surviving multiple‑comparison correction), and no significant pelican‑bicycle interaction. • All 21 pelican‑bicycle images face right, but this matches overall right‑facing frequency (≈60%). • Feature extraction reveals no unique repeated scene elements. Limitations include a single LLM judge, potential SVG‑maxxing, and a modest $80 API budget. The study concludes there is little evidence of targeted “pelican‑maxxing” by AI labs.

Scientists used ancient DNA from the skeletal remains of Grand Duke Francesco I de’ Medici (d. 1587) and his brother Giovanni to test whether malaria, not arsenic poisoning, caused their deaths. DNA extraction from rib bone revealed genetic signatures of *Plasmodium falciparum* and *P. malariae* in Francesco, indicating a double infection, and a previously unknown *P. falciparum* strain in Giovanni. These results corroborate contemporary physician reports describing intermittent fever and other malaria‑consistent symptoms, and they align with earlier paleo‑immunology findings that detected malaria antigens. While the DNA evidence strongly supports malaria as a major factor, researchers acknowledge it does not exclude concurrent poisoning; previous toxicological studies reported arsenic traces from separate tissue samples. The study demonstrates the utility of paleogenomics for resolving historical medical debates and provides rare data on Renaissance‑era malaria strains in central Italy.

ascdraw is a keyboard‑first, Unicode‑based diagram editor supporting infinite canvases, connected lines, symbols, shapes, rectangular editing, layers, and export to TXT, JSON, and PNG. It runs at 120+ FPS and is distributed under GPLv3, with optional personal licensing ($9.99/€9.99) and commercial licenses available. Installation uses the Rust toolchain (e.g., `cargo build --release --locked` or `cargo install --path . --locked`). The default interface opens in Stamp mode; navigation uses arrow keys (h j k l) and a set of shortcuts for Stamp, Line, Shape, Utils, and file operations. Optional features include 16‑color ANSI styling (preserved in PNG/JSON), multi‑layer management, and a dark mode. Input can be filtered from stdin and output to stdout, enabling use as an external editor (`ascdraw -output.txt`). Configuration merges bundled `ascdraw.toml` and `theme.toml` with user overrides in `$XDG_CONFIG_HOME/ascdraw/config.toml` or `~/.config/ascdraw/config.toml`, reloading live. The project cites OpenAI GPT‑5.5/5.6 assistance and is copyrighted 2026 by Przemysław Alexander Kamiński.

GitHub is revising its bug bounty program to prioritize high‑impact findings over report volume. A permanent, invitation‑only VIP program will reward researchers who consistently deliver critical or multiple high‑medium‑low severity bugs with higher payouts, faster response times, and direct interaction with the security team. Qualification requires at least one critical, two high, four medium, or seven low findings, with criteria posted on HackerOne. The public bounty table shifts to static payouts per severity level, eliminating variable ranges and allowing discretionary bonuses for exceptional work. A new HackerOne “signal” threshold limits submissions from researchers without an established track record to four initial reports, aiming to filter low‑effort or AI‑generated submissions while still permitting newcomers to demonstrate value. Existing reports remain under the old structure; the new model applies to reports submitted on or after July 27 2026. GitHub also commits to faster response times, clearer severity rationales, and ongoing community engagement through conferences and outreach.

- Cactus‑Hybrid augments on‑device LLM checkpoints with an internal probe that outputs a confidence score (0‑1) for each answer, enabling automatic handoff to a larger model when confidence < 0.85. - The initial release is Gemma 4 E2B Hybrid, the smallest Gemma variant; it answers locally for ~65‑85 % of queries and routes the remainder to Gemini 3.1 Flash‑Lite, matching its benchmark performance while reducing cloud usage. - Quantisation quality is evaluated on Cactus Quants; developers may also test Unsloth, GGUF, and MLX quantizations. - Example usage is provided for three runtimes: • Python cactus‑compute bindings (`cactus_complete`) with optional auto‑handoff. • MLX‑LM (`load`, `generate`) where `model.last_confidence` returns the score. • Transformers ≥ 5.5.4 (`AutoModelForCausalLM`) with `return_confidence=True`; the model must be loaded via explicit `.to(device)` rather than `device_map="auto"` to keep the probe functional. - For llama.cpp, a C++ patch is compiled into the server; responses include a top‑level `confidence` field. - AUROC evaluation shows the probe distinguishes correct from incorrect outputs with 0.79‑0.88 on four audio benchmarks, indicating a modality‑independent correctness signal. - The code is MIT‑licensed; Gemma usage follows Gemma’s licensing terms.