HackerNews Digest

The study led by Stanford Medicine shows that the human brain derives from two separate embryonic progenitor populations, not a single common precursor. One progenitor expresses Otx2 and forms the forebrain and midbrain; the other expresses Gbx2 and forms the hindbrain (brain stem). Chromatin analysis revealed mutually exclusive configurations that lock each lineage, explaining past failures to generate hindbrain neurons from forebrain‑midbrain progenitors. Using human pluripotent stem cells, the team successfully produced functional hindbrain motor neurons that display action potentials and region‑specific protein markers. Comparative analysis of mouse embryos, chickens, zebrafish, and acorn worms indicates this dual‑origin pattern dates back >550 million years, paralleling the two‑system organization seen in jellyfish. The findings clarify developmental biology, enable in‑vitro models for brain‑stem diseases such as spinal muscular atrophy and amyotrophic lateral sclerosis, and suggest new avenues for therapeutic research. The work was published in Nature Neuroscience (Sept 18) with contributions from Caltech, UCSF, and multiple NIH, NSF, and private grants.

The post describes how GPT‑6 “Astra” decoded a World‑I German ADFGVX radio cipher transmitted on 27 Nov 1918, one of the unsolved messages listed on Scienceblogs.de. Using the key “TRUPPENVERSCHIEBUNG”, which must be alphabetically reordered, Astra reconstructed the 6×6 polybius square, arranged the 170‑symbol ciphertext in rows of 19 under the key, and applied columnar transposition based on the key’s alphabetical positions. Decoding yielded the German plaintext: “EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X”, interpreted in English as “An English cruiser arrived at Sevastopol on the ?4th and an Allied squadron follows on the 26th.” Historical logs confirm HMS Canterbury reached Sevastopol on 24 Nov 1918, with an allied squadron arriving on 26 Nov, matching the decoded message. The article notes the key’s later documented use (from 9 Dec 1918) may explain why this cipher remained unsolved.

The essay argues that mathematics should value “motivated explanations” alongside traditional proofs, because the discipline’s ultimate aim is human understanding rather than merely generating formal proofs. A motivated explanation is defined by introducing definitions only after the problem context is clear, allowing provisional ideas that are later refined, and emphasizing why a theorem is the right question and how it fits within broader mathematics. The author distinguishes this genre from proofs, noting its non‑binary validity and the need for a practical rubric to assess motivation. Examples cited include the Princeton Companion to Mathematics, Bill Thurston’s essay on proof and progress, Timothy Chow’s beginner’s guide to forcing, and recent AI‑assisted work on Erdős Problem 1196, where the explanatory follow‑up paper was deemed more valuable than the proof itself. Proposed actions: require PhD students to present clear explanations of results, identify “open exposition problems” analogous to Hilbert’s problems, create journals focused on exposition, and adjust hiring/tenure criteria to reward high‑quality textbooks and exposition. The goal is to shift cultural incentives, reassure future mathematicians, and reinforce mathematics as a public good.

The San Francisco Onion Futures Company sells private, transferable contracts for future physical delivery of yellow onions, with contracts available up to six months in advance at prices of $3–$12 per onion. Pricing is updated daily at midnight (San Francisco time) via a proprietary algorithm and can be viewed only with JavaScript enabled; purchases are processed through Stripe. Delivery is arranged in the first week of the contract’s delivery month; the company contacts the registered owner, offers scheduled times and locations in San Francisco, Toronto, Seattle, and Chicago, and may use third‑party services at no extra cost. Contract owners receive a unique key that can be transferred; the new holder updates contact details, voiding the previous key. No refunds or cash settlements are provided, though contracts may be resold until delivery begins. The operation is claimed to be legal under 7 U.S.C. § 13‑1 because it is a private sale, not a board of trade. Contact is via [email protected]; the firm is owned by Sophia Vysparov (homepage a11ce.com).

A Mastodon post titled “GrapheneOS: ‘Android 17 QPR1 is the first release since Androi…’” appears on the GrapheneOS Mastodon instance. The page includes standard Mastodon interface text prompting users to enable JavaScript for the web application and references platform-specific apps. An image is present with alt text simply stating “Mastodon.” No additional technical details, announcements, or discussion are provided in the scraped content.

The repository implements a low‑cost macOS computer‑use agent that follows plain‑English goals by repeatedly capturing the screen, performing OCR (Apple Vision), extracting actionable UI elements from the Accessibility (AX) tree, and selecting the next action with a TypeSafe decision model (up to 255 options, returning a calibrated probability distribution in a few hundred milliseconds). Only when a text field truly requires free‑form input does it invoke a writer LLM; otherwise actions are deterministic clicks, scrolls, or browser navigation, keeping per‑step cost around $0.0002. The loop stops on completion, low confidence (< 0.4), repeated no‑ops, or step limits, then uses a stronger writer model to produce a final answer. The codebase (Python 3.12+, uv) includes modules for macOS Quartz/AppleScript integration, perception (capture, OCR, changed‑tile caching, block merging), date parsing, decision making, action handlers, runner, logging, and CLI tools (clicker, clicker‑inspect). Runs are logged with screenshots, payloads, and timing breakdowns. A Linux port swaps macOS APIs for xdotool/AT‑SPI and Vision OCR for PaddleOCR/RapidOCR. MIT‑licensed.

The author argues that contemporary scientific research is inseparable from open‑source software and that such software should be considered a core component of computational science. Science is defined as a systematic discipline producing testable, reproducible hypotheses; reproducibility therefore requires not only published results but also executable, modifiable code that encodes predictive models. Bugs in research software have already caused retractions, demonstrating that unreliable code undermines scientific validity. Open‑source licenses provide two essential properties: (1) reproducibility—code can be run in identical environments (e.g., via containers or NixOS) and adapted by others; (2) reliability—public scrutiny allows errors to be detected and corrected. A future vision is proposed where every paper links to runnable analysis, software evolves collaboratively like Wikipedia, and transparent code strengthens public trust. The author urges researchers to (a) publish well‑documented code with papers, (b) use reproducible environments (NixOS, Docker, Conda), (c) leverage existing tools, and (d) recognize software contributions in academic promotion.

Quick Tunnels, a Cloudflare service, provides a single‑command tool that exposes a local server (e.g., on a laptop) as a public, encrypted URL hosted on Cloudflare’s edge network. The feature requires no Cloudflare account, DNS configuration, or open inbound ports. It includes JSON‑formatted output for programmatic use. Cloudflare advertises global coverage in over 335 cities, typical connection latency of about three seconds, and zero ports opened on the client side. The service is offered free of charge.

Rust LSPs must serve queries while the project is only partially indexed. After the client’s initialize handshake, the server must respond before any workspace discovery, then schedule indexing asynchronously. Because LSP treats documents, not the filesystem, a virtual file system (VFS) with source generations is required to keep in‑memory state consistent and to cancel in‑flight read queries when mutations occur. Queries differ in cost: hover and document symbols need only parsing and an item tree, while inlay hints and references demand semantic analysis, definition maps, and body‑level type inference. Rust‑analyzer uses salsa (an incremental, memoized database) with lazy computation and optional cache‑priming to pre‑compute the semantic layer, keeping bodies on demand. Rust Glancer favors eager full indexing, persisting state to disk for low RAM usage and instant restarts; it answers queries as soon as the relevant semantic data is ready, prioritizing the open file. Workspace discovery is ambiguous—clients may open multiple Cargo workspaces or files without Cargo.toml. rust‑analyzer eagerly discovers and indexes all reachable workspaces, while Rust Glancer requires an explicit Cargo.toml and indexes a workspace only after it is opened, avoiding unwanted background work.