HackerNews Digest

• The author evaluated GPT‑6 “Astra”, noting its impressive multimodal abilities but poor suitability for practical software engineering.
• A “software factory” experiment let Astra manage its own workflow, spawning sub‑agents and storing notes. Over 35 hours it consumed ~4 billion tokens (≈1 B tokens for the run), costing about $1,200, producing 75 k lines of code, 79 commits, ~1 400 messages, and $15.5 per commit.
• Astra heavily favors Python for tool calls, often generating convoluted, token‑efficient scripts (e.g., extensive pathlib manipulations, socket‑FD passing, and auto‑generated C extensions) that are unreadable and deviate from CPython style.
• The model repeatedly inserted low‑level code, random constants, and bizarre control‑flow constructs, suggesting training rewards prioritize token efficiency and task completion over human readability.
• Despite successful reverse‑engineering demos, the output frequently consisted of “slop” – code that technically runs but is objectively poor for maintainable projects.
• The author concludes that Astra’s current trajectory inflates costs and produces output misaligned with conventional software‑engineering practices, questioning the value of such models for developers.

Shopify migrated to React Native in 2020 to avoid duplicating feature work, enable cross‑stack developers, and reduce parity maintenance. By 2025 LLM‑based agents dramatically lowered the effort required to implement and test equivalent Swift and Kotlin code, prompting a reassessment of the mobile stack. Prototypes showed agents could translate features between iOS and Android, maintain parity through shared specs and tests, and rebuild core apps faster than before. Consequently, Shopify is transitioning back to native development, using a greenfield approach: the Shop app was rebuilt and released in 12 weeks, and the larger Shopify app (300+ screens, widgets, watch extensions) is slated for release later this year, with remaining apps to follow. Migration is managed by the Helix system, which breaks work into test‑validated checkpoints, and a CLI that decouples business logic from UI for millisecond‑scale feedback loops. Open‑source React Native libraries (FlashList, Restyle) will be supported through 2026 or archived, with stewardship hand‑off planned. Future success will be measured by product velocity, app quality, and autonomous agent contribution.

The team at Conviva built a Rust query engine on DataFusion, Arrow, Rayon and Tokio that reads large (3–5 GB) Arrow IPC files stored on NVMe. Using mmap gave zero‑copy reads but under heavy concurrent query loads the shared Linux page cache thrashed: RSS approached 99 % of RAM, major faults rose to >1 k /s, minor faults exceeded 2 M /s, and kernel lock contention caused p95 latency to jump from ~30 s to >150 s. A controlled test showed a single pod outperforming four pods by >20 % because all pods competed for the same page‑cache.

To avoid the page‑cache bottleneck they switched to io_uring with O_DIRECT via the compio Rust wrapper, issuing a separate read for each of ~40 columns per day‑long query. Initial runs on Linux cut major faults 35× but increased minor faults 8× and slowed total query time from 13.6 s (mmap) to 21.8 s. The slowdown stemmed from:

• a single “Batch Materialization Layer” handling prefetch, submission, decoding and caching on one thread,
• Arrow’s default Buffer::from_slice_ref copying data and causing many minor faults,
• insufficient parallelism (40 concurrent SQEs) and extra memcpy work.

After removing the copy (building Buffers directly from io_uring buffers) and enabling O_DIRECT, runtime improved to ~16 s—still ~20 % slower than mmap—highlighting that io_uring’s gains are negated by architectural overheads and memory‑management costs.

The Gemini app for Windows launches for Windows 10 and 11, providing on‑desktop AI assistance via three primary modes:

• Quick access shortcut: Press Alt + Space to invoke Gemini over any active window for instant queries, fact‑checks, or idea generation without leaving the workflow.
• Dedicated workspace: A full‑screen interface lets users assign multi‑step tasks to Gemini Spark, retrieve data from Google services (Gmail, Drive), and generate project summaries.
• Creative generation: Integrated tools enable image creation with Nano Banana and video production with Gemini Omni, all within the same desktop environment.

The app is described as lightweight, non‑resource‑intensive, and globally available for download at gemini.google/desktop. Future updates will add more native desktop capabilities.

The page is titled “Notion” and displays a notice that JavaScript must be enabled to use the site, instructing users to enable it to continue. No functional content is presented without JavaScript. The only visual element referenced is an image with the alt text “Notion,” indicating that the page includes a graphic representing the Notion brand or interface. No additional text, links, or interactive features are described in the scraped excerpt. The content therefore consists solely of a JavaScript dependency warning and a single branded image placeholder.

The post recounts the author’s reaction to a meme shared by Tyler Gaw, which humorously highlighted the long‑term usefulness of a “Big Box of Cables” after ten years of storage. Inspired, the author documented the meme (screenshot, printout) and physically attached it to his own cable storage container, labeled “FAMILY TECHNO BOX.” He describes the steps taken—cutting clear tape, affixing the printed meme—to create a visual reminder of the box’s value and to discourage family members from discarding it. The author intends the taped message to serve both as personal motivation when adding new cables and as a caution to others about throwing away seemingly obsolete hardware. An accompanying image shows the box with cables protruding and the handwritten label, alongside the taped meme.

Git worktrees provide additional working directories attached to a single repository, sharing the object database, refs, stashes, and remotes while giving each worktree its own HEAD and index. They allow parallel development on multiple branches without repeated check‑outs, useful for long test runs, PR reviews, or isolating AI coding agents. Creating a worktree is fast (git worktree add <path> -b <branch>); a branch can be checked out in only one worktree at a time. Untracked files (e.g., node_modules, build caches) must be set up per worktree, which can be a downside for large projects. Magit integrates worktree management via commands such as magit-worktree-checkout, magit-worktree-branch, magit-worktree-status, magit-worktree-delete, and can display all worktrees in the status buffer with a hook. Each worktree is treated as a separate project by Emacs tools like project.el or Projectile. Deleting a worktree should use git worktree remove or Magit’s Z k; leftover bookkeeping can be cleaned with git worktree prune. The article also notes Jujutsu (jj) as an emerging VCS that mitigates worktree limitations by treating the working copy as an auto‑snapshotted commit.

The OpenAI Agents API provides programmatic access to a managed Codex harness that runs agents in sandboxed environments. Core concepts are Agent (model, instructions, tools, MCP servers), Environment (optional sandbox or self‑hosted workspace), Session (a durable instance handling multiple turns), and Events/Items (the inputs and outputs exchanged. OpenAI‑hosted sandboxes provisioned by the service handle code execution, file edits, MCP connections, and artifact creation; self‑hosted environments can be specified with a workspace directory and capability paths. Billing follows the selected model’s API rates and standard container rates for sandbox usage. Agents can use tools such as programmatic tool calling, MCP access, and web search, and support multi‑agent operation with up to four concurrent sub‑agents. Sessions retain state across turns, allow steering, context summarization, sub‑task delegation, and can be deleted when finished. Data residency is limited to the United States; Zero Data Retention is not supported, even with self‑hosted sandboxes. The API is accessed via SDKs or direct HTTP calls, with example payloads provided for JavaScript, Python, Go, Java, Ruby, and cURL.

The author demonstrates a denial‑of‑service vulnerability that can be triggered on macOS by loading a malicious WebGPU shader from an untrusted website. The shader consists of a compute stage with an infinite busy‑loop that repeatedly copies data within a storage buffer, while a vertex shader reads from the same buffer, causing the GPU pipeline to stall. Because macOS’s WindowServer also relies on GPU resources, the stall propagates, rendering the desktop UI unresponsive and eventually invoking the system watchdog, which may trigger a kernel panic or require a forced power‑off. The effect reproduces on Chrome, Firefox and Safari on M‑series Macs but not on other operating systems. A similar WebGL‑based attack (ShadyShader, CVE‑2023‑40441) was previously disclosed and partially mitigated by Apple through input‑validation improvements, but WebGPU lacks comparable safeguards. The author notes that GPU pre‑emption on Apple silicon is handled by the ASC coprocessor firmware, limiting the kernel’s ability to abort runaway shaders, and that Apple has classified the issue as low‑priority.

The decorrelation‑stretch algorithm, originally developed at NASA JPL for enhancing satellite data, maps original colors into an expanded range via matrix multiplication and diagonalization of color matrices—a process rooted in the Karhunen–Loève Transform. Ronald Alley’s 1996 paper refined the method for the ASTER multispectral imager, enabling faster, more accurate color enhancement used today to monitor volcanic plumes and map lava flows.

Jon Harman adapted the technique into the Dstretch plug‑in for ImageJ and later Android/iOS apps, leveraging his background in medical imaging and algebra. Dstretch has uncovered dozens of previously invisible rock‑art figures at sites such as Angkor Wat, Baja California, Norway’s Årsand 1, Egypt’s Beni Hassan, Canada’s Writing‑on‑Stone Provincial Park, and Greek urban settlements at Vlochos. It also revealed tattoos on mummified remains and aided archaeological surveys where conventional geophysical methods fail. Harman provides paid processing services (≈200 requests yr⁻¹ at $50 each) and the apps have been downloaded thousands of times, demonstrating broad utility beyond its space‑technology origins.