HackerNews Digest

Austin added 120 000 housing units (≈30% growth) from 2015‑2024, driven by zoning changes (Vertical Mixed‑Use, targeted rezoning, density‑bonus programs), ADU reforms, citywide parking‑requirement elimination, and municipal bonds ($250 M in 2018, $350 M in 2022). These policies sped permitting (Site Plan Lite, Infill Plats, AI pre‑check) and eased height, lot‑size, and compatibility rules, while the HOME initiative (2023‑24) expanded duplex/triplex/ADU construction and introduced single‑stair‑way midrise apartments. Rents fell after a 93 % surge in the 2010s: median rent dropped from $1 546 (Dec 2021) to $1 296 (Jan 2026), 4 % below the U.S. median. Large‑apartment rents fell 7 % (2023‑24); Class C (older, non‑luxury) units fell 11 %; inflation‑adjusted rents fell 19 % versus a national +10 % rise. Affordable‑housing output rose to 4 605 units in 2024 (double 2023) via density bonuses and bond funding. Housing composition shifted: <50 % of Austin units are single‑family homes (vs 71 % U.S.), with apartments accounting for half of new units. Yet an under‑production gap of ~23 000 units (2022) remains, prompting further reforms.

Cook defines a lightweight DSL for orchestrating AI‑agent tasks. Its primitives are **work** (a single prompt), **loop operators** (repeat xN, review, ralph) and **composition operators** (versions vN/race, vs, pick/merge). Loop operators wrap preceding work left‑to‑right: `xN` repeats a task N times passing prior output forward; `review` adds a gated review step that iterates until a “DONE” condition is met; `ralph` provides an outer gate that advances through a task list, resetting iterations after each successful step. Composition runs parallel isolated git worktrees: `vN` spawns N identical cooks with a default “pick” resolver; `vs` runs distinct cooks side‑by‑side, each may contain its own loops, and results are merged via `pick` or `merge` criteria. Configuration is scaffolded with `cook init`, producing `COOK.md`, `.cook/config.json` (agent/model defaults, per‑step overrides, env vars), a Dockerfile, and a log directory. Execution can use an agent sandbox (`--sandbox agent`) or Docker sandbox (`--sandbox docker`), the latter required for OpenCode.

The repository implements an autonomous AI agent (e.g., Claude Code) that self‑trains to become the leading MaxSAT solver using 229 weighted MaxSAT instances from the 2024 MaxSAT Evaluation (anytime weighted track). The agent reads program.md for execution instructions, expert.md for accumulated knowledge, and a library directory containing Python solver modules (e.g., solvers.py, greedy_sat, tabu_search, walksat_hard/soft, core_guided, clause_weight_ls). It runs solvers on each instance, logs experiments, updates best‑solution files, and commits changes to the GitHub repository, enabling multiple agents on separate VMs to collaborate via git pull/push without explicit coordination. Deployment is scripted (run.sh) for EC2, requiring a .env file with CLAUDE_CODE_API_KEY and GITHUB_ACCESS_TOKEN. The agent autonomously discovers new strategies, improves solutions (e.g., reducing instance pa‑1 cost from 5445× to 612×), and expands its toolbox, though it exhibits low parallelism (1–6 concurrent scripts), can fixate on single hard instances, and typically stops after a few hours despite “never stop” prompts. Nine large instances remain unsolved.

Tom Hartmann, a former agricultural equipment technician, became a “Software Mechanic” after the post‑transition shift where software is regenerated from plain‑language specifications rather than repaired. In this new economy, expertise lies in the domain (farming, medicine, etc.) and in diagnosing gaps between specifications and generated code. Key case studies illustrate recurring problems: - **Margaret Brennan** lost $25 k because a weather‑model update altered crop‑maturity inference; the spec lacked a clause to detect upstream data changes (“ground moved” issue). Tom added a monitoring clause, charging $180 for the fix. - **Ethan Novak** suffered $14 k losses when a regenerated feed‑tool changed output format, causing a pricing tool to misparse data. Tom patched the spec and recommended a “software choreographer” to map and validate all tool interfaces, a service that ultimately prevented further failures. - **Carol Lindgren** resisted a grandson‑installed irrigation optimizer. Tom confirmed it saved ~15 % water but highlighted conflicts with her 30‑year tacit knowledge. He installed a manual override switch and logged overrides, preserving her control while retaining automation benefits. The narrative emphasizes that software specifications must be explicit, continuously monitored, and integrated with domain expertise; maintenance (pit‑crew or choreography) is economically cheaper than failure but often resisted due to psychological bias toward emergency fixes.

The page is a GitLab repository titled “nvidia_greenboost” owned by user Ferran Duarri. The title indicates the project is hosted on GitLab, a platform for source‑code management and collaboration, and that its name combines “nvidia,” suggesting a relation to NVIDIA hardware or software, with “greenboost,” implying a focus on performance or efficiency enhancements. No additional description, documentation, or code details are provided in the excerpt, so the only concrete information is the repository’s name, its hosting service, and its owner’s GitLab account.

OpenRocket is a rocket‑design and simulation tool that provides both 2‑D and 3‑D visual editors for building models. It includes a parts library, motor selection dialogs, and a simulation engine capable of generating flight plots, multi‑level wind profiles, and scripted runs. Additional utilities comprise PhotoStudio for rendering, a Rocket Optimizer for design refinement, a Component Analyzer for performance metrics, and export functions for data sharing. The project is released under a Creative Commons BY‑SA license. Community interaction is facilitated through a Discord server where users can discuss designs, ask technical questions, and communicate with developers.

Rob Pike’s five programming rules emphasize pragmatic performance and simplicity: 1. **Unpredictable bottlenecks** – don’t assume where a program will spend time; avoid premature speed hacks. 2. **Measure before optimizing** – only tune after profiling and when a single part dominates runtime. 3. **Avoid fancy algorithms for small n** – complex algorithms have large constant factors; use simple approaches unless large inputs are proven. 4. **Prefer simple algorithms and data structures** – they are less error‑prone and easier to implement. 5. **Data structures dominate** – well‑chosen structures and organization make algorithms evident; focus on data, not clever algorithms. Pike’s first two rules echo Hoare’s “premature optimization is the root of all evil.” Thompson paraphrases rules 3‑4 as “when in doubt, use brute force,” reflecting the KISS principle. Rule 5 parallels Brooks’s advice from *The Mythical Man‑Month*: write straightforward code that leverages robust data abstractions.

A Wander console enables users to browse random sites and pages contributed by the Wander community, a network of personal websites. While a user can switch to another console on a different domain, the current console can also retrieve recommendations recursively from linked consoles without leaving the page. To host a personal Wander console, one must download the provided ZIP file, extract `index.html` and `wander.js`, place them in a `/wander/` directory on the web server, and edit `wander.js` according to the instructions on codeberg.org/susam/wander. After deployment, the console’s URL can be shared in the community thread; other users may add it to their consoles, integrating it into the broader Wander network. Additional documentation and source code are available at codeberg.org/susam/wander.

The article explains that the prevalence of bell‑shaped (normal) distributions in empirical data stems from the central limit theorem (CLT). Originating with Abraham de Moivre’s work on gambling outcomes, the theorem was formalized by Pierre‑Simon Laplace and states that the sum or average of a large number of independent, identically distributed random variables converges to a normal distribution, regardless of the original distribution’s shape. This property underlies the regularity observed in measurements such as human heights, weights, test scores, and physical phenomena, because many underlying factors act additively and independently. The CLT enables statisticians to assess deviations from expected randomness (e.g., detecting a biased coin) and to construct confidence intervals without detailed knowledge of the underlying process. Limitations arise when samples are not independent or when extreme outliers dominate, prompting the use of specialized variants of the theorem or alternative models for rare events. Consequently, the CLT remains a foundational tool across scientific disciplines for inference based on aggregated data.

NVIDIA NemoClaw is an open‑source plugin that installs the NVIDIA OpenShell runtime to run OpenClaw agents inside a sandboxed environment. The sandbox enforces security policies (Landlock, seccomp, network namespaces) and routes inference calls through NVIDIA cloud APIs, e.g., the nemotron‑3‑super‑120b‑a12b model. Installation is performed via a single script (`curl … | bash`) which also installs Node.js if needed, then runs an onboarding wizard to configure the sandbox, inference provider, and security policies. The compressed sandbox image (~2.4 GB) requires ≥8 GB RAM or sufficient swap to avoid OOM termination. Primary CLI commands include `nemoclaw connect`, `status`, `logs`, and OpenShell subcommands for sandbox management. Interaction with the agent can be done through the OpenClaw TUI (`openclaw tui`) or CLI (`openclaw agent …`). Local inference options (Ollama, vLLM) are experimental and may need additional host routing support. Known limitations note ongoing development of the OpenClaw‑NemoClaw plugin and potential platform‑specific workarounds. The project is released under the Apache License 2.0.