HackerNews Digest

Claude Sonnet 5 is Anthropic’s latest Sonnet‑class model, positioned as the most “agentic” in the series. It can plan, browse, use terminals, and execute multi‑step coding tasks with performance near Opus 4.8 but at lower cost. Compared with Sonnet 4.6, it shows measurable gains in reasoning, tool use, coding, and knowledge work, while safety assessments indicate fewer undesirable behaviors, better refusal of malicious prompts, and reduced hallucination and sycophancy. Cybersecurity capability remains limited; the model cannot produce full exploits and only shows a slight rise in partial success over its predecessor, prompting default cyber‑safeguards identical to those in Opus 4.7/4.8. Pricing is introductory $2 / M input tokens and $10 / M output tokens through 31 Aug 2026, rising to $3/$15 thereafter, and it is the default on Free, Pro, Max, Team, and Enterprise plans, including Claude Code and the Claude Platform. Early‑partner feedback highlights reliable end‑to‑end task completion, efficient multi‑step software engineering, and consistent adherence to safety and cost constraints.

None

The U.S. Department of Transportation plans to replace the long‑standing civil prohibition on over‑land supersonic flight with a noise‑based certification standard, permitting aircraft to exceed Mach 1 provided their sonic signature stays below a prescribed limit. The change follows a June 2025 executive order by President Trump directing the FAA to repeal the ban, establish interim noise criteria, and remove other regulatory obstacles. The FAA aims to finalize the rulemaking by mid‑2027. Historically, the 1973 ban was enacted because 1960s sonic‑boom tests caused window damage, property loss, and widespread public complaints; consequently, aircraft such as Concorde were required to remain subsonic over U.S. territory. Current U.S. developers include Boom Supersonic, which has pre‑orders from United, American, and Japan Airlines for its 60‑80‑passenger Overture, and Spike Aerospace, building a 18‑passenger Diplomat. Both firms claim quieter booms and improved fuel efficiency, targeting trans‑Atlantic trips under four hours.

Copybara is a Google‑internal tool for transforming and syncing code between repositories, primarily supporting Git (with experimental Mercurial read support). It designates one repository as authoritative while allowing contributions from any repository and enabling releases from any of them. Core capabilities include: - Importing sections of code between confidential and public repos, and vice‑versa. - Propagating changes from non‑authoritative to authoritative repos with automatic conflict handling. - Stateless operation: state is stored as a label in destination commit messages, ensuring consistent results across users. Configuration uses a Skylark‑style workflow (e.g., `core.workflow`) specifying origins, destinations, file globs, authoring, and transformations (replace, move). Installation options: - Use weekly snapshot binaries (requires Java 21+, add `--java_runtime_version=remotejdk_21` to `.bazelrc`). - Build from source with JDK 11 and Bazel (`bazel build //java/com/google/copybara:copybara_deploy.jar`). - Add as an external Bazel repository via `http_archive` in `WORKSPACE`. - Run via Docker (experimental) with environment variables to set subcommand, config, workflow, etc. Documentation is evolving; contact the mailing list for queries.

Anthropic announced on X that the U.S. Department of Commerce has removed export‑control restrictions on its Claude Fable 5 and Mythos 5 language models. The company said it will start re‑enabling user access to these models the following day and will provide a further update once the restoration process is underway. Anthropic expressed appreciation for users’ patience during the restriction period and thanked the individuals and teams who assisted in redeploying the models. No additional technical details, timelines beyond “tomorrow,” or policy explanations were provided. The post included a user‑avatar image but no further visual content.

The Forestiere Underground Gardens in Fresno, California, are a subterranean complex constructed by Sicilian immigrant Baldassare Forestiere between 1906 and his death in 1946. Using hand tools and mules, Forestiere excavated ten acres of hardpan, creating three levels (≈10 ft, 20 ft, and 23 ft deep) and 65 interlinked rooms without formal blueprints. The space includes a summer bedroom, winter bedroom, bath, kitchen, fishpond, parlor with fireplace, grottoes, courtyards, and conical skylights that regulate airflow and temperature. Excavated soil and hardpan were repurposed for planters, bricks, and structural supports. The underground environment supports a diverse planting program—citrus, berries, kumquat, loquat, jujube, and grafted multi‑fruit trees—many over a century old, protected from frost and insulated by above‑ground vegetation. The gardens were listed on the National Register of Historic Places in 1977 (No. 916) and designated a California Historical Landmark in 1978. T. Coraghessan Boyle later fictionalized the site in a 1998 New Yorker short story.

Claude Science beta is an AI assistant designed for scientific workflows. It creates isolated compute environments for each specialist domain and records full provenance of all results. The platform includes built‑in analysis specialists for genomics, single‑cell sequencing, proteomics, structural biology, cheminformatics, and additional fields. It connects natively to over 60 scientific databases and domain‑specific open models. Integration leverages NVIDIA’s BioNeMo Agent Toolkit to access life‑science models and libraries in BioNeMo, including Evo 2, Boltz‑2, and OpenFold 3. The system enables pipeline execution, database navigation, cluster job orchestration, and session‑persistent tracking, extending AI capabilities beyond conversational discussion to fully reproducible computational research.

The page presents a Gemini 3.1 “Flash‑Lite” image titled “Nano Banana 2 Lite”. The primary prompt describes a top‑down, abstract coastline split diagonally by land and sea, with navy‑blue surf turning to white foam, a narrow pale sand strip, aquamarine water, and rugged ochre‑beige terrain, emphasizing color, texture, and diagonal composition rather than realistic perspective. The accompanying visual assets consist of 22 alt‑text entries: - Macro and wildlife photographs (spider with droplets, spiny lizard on bricks). - Action sports shot (male swimmer in butterfly stroke). - Conceptual illustrations (man on books viewing a paper‑wave tidal wave, minimalist icons for low‑latency, cost‑efficiency, high‑quality on light/dark blue backgrounds). - Numerous side‑by‑side comparisons of images generated by Nano Banana 2 and Nano Banana 2 Lite, covering themes such as astronaut in futuristic corridors, aerial coastlines, hawk, black‑and‑white bird over pine valley, glass sphere on water, hair transforming into birds, and a vintage sedan in fog. - Footer graphics for “gemma” in light and dark variants.

The passage explains how gears provide mechanical advantage by multiplying torque and altering speed. In the context of a wind‑up toy for children, a stiff spring requires substantial torque to wind. Incorporating gears reduces the force a child must apply, allowing a modest input force to generate the necessary torque to compress the spring. This is achieved by using gear ratios that increase torque while decreasing the input speed, making the winding process easier for young users. The discussion emphasizes gears’ role in translating small forces into larger torques, thereby improving usability of torque‑intensive mechanisms in consumer products.

The post presents WGO‑Bench, a new benchmark for robotic subtask annotation comprising 100 egocentric and robot‑video episodes with 743 human‑annotated segments across 62 high‑level instructions. Over 60 experiments identified the most effective pipeline: subtask boundary detection (0.306 F1), subtask labeling (61.0 % accuracy), and end‑to‑end annotation (0.168 F1). Gemini models, especially Gemini 3.5 Flash, outperform the best non‑Gemini model (GPT‑5.5) by 24.5 %. The optimal method uses visual “contact sheets” that pack 20 frames with embedded timestamps, achieving 0.263 F1 for segmentation while reducing inference cost to $2.64 per video hour (≈19× cheaper than human labeling). Direct frame‑by‑frame prompting reaches only 0.193 F1 and is far more expensive. Experiments show visual timestamps on contact sheets markedly improve boundary detection, whereas higher resolution, denser sampling, or larger sheets do not. The complete open‑source pipeline is available in Refiner for users to apply to their own robot videos.