HackerNews Digest

Container machine delivers a persistent, lightweight Linux environment on macOS using standard OCI images. It automatically maps the host username and home directory ( $HOME → /Users/ in‑container) so macOS editors and IDEs can edit files that are compiled and run inside the container. macOS tooling (profilers, browsers, debuggers) accesses the same files without copying. Systemd‑based images support real Linux services (e.g., systemctl start postgresql). Multiple machines can target different distros (alpine, ubuntu, debian) while sharing the same home and dotfiles. Key CLI operations: - `container machine create --name ` - `container machine run [-n ] [command]` (interactive shell or one‑off command) - `container machine set-default ` - `container machine ls`, `inspect`, `stop`, `rm` - `container machine set -n cpus= memory=` (applies after restart) - Home‑mount mode: `rw`, `ro`, or `none`. Custom images must include `/sbin/init`; a sample Ubuntu 24.04 Dockerfile installs systemd and common tools. Adding `/etc/machine/create-user.sh` lets a custom script run on first boot with environment variables CONTAINER_UID, CONTAINER_GID, CONTAINER_HOME, CONTAINER_USER, CONTAINER_MACHINE_ID.

Claude Fable 5, a Mythos‑class model, is released for general use with state‑of‑the‑art performance across software engineering, knowledge work, vision, long‑context memory, and life‑science research. Benchmarks show it leads other Claude models on coding (FrontierCode, CursorBench), finance (Hebbia), vision (ViBench, Pokémon FireRed), and scientific tasks, often requiring fewer tokens and turns. To mitigate misuse, Fable 5 employs conservative safety classifiers that redirect queries on cybersecurity, biology/chemistry, or model‑distillation to Claude Opus 4.8; fallback occurs in <5 % of sessions. Claude Mythos 5 uses the same underlying model but lifts cybersecurity safeguards for a limited group of cyber‑defenders via Project Glasswing and will later expand through a trusted‑access program for biology researchers. Pricing is $10 per M input tokens and $50 per M output tokens, half the cost of Claude Mythos Preview. Availability: Fable 5 is open via the Claude API; Mythos 5 is currently restricted to Glasswing partners with planned broader trusted access. Data from these models is retained for 30 days for safety analysis only.

npm v12 (scheduled for July 2026) changes default install behavior to improve security. Scripts (preinstall, install, postinstall, node‑gyp rebuild, prepare) are disabled by default; only packages explicitly allowed via `npm approve-scripts` remain executable, with the allowlist stored in `package.json`. Git dependencies are blocked unless `--allow-git` is used, closing a code‑execution vector where a dependency’s `.npmrc` could override the Git executable. Remote URL dependencies (e.g., HTTPS tarballs) are also blocked by default and require `--allow-remote`; the defaults for `--allow-file` and `--allow-directory` stay unchanged. These defaults are already available as warnings in npm 11.16.0+ (or later) so developers can test, review blocked items with `npm approve-scripts --allow-scripts-pending`, approve trusted packages, and commit the updated allowlist before upgrading. Documentation and community discussion links are provided for further guidance.

Richard Sutton posted on X a brief announcement of a new video in which he argues that generative AI systems trained solely by supervised learning are fundamentally unable to produce genuinely novel discoveries. He labels the viewpoint “possibly controversial” and provides a link to the video (https://t.co/LhAU6AyDkh). The accompanying text includes the opening line of the speech, “Good day ladies and…,” indicating a formal presentation. An image is attached, showing a user avatar as the visual element of the post. No further technical details, data, or arguments are included in the scraped excerpt.

A Munich Regional Court ruled that Google is directly liable for statements made in its AI‑generated search overviews, treating them as the company’s own content rather than merely linking to third‑party sites. The court found that the overviews falsely linked two local publishers to scams and other illicit practices, creating new, self‑contained assertions not present in any cited source. Existing German jurisprudence granting limited liability to traditional search engines was deemed inapplicable because AI overviews generate independent statements that only Google can verify. The court rejected Google’s defense that users could check sources, noting that users rarely follow links and that the overviews are understandable on their own. Consequently, Google cannot rely on host‑provider protections under the Digital Services Act or the notice‑and‑take‑down regime. The injunction bans the specific false claims, orders Google to cover 80 % of legal costs, and may set a precedent with international relevance, highlighting potential liability for AI‑driven services that produce unverified paraphrases.

Lotte Reiniger, a 26‑year‑old German filmmaker, created the world’s oldest animated feature, *The Adventures of Prince Achmed*, by employing a minimal crew. She personally crafted silhouette puppets from cardboard and lead, joined with wire hinges, and operated the camera with her husband’s assistance. Reiniger directed, designed the film’s scenario—melding several Middle Eastern fairy tales—and oversaw frame tracking. The production used stop‑motion animation on flat silhouettes placed on a glass plate lit from below, a technique that allowed intricate, fluid motion without conventional drawing. Despite limited resources, Reiniger’s hands‑on approach resulted in a landmark work that predates Disney’s early features by a decade, establishing a foundational method for silhouette animation and influencing future cinematic storytelling.

The author describes a recent 48‑hour hackathon in Vilnius where their two‑person team retrofitted an old rotary phone with a Raspberry Pi, exposing its I/O and linking it to a server via a single WebSocket. The server handled two‑way audio, custom bell tones, and hang‑up control, while an AI agent used the Spotify API to generate playlists from natural‑language prompts (e.g., “play music by artists alleged to be on the Epstein files,” “create a 1970s Zambian psychedelic rock playlist”). Voice output was provided by ElevenLabs, giving the phone a “warm Yorkshire gentleman” persona. The team relied on AI‑generated code, noting that modern hackathons prioritize system architecture and rapid integration over manual coding. The author argues that as software problems become “solved,” future hackathons will emphasize hardware and legacy tech, proposing projects such as Apple II applications, fax‑to‑social‑media converters, Game Boy Advance Bloomberg terminals, LLM‑driven cash registers, and AI‑controlled microwaves. The piece calls for “ridiculous”, overbuilt prototypes that blend APIs, wiring, and consumer electronics.

The article outlines hardware architectures for ultrafast inference and online learning on FPGAs using Kolmogorov‑Arnold Networks (KANs). Fixed‑point quantization encodes real‑valued neural computations as binary functions, enabling lookup‑table (LUT) implementation of univariate activation functions. By representing each KAN edge activation as a separate LUT, the design avoids exponential growth of LUT entries that occurs with multivariate functions, and leverages B‑spline bases that are local (only S + 1 active basis functions per interval) and bounded (∑B_i(x)=1), which simplifies fixed‑point range selection and stabilizes gradient updates. Inference proceeds by parallel LUT reads for activations followed by an adder tree for the summed outputs; multi‑layer networks cascade these blocks. For online learning, B‑spline values remain in LUTs while coefficients are updated in‑place on the FPGA, allowing forward, backward, and weight‑update passes to run at sub‑microsecond latency. Reported results show up to 2700× speedup over prior KAN‑FPGA implementations, scaling to >50 k parameters with near‑constant resource usage and improved convergence on benchmarks such as function approximation, qubit readout, and non‑stationary control.

The post surveys “molly guards”—physical or UI mechanisms that prevent accidental activation of controls. It begins with industrial examples, such as a perspex‑covered power button on an IBM electronic typewriter and softer guards that block access without removal. A red “writing to SD card” indicator is placed beside the card slot to deter ejection during write operations, and a floppy‑drive head‑lowering handle simultaneously secures the disk. The article also notes UI‑level guards that render actions disabled or require confirmation, citing Finder’s warning when opening many files, iPhone’s slide‑to‑unlock used only for silencing alarms, and Chrome’s press‑and‑hold‑Q prompt. Early iTunes featured a skeuomorphic CD‑burning guard, while a university alumni magazine displayed a literal “Molly” with her father as a visual gag. Across hardware and software, these designs employ visual cues, physical barriers, or interaction constraints to reduce user error.

The author evaluated Claude 5 Fable, the first Mythos‑class AI model, across a range of tasks. In repeated experiments Fable consistently outperformed other publicly available models, handling long, multi‑hour sessions and generating complex outputs such as a scholarly social‑science paper from a single prompt and a 10‑page all‑S rhyming poem. Notable demonstrations include: - Generation of playable games (e.g., a coin‑flip variant of “Balatro,” a self‑aware snake game) using only mathematical descriptions, as Claude cannot produce images. - Construction of an isochrone map: Fable coordinated dozens of subsidiary Claude Sonnet agents to retrieve over 2,200 flight records, global rail schedules, and road‑speed data, then coded, tested, and refined the visualization, iterating on remote‑location estimates. - Development of “Concord,” a 19‑page software design that calibrates human and AI judgments across datasets and performs statistical analysis, produced after a 9.5‑hour autonomous run. Limitations noted are high token consumption (approximately twice Opus cost) and frequent guard‑rail fallback to Claude 4.8 Opus. The workflow shifts human involvement from step‑by‑step control to high‑level specification and final approval, raising concerns about transparency and the evolving human‑AI relationship.