HackerNews Digest

Naphtha shortages, intensified by the Iran war, are disrupting Japanese manufacturing. Calbee, a major snack producer, announced on May 12 that 14 flagship products will shift to black‑and‑white packaging because the shortage limits ink and solvent supplies; mascot imagery will also be removed, with rollout beginning May 25. A Teikoku Databank survey identified 52 companies producing basic chemicals (ethylene, synthetic rubber, PVC) from naphtha, feeding a network of 46,741 manufacturers—about 30 % of Japan’s 150,000 surveyed firms. In the chemicals‑petroleum‑coal sector, 67.2 % of 4,700 companies rely on naphtha; 88.4 % of cyclic‑intermediate producers (used for plastics, fibers, dyes, pharmaceuticals, cosmetics, agro‑chemicals) are integrated, as are 87.3 % of gelatin/adhesive makers and 84.0 % of surfactant producers. Related impacts appear in pulp‑paper (80.1 % of coated‑paper producers), food packaging, medical supplies (syringes, gloves), insulation, and beverage containers. Other firms responding include Mizkan (suspending nattō containers) and Nisshin Seifun Welna (plain spaghetti tape). Continued Middle‑East conflict could broaden these shortages across daily‑life products.

The article introduces a “dead economy theory” extending the “dead internet” idea: massive AI deployment replaces human labor across the global workforce, creating a feedback loop that erodes the very market that sustains AI firms. Key points: - AI giants (OpenAI, Anthropic, Google DeepMind, Meta AI, Microsoft) have invested hundreds of billions, valuing the sector at $800 billion+ and targeting the entire labor market as the addressable base. - Benchmarks (OpenAI’s GDP‑Val, AI Productivity Index) claim AI can outperform humans in many professional tasks, suggesting a path to full cognitive‑labor automation. - The “AI layoff trap” model shows firms cut costs by automating, boosting margins and share prices, but displaced workers reduce aggregate demand, eventually harming the firms’ own customer base. - Economists warn this creates a prisoners‑dilemma: each firm’s incentive to automate exceeds the collective cost of demand contraction, accelerating a race toward “excessive automation.” - Historical analogues (industrial and agricultural revolutions) indicate short‑run displacement can last decades; AI’s speed may compress this to a few years, risking widespread precarity and democratic instability. - The author argues current policy proposals (UBI, retraining) ignore structural loss of labor‑based tax revenue and civic leverage; suggested remedies include public ownership of AI infrastructure, aggressive antitrust, and higher taxes on automated capital.

None

Snowboard Kids 2 has reached 100 % decompilation: every original N64 MIPS function now has a matching C implementation that assembles to the same machine code. The effort spanned under two years, beginning with the first commit in September 2024, and concluded while the author was hospitalized with a newborn. Community support from the N64 decompilation Discord (notably Bl00D4NGEL, inspectredc, SlaveOfIDO, queueRAM) was essential, and coding agents—Claude, GLM, and Codex—accelerated the final ten functions, with Codex 5.5 xhigh proving most effective for the hardest cases. The next phase is a high‑quality recompilation, already functional but still containing bugs, widescreen adjustments, and extended draw distance artifacts. Remaining decompilation tasks include renaming functions, cleaning up structures, and extracting graphics/audio assets. The author also contemplates a Snowboard Kids 1 decompilation to enable a combined “Super Snowboard Kids” version on the newer engine. Further information and contribution tasks are listed in the project README.

Perry v0.5.306 is a TypeScript-to-native compiler that produces GUI and CLI binaries for macOS, iPadOS, iOS, Android, Linux, Windows, watchOS, tvOS, WebAssembly, and the Web. It eliminates runtime dependencies such as Electron, delivering native performance with a generational garbage collector and lazy JSON tape handling, which benchmarks faster than Node.js and Bun on most tests. A single TypeScript codebase can target all supported platforms, generating compact executables (e.g., a 2.3 MB binary for a “Hello, World!” program). The tool is invoked via `perrycompile`, producing ready‑to‑run native files without additional packaging steps.

Math‑To‑Manim (M2M2) is a pipeline that converts math‑or physics prompts into fully inspectable Manim explainer videos and accompanying artifacts. Each run follows a fixed stage chain—IntentAgent → PrerequisiteGraphAgent → CurriculumAgent → MathAgent → StoryboardAgent → SceneSpecAgent → ManimCodeAgent → StaticReviewAgent → RenderAgent → VideoReviewAgent → PublisherAgent—producing JSON contracts, a knowledge graph, curriculum, math packet, storyboard, scene spec, generated Python code, validation reports, render results, and review notes. The system supports audience‑aware requests, reverse‑reasoning from final concepts to prerequisites, and typed Pydantic artifacts between stages. A reinforcement‑learning “Prime Intellect” environment uses the generated scene spec, code, and validation evidence to train models to repair failing Manim code; rewards check parseability, safety, expected terms, and layout. The repository includes CLI commands for smoke tests, model‑driven generation (OpenAI or Codex), and rendering (Manim, FFmpeg, LaTeX). Artifacts are stored under runs// bundles, enabling traceability, debugging, and recursive editing via agents such as Hermes. MIT‑licensed.

Mistral AI presented a full AI stack—compute, models, platforms, and consultancy—centered on a 40 MW Paris data center with additional sites planned, including Sweden. Their focus is on efficient, open, on‑premise models, positioning them against Anthropic and OpenAI. Partnerships highlighted include collaborations with ASML, BNP Paribas, and Amazon Alexa+, emphasizing practical AI solutions over new model announcements; they introduced “Vibe for Work,” akin to Claude for Work. Key technical points: a “harness” adds context, persistence, and learning to models, enabling reasoning, error recovery, and transparency; skills capture organizational best practices. Mistral showcased specialized small models that outperform larger ones in energy efficiency and speed: Document AI for EU Patent Office OCR, Voxtral for multilingual voice (Alexa+ Europe), and Robostral for industrial robotics with ASML. On‑prem deployments enable data sovereignty, illustrated by BNP Paribas using Mistral for KYC and Abanca handling over 1 million customers via agent orchestration. A case study demonstrated fine‑tuning Codestral to transcribe ancient papyri, accelerating a 180 k‑document digitization project. Overall, Mistral aims to become Europe’s primary full‑stack AI partner, offering open models, on‑premise deployment, and enterprise collaborations.

MCP (Model Context Protocol) connects LLMs to external services but incurs significant overhead. In the Quandri stack, four MCP servers consume ≈10.5 % of the model’s context window; the Linear server alone loads 12,800 tokens (42 tool definitions) even when only “get_issue” or “save_issue” are used. Benchmarks show MCP calls are ~3× slower per request and ~9.4× slower on the first call because each server adds a processing layer. Token‑wise, a Linear issue lookup via MCP uses ≈12,957 tokens versus ≈200 tokens with a direct CLI curl command—a ≈65× increase. Alternatives proposed are: * **CLI‑first strategy** – invoke existing CLIs/APIs directly, eliminating tool‑definition context and simplifying debugging. * **Skills pattern** – embed concise CLI instructions in on‑demand “skills,” loading only the needed commands. MCP remains useful when no CLI exists, for non‑developer users, or for real‑time bidirectional interactions, and it can enforce query safety and credential protection for databases. Quandri now mixes Bash/CLI for common tools, Skills for repeatable workflows, and MCP only for services lacking robust CLIs. Updated Claude Code deferred loading reduces MCP context bloat by ≈85 % but the performance and architectural concerns persist.

Prusa has released an open‑source “Prusa ColorMix” model that enables multi‑material FDM printers to reproduce dozens of colors using only five filaments (Cyan, Magenta, Yellow, Black, White). The approach adapts traditional CMYK halftoning: colors are mixed by alternating thin layers rather than dots, relying on the eye’s limited resolution to blend adjacent layers. A calibrated Yule‑Nielsen‑based algorithm predicts resulting colors; systematic errors (darkening of light‑dark mixes, saturation loss, cyan hue drift) are corrected with lightness‑scaled adjustments derived from measured test cards. Calibration is performed per batch using a consumer colorimeter on five base filaments and 20+ mixes, ensuring consistency without expensive spectrophotometers. The model, licensed under MIT, is integrated into PrusaSlicer 2.9.6 and EasyPrint, with a dedicated CMYKW filament set (Prusament) and tools for creating and testing mixes. Source code and a TypeScript/C++ port are on GitHub, and the community is invited to contribute additional measurements for broader material support.

Shift, an AI‑training startup, offers free home‑cleaning services in exchange for video data of the cleaning process. Cleaners wear a “magic hat” equipped with a camera that records their actions from a first‑person view; the footage is anonymized (faces, names, screen content blurred) before being used to train household‑robot models. The company claims the generated training data outweighs the cost of providing the service. Currently limited to New York, Shift plans to expand to San Francisco, London, Zurich and Munich, with a limited‑time free‑cleaning promotion. Cleaners are vetted through partner firms but are not Shift employees and may decline tasks they find uncomfortable. Shift’s broader roadmap includes extending data collection to other domestic tasks such as plumbing, cooking, and construction, aligning with a market for human‑task recordings to improve AI and robotic systems. The initiative is part of a larger effort to accelerate autonomous home‑service robots.