Hacker News Morning Brief: 2026-05-17
This morning’s brief moves from AI adoption pressure and inference research to hardware projects, storage density, scientific oddities, and a few sharp reminders that systems fail where human behavior meets rough interfaces. The strongest discussions were not about raw novelty, but about tradeoffs: what abstractions hide, what vendors leave behind, and where tooling should meet users where they already are.
AI & Tech Policy
OpenAI and Government of Malta partner to roll out ChatGPT Plus to all citizens
Summary: OpenAI says it is partnering with Malta’s government to make ChatGPT Plus available to all Maltese citizens. The compact source pack could not fetch the full announcement because the page returned a 403, so the supported detail is limited to the national-scale access program described in the title. Even at that level, the story is notable because it frames a consumer AI product as public digital infrastructure rather than as an individual subscription.
HN Discussion: Commenters focused less on the coupon mechanics and more on institutional use. They questioned whether broad access would lead to useful training or just copy-paste demos, and several raised privacy concerns about routing citizen interactions through a private AI provider. Malta-specific trust issues also entered the discussion, including corruption and money-laundering references.
I believe there are entire companies right now under AI psychosis
Summary: The post argues that some companies have entered an irrational state around AI adoption, making sober conversation difficult. It is a short social-media claim rather than a detailed essay, but the core point is clear: AI enthusiasm can become an organizational pressure system. The concern is not that AI tools are useless, but that management fashion and fear of missing out can override judgment about where they help and where they damage foundations.
HN Discussion: A FAANG commenter described a $300-per-day token quota and steady management pressure to show AI use in meetings and daily work. Others described a conflicted stance: coding assistants are powerful, yet they can make software work feel worse and may encourage risky infrastructure changes by lightly supervised users. A concrete incentive story involved a CFO pushing AI after hearing peers brag about usage.
Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
Summary: Orthrus describes itself as fast, lossless LLM inference via dual-view diffusion decoding. The headline claim is that Orthrus-Qwen3 can produce up to 7.8 times more tokens per forward pass on Qwen3 while preserving the same output distribution. The compact GitHub excerpt does not expose benchmark tables or implementation details, so the supported summary is the repository’s high-level method and its inference-efficiency target.
HN Discussion: The first question was whether more tokens per forward pass actually means a similar reduction in compute cost, and what hidden tradeoff makes that possible. Commenters wondered why the idea had not appeared earlier and mentioned related DTree-style tricks. Practical interest centered on whether the approach could reduce hosted-model congestion or work with GGUF and quantized local models.
Security & Privacy
Grafana Labs internal source code accessed
Summary: Grafana said an unauthorized party obtained a token with access to its GitHub environment. That token reportedly enabled the attacker to download the company’s codebase. A quoted follow-up in the HN thread says the attacker then tried to blackmail Grafana by demanding payment to prevent release of the codebase, and that Grafana chose not to pay.
HN Discussion: Commenters focused on incident language as much as the breach itself, including the awkward distinction between “unauthorized party” and “threat actor” in the disclosure. Others shared non-Twitter links for people who could not read the thread. The concrete security concern was narrow but serious: a token compromise became repository access and then extortion leverage.
Business & Industry
Tesla Solar Roof is on life support as it pivot to panels
Summary: Electrek argues that Tesla’s Solar Roof has failed to scale anywhere near its 2016 promise of attractive solar tiles replacing conventional roofing. Elon Musk had targeted 1,000 new Solar Roofs per week by the end of 2019, while the article says Tesla has installed roughly 3,000 systems total nearly a decade later. Tesla is described as having stopped reporting deployment numbers and quietly pivoted back toward conventional solar panels.
HN Discussion: Commenters debated whether Solar Roof was partly a stock-market narrative during a weaker period for Tesla. Technical criticism centered on the tiny-tile approach: smaller units may look integrated, but they complicate installation, wiring, reliability, and service. Others asked whether a middle ground could exist between solar shingles and ordinary roof-mounted panels.
Fisker went bankrupt and owners built an open source car company from the ashes
Summary: Electrek reports that after Fisker Inc. filed for Chapter 11 bankruptcy in June 2024, about 11,000 Ocean SUV owners were left with vehicles whose software support was disappearing. Those cars had cost roughly $40,000 to $70,000, while over-the-air updates, connected services, and warranty support vanished with the manufacturer. The article says owners organized, reverse-engineered proprietary software, and built an open-source support effort around the Fisker Ocean.
HN Discussion: Commenters quickly clarified that the headline refers to owners of the vehicles, not owners of the failed company. A major theme was cloud dependency: this failure mode is not unique to EVs if core functions in any modern car require vendor check-ins without fallback. Some readers also criticized the article’s prose as possibly AI-generated, separating the underlying story from its presentation.
Web & Infrastructure
The Third Hard Problem
Summary: The essay extends Phil Karlton’s joke about naming and cache invalidation by proposing a third hard problem: mapping a general graph onto a hierarchy. It calls this “tree mapping” and argues that the problem appears across file systems, writing, programming, architecture, and biology. The mechanism is that humans like localized, hierarchical structures, while many real relationships are web-like and cross-linked.
HN Discussion: Commenters compared the idea with the “one true taxonomy” problem, where stakeholders cannot agree because each wants to classify on a different dimension. A stronger critique argued that hierarchies accumulate authority: file systems become navigation, organization, ownership, access-control, and governance infrastructure. Others framed trees as lossy dimensional reductions of high-dimensional graphs.
MCP Hello Page
Summary: The post describes a practical onboarding failure for MCP servers: users open an MCP endpoint in a browser, see a 401 Unauthorized JSON response, and assume the service is broken. The author’s workaround is to return a human-readable HTML page for GET /mcp when the Accept header asks for text/html and not JSON or text/event-stream. The page tells users they are looking at an MCP server and should add it to their client instead.
HN Discussion: Commenters mostly approved of using HTTP content negotiation this way, arguing that this is what the Accept header is for. Enterprise MCP users raised broader concerns about the specification, especially authentication, wrong-transport responses, and identity-provider pressure. The support-design theme was concrete: documentation works best when it appears at the exact point of confusion.
Content-defined chunking added to Bazel
Summary: BuildBuddy describes adding content-defined chunking to remote caching so large build outputs can reuse unchanged byte ranges instead of transferring whole files. The target is binaries, bundles, packages, and archives that are mostly unchanged between builds. In BuildBuddy’s own Bazel chunking PR benchmark, the implementation reportedly reduced uploaded data by 40% and disk cache size by 40%. Client-side use requires Bazel 8.7 or 9.1+ with the experimental remote cache chunking flag.
HN Discussion: Commenters immediately mapped the idea to Docker builds, where CDC could reduce image outputs or improve build-cache behavior. The main tradeoff was chunk granularity: smaller chunks may save storage but increase request count when pulling cached artifacts. A robustness concern asked whether malicious inputs could force extremely tiny or huge chunks.
Points are a weird and inconsistent unit of measure
Summary: Hillel Wayne traces a layout bug found while redoing Logic for Programmers diagrams: LaTeX and Inkscape both use points, but not the same points. LaTeX defines a point as 1/72.27 inches, about 0.3515 mm, while Inkscape uses 1/72 inch, about 0.3528 mm. The difference is only about 0.4%, but it matters when a LaTeX pseudo-grid and an Inkscape diagram grid need to align exactly.
HN Discussion: Commenters supplied CSS unit tables showing how web standards define physical units in exact relationships. The thread mixed practical unit history with jokes about imperial, metric, and fictional SI points. Other examples broadened the topic to EMUs and early dot-matrix printer spacing at exactly 1/72 inch.
System Administration
Kioxia and Dell cram 10 PB into slim 2RU server
Summary: Blocks & Files reports that Dell is using Kioxia LC9 high-capacity QLC SSDs in a 2RU PowerEdge R7725xd storage server. The described configuration puts 40 Kioxia LC9 E3.L 245.76 TB NVMe SSDs into an AMD EPYC 9005-powered system for about 9.8 PB of capacity. It supports up to five 400 Gbps NICs, and the article frames the design around dense, power-efficient storage for AI infrastructure.
HN Discussion: Commenters spotted an apparent article typo confusing terabytes and petabytes, which became a discussion about technical reporting quality. Cost estimates put the drives alone somewhere in the $500k to $1M range, making this hyperscaler, defense, or research infrastructure rather than ordinary enterprise gear. Speculative comments imagined satellite caches and calculated fill time at full NIC speed.
Tech Tools & Projects
Zerostack – A Unix-inspired coding agent written in pure Rust
Summary: Crates.io exposes Zerostack 1.0.0 as a Rust package, but the compact source excerpt does not include a project README or design document. The title positions Zerostack as a Unix-inspired coding agent implemented in pure Rust. Because the available article content is thin, the supported claim is limited to the package identity and framing, not detailed architecture or benchmark results.
HN Discussion: Commenters asked how Zerostack would handle skills, especially whether vendor-provided skills would need to be migrated into prompt text. Early users praised its speed while reporting model-provider compatibility friction, including Azure OpenAI’s max_completion_tokens parameter and custom-header support. Resource use was another theme, with one commenter contrasting its reported low RAM footprint with multi-gigabyte Claude Code sessions.
Hosting a website on an 8-bit microcontroller
Summary: The post describes serving a real website from an AVR64DD32, an 8-bit AVR microcontroller with a 24 MHz core, 8 KB of SRAM, 64 KB of flash, and a roughly $1 part cost. The author explains why raw Ethernet is difficult: 10BASE-T uses Manchester encoding, so a 10 Mbit/s link becomes a 20 Mbit/s wire signal. That is beyond what the AVR’s IO timing can realistically generate, making the project a constraint exercise rather than a sensible hosting architecture.
HN Discussion: Commenters compared AVR DD, EA, and EB parts with newer Microchip PIC32 CM ARM M0+ chips, debating whether modern 32-bit microcontrollers make AVR DD less attractive. Others enjoyed the visible slowness of the demo, describing the page as streaming in like dial-up-era rendering. The tone treated the project as charming engineering theater with real hardware lessons.
A nicer voltmeter clock
Summary: The author revisits a 2019 desk clock design that displays time using analog panel voltmeters instead of ordinary clock faces. The new build uses three generic 90-degree 5 V panel meters, disassembles them, measures their faces, and replaces the stock decals with custom time scales. The hour gauge has 13 divisions and the minute and second gauges have 61 divisions so the hands can move continuously instead of jumping at exact boundaries.
HN Discussion: Commenters admired the physical craft and treated the project as a source of ideas for their own instrument-panel builds. Several focused on motion behavior, especially the second-hand reset and the 11:59:59 rollover. One related build compared real panel meters with simulated LCD panel meters for an analog computer, weighing flexibility, cost, and authenticity.
C++26 Shipped a SIMD Library Nobody Asked For
Summary: The article criticizes C++26’s std::simd, the P1928 portable SIMD abstraction meant to target AVX2, AVX-512, NEON, and SVE without architecture-specific intrinsics. Its central claim is that std::simd compiles much slower, can run slower than scalar loops, chooses poor vector widths by default, and cannot express many operations used in real SIMD code. The author cites a satirical repository and reproduced benchmarks as evidence, while tracing the library lineage back to Matthias Kretz’s work around 2009.
HN Discussion: Experienced SIMD programmers argued that abstractions and auto-vectorizers often fail outside narrow cases because microarchitectures differ too much for one interface to optimize well. A standards-history commenter said earlier SIMD proposals faced similar objections around SVE mapping and control-flow expression. The thread also debated article quality, with one reader dismissing it as AI-generated despite conceding there might be a worthwhile story.
Show HN: Rocksky – Music scrobbling and discovery on the AT Protocol
Summary: Rocksky is a decentralized music tracking and discovery platform built on the AT Protocol. The repository metadata lists Spotify, ATProto, Last.fm, MusicBrainz, scrobbling, and ListenBrainz among its integrations and concepts. Its codebase is mostly TypeScript with a notable Rust component, and visible branches suggest active work on likes, feed generation, ListenBrainz/Navidrome support, MusicBrainz, Spotify, artist listeners, avatars, and Discord embeds.
HN Discussion: Commenters compared Rocksky to Last.fm, with one saying Last.fm remains the best social music platform and recommendation model. AT Protocol interest was explicit, including enthusiasm for seeing an “Atmosphere” app on HN. Integration questions focused on authentication, especially whether native iOS apps need Rocksky API keys or can support a Bluesky-style OAuth flow.
3D Gaussian Splatting in a Weekend
Summary: Benjamin Feldman’s tutorial builds a simple 3D Gaussian Splatting renderer and explains the technique from first principles. It frames 3DGS as reconstructing a 3D scene from pictures by rendering the scene from known camera angles, comparing with ground-truth images, and updating the scene to reduce differences. Unlike triangle renderers, 3DGS uses Gaussian splats as primitives, so the post walks through spherical harmonics, covariance projection, shaders, bounding quads, sorting, and differentiability.
HN Discussion: One technical thread focused on spherical harmonics, including whether high-order coefficients can be dropped or sparsely encoded to trade detail for speed and size. Another commenter asked how 3DGS differs from point clouds produced by LiDAR or Kinect-style depth sensors. The concrete themes were representation and performance: splats versus points, color-frequency encoding, and what rendering detail can be compressed away.
Ploopy Bean: a trackpoint for every computer
Summary: Ploopy’s Bean Pointing Stick is a 3D-printed, open-source pointing-stick mouse offered as a preorder product. The page says it adds high-precision pointing-stick functionality to any setup and includes four Omron D2LS-21 buttons for clicks. It runs QMK and supports VIA, which should allow portable customization, and the unit is sold fully assembled with preorder pricing shown at CAD 69.99 before cable options.
HN Discussion: TrackPoint enthusiasts focused on ergonomics, especially whether projects should use 6 mm assemblies and soft-rim caps instead of smaller caps that require more pressure. Several questioned the form factor because a pointing stick’s value usually comes from being inside the keyboard, avoiding hand movement from the home row. Alternatives included Ultimate Hacking Keyboard pointer modules and Kanata mouse emulation.
Academic & Research
Unknowable Math Can Help Hide Secrets
Summary: Quanta reports on a graduate student’s use of the complexity of mathematical proofs as a cryptographic tool. The article sits at the intersection of computational complexity, proof theory, foundations of mathematics, and cryptography. Its core framing is that unknowability itself, not just hard-but-knowable structure, can help hide secrets, with non-interactive zero-knowledge proofs appearing as a highlighted goal in the surrounding discussion.
HN Discussion: Commenters tried to map the idea onto familiar primitives, asking whether it behaves like a one-way hash over mathematical statements. Skepticism centered on what practical cryptographic tool the article actually delivers and how it differs from existing non-interactive systems such as ZK-SNARKs. Another concrete concern was whether the proposal is just security through obscurity under a mathematical label.
Self-Distillation Enables Continual Learning [pdf]
Summary: The arXiv paper proposes Self-Distillation Fine-Tuning, or SDFT, for continual learning in foundation models. Its motivation is that models should acquire new skills without degrading old capabilities, while on-policy reinforcement learning often requires explicit reward functions that are unavailable. The abstract positions SDFT as a way to obtain on-policy-style benefits from expert demonstrations, where standard supervised fine-tuning is off-policy.
HN Discussion: Commenters dug into empirical setup details, including Qwen-2.5-7B-Instruct and the ToolAlpaca benchmark for selecting correct tool calls from API specs and user requests. Some skepticism targeted the paper’s wording, especially “enables” and “establishing” in the title and abstract. A naming thread compared SDFT with Apple’s Simple Self-Distillation and complained that SSD is an overloaded acronym.
Illusions of Understanding in the Sciences
Summary: The Springer page is for an article titled “Illusions of Understanding in the Sciences” in Computational Brain & Behavior. The compact pack does not expose the abstract, so specific claims about the paper’s evidence are not supported here. The title and venue suggest a cognitive-science or philosophy-of-science focus on when scientific models create confidence that exceeds actual understanding.
HN Discussion: Discussion was very thin, but the visible comment asked what a model is and noted that models can sit at different abstraction distances from experienced reality. The concrete theme is model interpretation: whether simplified scientific representations genuinely explain phenomena or mainly create a feeling of explanation. With so little thread material, the safest read is that the submission prompted conceptual rather than empirical debate.
A molecule with half-Möbius topology
Summary: The Science link returned a 403 in the compact pack, but the title identifies the reported result as a molecule with half-Möbius topology. A quoted explanatory comment describes half-Möbius electronic topology as a cyclic molecule whose pi-orbital basis undergoes a 90-degree phase twist per revolution. That makes it distinct from a trivial Hückel case with no net twist and from a classical Möbius case.
HN Discussion: Commenters translated the topology into more familiar terms: 90-degree phase twists, multiple circuits around the ring, and comparisons with Möbius or Klein-bottle-like structures. A methodology theme appeared around scanning tunneling microscope observations of Dyson orbitals, presented as evidence that researchers were not just imagining the topology. The tone mixed technical curiosity with aesthetic awe.
History & Science
Colossus: The Forbin Project
Summary: The linked Wikipedia page covers Colossus: The Forbin Project, a 1970 film directed by Joseph Sargent with a screenplay by James Bridges. The page is organized around plot, cast, production, release, reception, remake notes, and references. The compact excerpt supports treating the story as a historical science-fiction reference point rather than as a new technical development.
HN Discussion: Commenters connected the film’s dark ending and computer-control premise to current AI guardrail debates. Several responses were personal recollections, including being warned about the film during the GPT boom or seeing machines named Colossus and Guardian in old CS labs. One commenter noted that the film came from the first book in a trilogy and wished the full sequence had been adapted.
Why did Clovis toolmakers choose difficult quartz crystal?
Summary: Phys.org reports on a study asking why Clovis people in North America sometimes made points and tools from quartz crystal. Quartz crystal is described as a difficult, low-quality knapping material because of its size constraints, hardness, and crystalline structure. The puzzle is that toolmakers selected it despite practical drawbacks, which suggests the choice may have involved aesthetic, symbolic, social, or other non-utilitarian motives.
HN Discussion: The thread was small but pointed. One commenter warned against defaulting to “it was religious” explanations in archaeology, while another offered the simpler explanation that the material looked cool. The concrete theme was interpretation risk: practical difficulty alone does not prove a symbolic or ritual reason.
Halt and Catch Fire
Summary: The post traces “Halt and Catch Fire” from computer-engineering humor to its later association with the AMC show about the 1980s and 1990s computer industry. In computing, HCF refers to machine code or invalid opcodes that make a CPU stop doing useful work and require reset or power cycling. The article cites an IBM System/360 story in which an invalid opcode repeatedly accessed a magnetic-core memory location, heating it enough to create a literal fire risk.
HN Discussion: Commenters used the post to reminisce about a period when computing hardware and software felt more comprehensible and directly controllable. Hardware anecdotes were specific, including a Commodore PET CRT controller that could park the beam and burn phosphor if misprogrammed. Another thread examined self-destruct-chip stories and whether military folklore grew out of accidental destructive opcodes.
PART Telescopes – Bringing radio astronomy within reach of rural schools
Summary: The PART Initiative is a student project to design low-cost radio telescopes and open software for rural schools. The team aims for a simple, reliable telescope with total production and assembly cost under $500. The design targets 21 cm line observations associated with galactic hydrogen, using a weather satellite dish, conductive plastic base, low-noise amplifiers, bandpass filters, a software-defined radio workflow, and a motor system.
HN Discussion: Commenters shared adjacent low-cost radio astronomy projects, including AstroChart, SETI ARISE, WVU DSPIRA lessons, MIT Haystack SRT material, Pictor, and Virgo. One criticism was that the current site appears partly boilerplate, referring to documentation that a commenter could not find. Technical curiosity focused on architecture and cost reduction, including projects using low-cost GNSS receiver chips as unusual RF frontends.
Fame! A Misunderstanding: A new translation of Albert Camus’s complete notebooks
Summary: The Los Angeles Review of Books essay reviews The Complete Notebooks by Albert Camus, translated by Ryan Bloom and published by University of Chicago Press in 2026. The review argues that Camus has long been misunderstood and that the complete notebooks offer a corrective to his public reputation. It places the notebooks among posthumously translated works including A Happy Death, The First Man, essays, journalism, lectures, correspondence, and earlier notebook collections.
HN Discussion: Commenters questioned whether misunderstanding Camus is mainly an English-language reception problem or whether French readers have similar disputes. One reader found The Myth of Sisyphus tedious and compared its embrace of the absurd with Stoic and Taoist themes. A quoted Camus passage about literary fame prompted discussion of reputations shaped by hurried journalism and a single talked-about book.
The bird eye was pushed to an evolutionary extreme
Summary: Quanta reports on new research explaining how bird retinas operate despite being among the most energetically expensive tissues in the animal kingdom. The article’s puzzle is that the bird retina does not use the energy advantage of oxygen in the expected way. The excerpt highlights a macaw eye with no visible blood vessels, raising the question of how such eyes work so well without oxygen delivery through retinal vessels.
HN Discussion: One thread connected anaerobic glycolysis in bird-eye metabolism to cancer biology, asking whether cancer cells revert to older, less alien metabolic behaviors. Commenters compared vertebrate and squid eye layouts, including blind spots and whether nerves sit above or below light-sensitive cells. There was also discussion of adaptation versus evolutionary contingency.
Other
We’ve made the world too complicated
Summary: The essay argues that modern life has become an abstract, compressed system of technologies, buildings, laws, infrastructure, and institutions that individuals cannot fully understand or control. The author links this complexity to background stress: clenched jaws, shallow breathing, higher blood pressure, and a persistent sense that the world does not make sense. It contrasts technological-savior narratives, including AGI, with the possibility that technology itself deepens abstraction.
HN Discussion: Commenters connected the essay to older critiques of civilization, especially the idea that humans adapted the environment to themselves and then made it too complicated. A recurring theme was alienation from abstract work: remote software work can be flexible yet feel less grounded than fixing a bike, baking bread, or solving a local problem. Others pushed toward existential meaning and humanity’s unusual ability to study the universe.
Bear spray is exploding in the trash near Yellowstone National Park
Summary: SFGATE reports that discarded bear-spray cans are exploding in trash near Yellowstone National Park. The practical message is direct: visitors should stop throwing away pressurized spray canisters after their trips. Workers at a garbage transfer station near the park encounter the problem as hikers and tourists dispose of safety gear used around grizzly and black bear habitat.
HN Discussion: Commenters suggested disposal alternatives, including mailing unused cans home by USPS ground and setting up donation bins at airports or park exit points. A waste-reduction theme emerged around tourists buying a can for a short visit and then discarding it instead of renting or handing it to another visitor. The thread also discussed capsaicin under pressure and heat, with comparisons to heating peppers in a microwave.