The newest HackerNews stories, desensationalized.

The IOCCC29 announcement details where to access the 2025 winning entries, including indexed pages, source code, author remarks, and a downloadable archive. Judges described submission volume and quality as unusually high despite the contest’s second consecutive run after a long hiatus, attributing results to cleaner workflows, stronger web presence, and stronger derivative work from prior winners. The post-contest process was documented more tightly, with YouTube segments for each winner, clearer updates to entry pages, and better submission and judging procedures. New fun challenges were added to entry pages with a GitHub workflow for contributors to improve solutions or submit alternatives when challenges remain open. The contest rules and guidelines were substantially rewritten for 2025 (rules 29.15 and guidelines 29.08), and judges said they may continue refining IOCCC30 operations through 2026 while preserving the same approach to contest documentation. The entry highlights include several notable winners across emulator, obfuscation, gaming, and data-generation themes, with multiple entrants from prior winners and a first-time geographic expansion. Non-winning participants are encouraged to iterate and resubmit, and the organizers acknowledge that several 2025 winners were improved forms of earlier non-winning work.

The proposal asks Anthropic for a published Linux desktop client for Claude, arguing that official support would complete a real gap rather than a cosmetic request. It notes that Claude Desktop ships only for macOS and Windows, while Claude Code already has a signed Linux pipeline and supports apt, dnf, and apk distribution, but lacks extension testing and Cowork parity in a supported graphical app. The report cites reverse-engineered Cowork behavior on macOS, which runs a Linux VM and then executes Claude Code inside containerized layers, to argue the technical path already exists and that only a Linux target appears missing. It also details user risk from the absence of first-party Linux builds: people rely on community repacks such as aaddrick/claude-desktop-debian, which are substantial but not Anthropic-signed or vendor-audited, yet widely used because they are otherwise the only practical option for desktop-like features. Proponents support the request with platform data showing Linux is materially used by developers and in markets like India and the US, making official support strategically relevant beyond a niche case. The author proposes official Ubuntu LTS plus Debian delivery through signed repositories and includes lower-cost alternatives, from a clear public roadmap statement to documented guidance for secure community usage. A steelman counterpoint is included, citing cross-distro fragmentation, support overhead, enterprise cost-benefit, and release-security complexity as reasons the effort might be deferred.

Yon describes a language model where values are immutable content-addressed entities stored in a fixed-geometry Leech-lattice heap with 196,560 slots per heap, so identical bytes map to the same address and equality checks become constant-time comparisons. The design frames worlds as categories, places as objects, and values as sections, with behavior expressed by arrows and identity handled explicitly rather than implicitly. It also removes traditional primitives such as garbage collection, threads, exceptions, and typeclasses, treating failure as data and using processes for concurrency via shared-memory interaction. The language emphasizes category-theoretic structure, internal logic through the subobject classifier, and the idea that identity is only requested when needed. Commenters note the concept is ambitious but criticize the documentation for being dense, metaphor-heavy, and lacking a clear problem statement that shows why this approach is better than existing systems. They ask for concrete motivation, prior-art comparisons, and practical memory-life-cycle discussion, especially for workloads with many unique values such as servers. The result is a technically inventive proposal whose practical feasibility is still difficult to assess from the presented explanation alone.

The article explains that classic 16-bit Windows memory management had to solve for fragmented and limited hardware long before modern virtual memory, effectively acting as an overlay-style manager. It describes how applications used segment handles instead of fixed addresses, with GlobalAlloc, GlobalLock, and GlobalUnlock used to pin and release movable memory while ensuring exported entry points and window procedures were patched safely. The NE executable format enabled per-segment loading and unloading, and segments were classified by mobility and discardability so Windows could compact RAM, reload code from disk, and preserve writable data semantics. The writeup details DLL behavior in this model, including two-level symbol resolution, stack/register caveats such as SS != DS, and why Windows-specific compiler switches (/Aw and /Gw) and prolog/epilog sequences were required for relocatable exported routines. It also clarifies how stack frame shaping with BP adjustment let the kernel detect active references when moving segments. In contrast, OS/2 on the 286 used selector-based addressing to reduce programmer burden, avoiding much of the lock/unlock and patching complexity. The article notes that Windows had stress tools like Shaker and HeapWalk, later Stress, to expose low-memory and fragmentation bugs that often stayed hidden because invalid accesses did not always fault on 8086 hardware.

The article argues that growing LLM context windows are increasingly bottlenecked by KV cache storage and transfer, and revisits lossless compression as a better path than lossy quantization that trades accuracy for uncertain quality. It explains an arithmetic-coding scheme where a cheaper predictor model runs on the same prompt and the target model’s true cache values are encoded against a per-scalar residual distribution, using a Gaussian-plus-mixture model to handle heavy-tailed errors. Because a KV cache is deterministic for a fixed prompt and model, any bitrate above zero is due to prediction mismatch, so better predictor fit directly reduces bits. The authors propose a practical predictor built from a structurally similar quantized model (FP8/INT4/MXFP4 variants) whose forward-pass noise is small and reusable, with residual variance estimated from a small calibration set. Experiments on Qwen3 targets show held-out C4 compression of 2.37× to 2.70× from BF16 caches, and 3.08× to 3.90× when starting from FP8 caches; with BF16-to-FP8 preprocessing, total savings reach about 6× to 8×, with larger models compressing better. They note arithmetic-coding throughput and bit-identical predictor outputs are required for deployment, and discuss use cases where bandwidth or memory dominates, such as cross-datacenter disaggregated prefill and prefix/offload caches, while cautioning that the predictor must justify its compute cost.

Kyushu is presented as an open source CLI that lets users compile JavaScript or TypeScript handlers into a single WebAssembly binary that runs through a Cloudflare Workers-like fetch API on a VPS or similar host. The project targets self-hosting with minimal dependencies, explicitly avoiding Node.js, Bun, and Docker while emphasizing isolation through Wasm sandboxing. Commenters read it as a lightweight, practical alternative for Cloudflare-style execution with easier local testing, since packaging into one binary lowers setup overhead. The discussion links this approach to related sandbox efforts, suggesting it as part of a broader shift toward simpler runtime-isolation layers. One perspective in the thread argues that although web technologies can be cumbersome, modern tooling has become tightly coupled with HTML, CSS, JavaScript, Wasm, and full-stack development. Another view from a long-term Wasm builder frames the value of such systems as adding choice among virtualization options alongside containers and microVMs, rather than replacing existing models. Another participant asks for practical use-case explanations, while others offer lightweight validation of the design and interface. Overall, the content and comments suggest Kyushu is aimed at developers who want portable, secure function execution without platform lock-in.

The post explains how to quantify late component rendering in Qt Quick by measuring the delay from making a QQuickItem visible to the moment it is first displayed. The author describes a TimedItem subclass that listens to QQuickItem::visibleChanged, starts an QElapsedTimer when visibility turns true, and connects to QQuickWindow::afterFrameEnd to capture the frame timing. To avoid recording an unrelated animated frame, the item calls ensurePolished and overrides updatePolish to set a readiness flag, so measurement happens only on the intended paint cycle. The measured milliseconds are normalized by frame duration (1000 / QScreen::refreshRate) and floored to frame-count latency, which is more stable across refresh-rate differences than raw time. The implementation disconnects from afterFrameEnd after one successful sample and emits a property change with the result. The author also recommends setting initial visibility to false in C++ so that QML initialization order does not suppress visibleChanged when an item may start true in some construction paths. The code is presented as a practical template for profiling and debugging perceived frame drops, with references to Qt scene graph and rendering docs for deeper context. The post includes a brief call to action for profiling help, training, and Qt migration services.

The author argues that software should be judged primarily by end-user utility, then supported by correctness, maintainability, and efficiency rather than ideology or tooling prestige. In that model, a project can fail even if technically safe or elegant when it is user-hostile, buggy, slow, or hard to maintain. The author downplays language guarantees and abstraction-heavy design unless they directly improve reliability and long-term utility. The piece insists on practical process and continuous improvement instead of static claims of correctness, with developer comfort considered only insofar as it helps deliver better software. Comment discussion reframes the central tradeoff: if every missing feature is effectively a bug, teams must prioritize highest-value work, not exhaustive defect elimination. Readers connect this argument to broader industry shifts, including AI coding and language ecosystem changes, while others stress that usability, clarity, and stability often matter more than technical purity in practice.

9 MOTHERS presents itself as a counter-drone engineering team in Austin that builds integrated systems spanning AI perception, kinetic engagement, and the full operational loop from testing to deployment. The posting argues that traditional defense timelines are too slow and claims its model is to design and build manufacturable hardware and software together, with owners responsible for entire systems rather than narrow tasks. It emphasizes short iteration cycles, with teams instrumenting range data and live-fire outcomes to drive daily design changes instead of planning against static roadmaps. The company says operators are central to requirements, prioritizing mass, setup time, and human factors from the start, while also setting constraints around cost per shot, serviceability, and affordability for U.S. and allied use. Hiring is open across audio machine learning, DSP, systems, robotics, hardware, and platform roles, with the team inviting otherwise unmatched candidates to reach out directly. The job listing portrays the role as physically based in Austin, near-range operations, with a hands-on culture of bench work, firmware updates, soldering, and immediate spec updates.

The article traces how polo, led by Adolfo Cambiaso, moved from a single hunch to a commercial cloning model after preserving cells from a champion horse before euthanasia. Equine cloning uses somatic cell nuclear transfer, but cloning still depends on difficult reprogramming and often suffers mitochondrial and epigenetic mismatch effects that can reduce viability and increase health problems. Technical practice has improved through embryo activation methods, culture conditions, and use of adult stem cells, yet efficiency remains low: only a fraction of transferred embryos become healthy foals, and stillbirths, placental issues, and developmental abnormalities are recurring risks. In horses, oocyte source and surrogate management matter; transvaginal aspiration yields markedly better blastocyst and healthy outcome rates than eggs from slaughtered mares. Argentina became the global center through companies such as Kheiron Biotech, which scaled production, reduced abnormalities, and by 2024 reported about 1,000 clones and a recent CRISPR proof-of-concept editing of the myostatin pathway in five foals. The practice now powers a high-cost industry, with births near $40,000, and large tournaments have showcased all-clone lineups as a competitive strategy. The ethical debate shifted from pure clone health concerns toward broader issues: objectification, welfare costs, hidden complications, and reduced genetic diversity within breeds dominated by narrow cloning lines. Experts differ on whether the field is now routine reproductive biotechnolgy or an unregulated performance-enhancement frontier that also requires strong international reporting and oversight, especially with gene editing. The article links these arguments to a future where performance engineering increasingly bl

Players describe a recurring gameplay regression in games using Steam networking where PC-to-PC matches in Israel show around 120 ms latency in Street Fighter 6, while the same players report 60–80 ms with Europeans and 5–10 ms on PC-to-PS5, with cross-play staying much smoother. They say multiple community members across different ISPs tried port-forwarding and ISP checks without finding local infrastructure issues, and they note similar issues do not appear in some non-Steam P2P games such as Tekken 8. The thread reports that the problem started after a March-era update and that several games can be made playable by manually replacing Steam WebRTC DLLs, indicating a likely client-side or transport-layer fallback behavior rather than pure geography. Commenters connect the symptoms to STUN failures that can push sessions onto relay paths with higher latency, which aligns with anecdotal patterns outside the original region. The original post framed the impact as concentrated in Israel and nearby countries, but comments suggest the issue may be more widespread or differently manifested by game, update, and connection path. No official investigation or clear explanation is provided in the submission, and participants repeatedly ask for a direct engineering response to identify root cause and fix strategy.

The post showcases a curated collection of 11,082 out-of-copyright works that is free to browse, download, and reuse, with new additions added regularly. It highlights a rights-model that distinguishes between underlying work rights and digital copy rights, while repeatedly cautioning that the information is guidance only and should be validated with the source institution. The site frames itself as a practical resource for creators and reusers by documenting “no known restrictions” where applicable. The comments reflect strong interest in its utility for research and creative projects, especially in game and design work. Commenters also acknowledge the value of explicit legal caveats and provenance documentation versus generic public-domain claims. Overall, the release is positioned as a useful reuse-focused archive, but it is not positioned as definitive legal clearance on its own.

The release describes ntsc-rs as a free, open-source Rust implementation that simulates NTSC transmission and VHS encoding behavior through signal-modeling algorithms instead of superficial LUT-and-overlay tricks, and it is optimized with multithreading and SIMD for real-time use at high resolutions. It is shipped as a standalone tool, a web application, and plugins for After Effects, Premiere, and OpenFX hosts such as DaVinci Resolve, Hitfilm, and Vegas. This positions it as a production-grade alternative to look-only presets. In discussion, commenters frame it as part of a broader analog-signature revival, comparing it to prior emulation work and highlighting that fidelity can depend on details beyond basic blur and color shifts. The thread also raises technical gaps in many existing solutions, including phase and burst handling, oscillator drift, PAL/SECAM behaviors, and deinterlacing edge cases. Community members suggest training workflows and dataset-driven AI pipelines that apply the effect to clean sources and learn reverse mappings or style transformations, showing demand for both deterministic and generative uses. Several participants also distinguish this tool from historical production reality by noting that broadcast-era composite quality and CRT-display behavior were often better than consumer VHS, while still valuing deliberate visual degradation when used for period-specific storytelling. Overall, the post is treated as a strong, standards-aware implementation whose adoption is driven by both creative control and interest in more physically accurate media emulation.

The episode highlights lcamtuf’s move to Substack and his current focus on teaching electronics, including promotion of his newly published book, The Secret Life of Circuits, with a June discount offer. It revisits his long-running influence on security culture, from early fuzzing-era work to modern readers on Hacker News, and contrasts how security books can become outdated faster than electronics foundations. He explains that many engineering ideas, including Rowhammer, have practical security and hardware implications beyond theory. The conversation surveys the evolution of computing from calculators and CRT-era displays through Nixie tubes to logic gates and processors, while also discussing the craft of transistor work from FETs to BJTs and why BJT sequencing after diodes can be conceptually awkward. He repeatedly emphasizes intuition-friendly derivations from basic trigonometry, challenging the expectation that electronics requires only calculus-heavy treatment. Practical observations include the niche status and value of old mechanical calculators, the educational appeal of dead-end designs, and the distinction between durable hardware concepts and rapidly aging security topics.

Symbolica 2.0 introduces a programmable-symbol model that lets users customize normalization, printing, differentiation, series expansion, and numeric evaluation through symbol-level hooks, and the release also simplifies both the Rust and Python-facing APIs. The Rust update emphasizes ergonomics through a richer prelude, builder-based configuration, clearer type handling, and easier evaluator construction, while Python users gain expanded notebook formatting, HTML/Typst outputs, and broader symbolic function support. Output readability has been improved with line wrapping, colored rendering, and structural formatting, and the evaluator pipeline now supports broader numeric domains through a redesigned interface, JIT compilation, and double-float arithmetic with roughly 106 bits of precision. The evaluation architecture now stores callbacks in a type-erased registry keyed by domain and materializes concrete functions during specialization, while domain fallback logic enables vectorization and compatibility between scalar and advanced types. Under the hood, normalization, Horner strategy, pattern matching, polynomial GCD, sorting, and pair elimination were optimized, with reported speedups from 2x to 10,000x on real workloads and requests for additional benchmark feedback. New mathematical hooks add practical support for gamma variants, polylogarithms, Bessel functions, zeta, geometric functions, and their local series behavior, including numeric evaluation inside generated kernels. The project lead notes that AI assistance helped with large refactors and documentation but not core algorithmic debugging, then points to an optional future GMP replacement route for easier WASM and Windows builds at a performance cost. Overall, the release is positioned as an attempt

The paper studies token consumption in LLM-based multi-agent software engineering by tracing 30 development tasks in ChatDev using a GPT-5 reasoning model across mapped SDLC stages: Design, Coding, Code Completion, Code Review, Testing, and Documentation. It breaks token use into input, output, and reasoning tokens to quantify where costs accumulate. On average, Code Review consumes 59.4% of tokens, making iterative refinement and verification the dominant cost center rather than initial code generation. Input tokens form the largest share at 53.9%, indicating repeated context replay and proposal expansion as the main driver of spend. The work argues this uneven distribution can explain unstable deployment costs and environmental concerns, and proposes the framework as a practical tool for forecasting expenses and redesigning agent collaboration protocols toward token efficiency.

Warren's Abstract Machine: A Tutorial Reconstruction is a book by Hassan Ait-Kaci that documents the Warren Abstract Machine and related execution mechanisms for Prolog, and it is now out of print. The source notes that the author retained copyright while a non-commercial electronic version was previously distributed through an external site, and that site is currently parked. In response, the poster has set up a new page to provide access to the electronic version for readers who need it. The update is framed as preservation and access restoration for a niche systems-level programming resource rather than a release of new content.

The project introduces Lathe, a Go-based CLI and LLM skill set designed to generate and manage hands-on technical tutorials from prompts, then let learners implement the steps themselves in a local web UI. The workflow is intentionally pedagogical: users can generate single or multi-part tutorials, ask targeted questions, verify builds, and extend parts, while the UI supplies the exact skill commands to keep stateful interactions inside Claude Code, Cursor, or Codex sessions. Installation is available as a single macOS Homebrew cask, a Linux/install script path, or Go source build, and tutorial assets are stored in a local directory tree under ~/.lathe/tutorials where metadata tracks status, tags, tools, sources, model, and voice. The author positions Lathe as a response to losing hands-on learning with modern coding assistants, arguing that building from first principles yields deeper understanding than passive code generation. Tutorials support explicit voice profiles, including plainspoken and companion, with optional custom voices constrained to avoid impersonation or fabricated credentials, and each tutorial records provenance and a byline naming the model and voice used. Verification is opt-in and runs in an interactive LLM session against temporary scratch directories, with statuses such as unverified, verifying, verified, failed, skipped, or extending, and recorded error output for failed checkpoints. The system also provides local search, filtering, version grouping, and source-link panels to help users audit tutorial quality and narrow their learning library. The author emphasizes that human tutorials remain preferable, but presents Lathe as useful for unfamiliar, sparse-documentation domains and notes limitations, including unverifiable behavior on non-MacOS/

The post argues that a small team built and shipped an internal beta over five months with no manually written code by treating Codex as the primary executor and humans as system architects. It describes shifting work from coding to building an agent-friendly repository: thin prompts, strict boundaries, rich docs in-repo, review loops, and automated observability, testing, and feedback workflows. The team reports scaling from three to seven engineers while sustaining growth in throughput and claims roughly 1,500 merged PRs and about 1.4 million lines of generated code across product, tooling, and docs. The process includes agent review chains, browser-driven validation, and execution plans that keep context local, while legacy-style constraints are replaced by typed interfaces, architectural layering, and mechanical linting. The author emphasizes that quality is preserved through predictable structure and continuous cleanup through codified "golden principles." They argue that human effort now centers on feedback loops, environment design, and judgment rather than direct implementation. They also note limits: they do not claim broad generality, and they continue to see unresolved questions around long-term coherence, drift, and the model quality bar. The post closes by warning that agent-first engineering makes scaffolding and governance more critical than ever.

Biohub announced three open releases—ESMC, ESMFold2, and ESM Atlas—framed as a world model for protein biology with broader discovery and design use cases than prior tools. ESMC is trained on roughly 2.8 billion protein sequences across life and is intended to capture conserved rules of folding, interaction, and function. ESMFold2 converts those learned sequence representations into atomically resolved structures and was reported in a preprint to produce candidate binders for five targets (EGFR, PDGFRβ, PD-L1, CTLA-4, CD45) in days instead of years, with hit rates reported between 36–88% for compact binders and 15–29% for antibody-derived designs plus experimental validation of binding, specificity, and stability. For PD-L1, designs reportedly restored T-cell signaling in laboratory tests, similar in direction to approved checkpoint therapy targets. ESM Atlas expands model navigation to 6.8 billion sequences and 1.1 billion predicted structures, surfacing links the release says are not yet captured in standard annotation systems. The initiative is positioned as open infrastructure for researchers, with explicit MIT-like open licensing details emphasized in discussion, and Biohub presenting this as a nonprofit mission to accelerate cures and preventive research. The release is presented as a practical attempt to shift early binder discovery from primarily empirical screening toward computation-guided design while keeping artifacts publicly accessible.

The article says liminal photography has become a crowd-curated expression of late-capitalist alienation, centered on empty malls, hallways, and service interiors like the shuttered Century III Mall and linked to the 2019 Backrooms mythos from 4chan. It argues that the effect comes from human absence, retro-commercial textures, and a limbo-like familiarity that feels both mundane and uncanny. The essay credits Facebook groups, Reddit, and TikTok for scaling the trend, citing large followings and community rules that reject AI-generated images to preserve a real-world documentary edge. It frames the movement as democratic cultural commentary outside elite institutions and traces visual precedents from Surrealism through Hopper, Wyeth, and Grant Wood. The piece also treats COVID-19 lockdowns and platform-driven isolation as accelerants, with spaces appearing as social no-places or simulations. Commenters broadly question the “defining aesthetic” claim, calling it more comparable to other niche revivals than a universal marker of the age. They say the strongest pull may be nostalgia for expected-but-vanishing public spaces, the uncanny emptiness of places once busy, and the 1990s memory field. Several expand origins to DOOM mod culture, House of Leaves, and earlier creepypasta. Others connect it to AI latent-space blending and mass cultural recombination, while some argue the late-capitalism framing is incomplete or overly theoretical. The discussion agrees the imagery is compelling but debates whether it is an epochal diagnosis or a fashionable online style.

The writer argues that the strongest relationships can span all parts of life, not just romance, and says they now satisfy social needs they once met through friends, family, and a blog. He and Alex use a weekly "Friday Demo" ritual modeled on startup all-hands presentations, where each shares something worked on that week, however rough, from a playlist to a landing page. This practice is framed as playful, low-pressure, and creatively productive, helping both stay motivated and accountable while navigating uncertainty like job hunting. Their shared examples include joint routines, cross-genre interests, family conversations, and personal projects such as book reviews and side websites, with practical constraints shaping habits. He contrasts initial fears of codependency with his current view that deep mutual respect and patience can support healthy identity growth for both people. The essay also links relationship quality to broader social categories, saying similar all-encompassing dynamics can enrich friendships, pets, and mentoring relationships. It invokes intellectual influences like Ayn Rand and startup culture as shared touchstones while emphasizing present-moment attention, trust, and conversation over performative productivity. The writer closes by describing a shift from performing for audiences to teaching and learning within the relationship, arguing this focus reduces distraction and speeds meaningful progress.

Five leading researchers were removed from the ADA annual meeting in New Orleans after handing out reprints of a Diabetes Care editorial criticizing U.S. administration actions toward scientific research. The group included editor-in-chief Steven Kahn and four other senior scientists, and they reported being physically escorted out and denied further attendance after staff removed their conference lanyards. ADA said the removals were for code-of-conduct violations, pointing to provisions against disruptive behavior, including protesting, harassment, and disorderly conduct, and said attendees were asked to stop but did not comply. The action occurred outside a session scheduled for NIH director Jay Bhattacharya, who later did not attend while another NIH official spoke. Critics in the comments treated the incident as evidence of censorship and political control, while others framed it as a policy-application dispute. The discussion links the episode to broader debates over scientific independence, the role of professional societies, and where advocacy ends and disruption begins in research settings.

A single-image generation method avoids conventional training by representing the input image as a patch dataset at multiple scales and computing score estimates with a closed-form denoiser, which keeps the patch dimensionality manageable. This replaces hours of optimization with a tractable, training-free diffusion step while preserving internal structural statistics of the source image. The approach is explicitly linked to classical patch-based restoration methods, showing a bridge between recent generative modeling and older denoising paradigms. Reported results claim state-of-the-art quality and diversity versus trained single-image diffusion baselines on the same problem setting. The method is demonstrated on several tasks, including unconditional generation, text-guided stylization, image symmetrization, and retargeting, and is said to work with latent-space diffusion. Additional acceleration techniques are presented to reach megapixel generation in about one second and gigapixel generation in minutes, signaling practical speed gains for high-resolution synthesis.

Chen proposed kernel spawn templates to reduce repeated fork()+exec() overhead for workloads that launch the same binary many times, such as frequent Git invocations. A template is created once with spawn_template_create() for an executable, then reused via spawn_template_spawn() with per-launch argv, environment, and action vectors for file descriptor, directory, and signal setup. The approach preserves existing exec validation checks while caching setup data and aims to avoid the full repeated cost of process creation in the expected pattern. Reported benchmarks showed about a 2% speed improvement, which is modest but relevant at high frequency. The patch generated strong discussion: Mateusz Guzik offered a detailed review, and Christian Brauner supported the direction only as a path toward native posix_spawn semantics. Brauner argued the ideal interface should avoid exposing fork() beneath the hood, while Chen said future work may move in that direction and the template mechanism itself may not land as originally proposed. The piece frames the proposal as incremental and practical, with recognition that it does not by itself retire fork()+exec().

Chemists reported experimental signals for B80, an 80-atom boron cage analogous to a carbon buckyball, generated through laser vaporization with argon-assisted cooling and validated against simulated spectra to identify three matching peaks. The study is significant because B80 has been discussed for years but long predicted unstable by many computational studies, especially standard density functional theory workflows. The authors argue that B80 is electronically comparable to C60 with similar valence electron counts, larger diameter, and potentially stronger electron affinity. They propose potential roles as a semiconductor, hydrogen-storage medium, or doped superconductor, while emphasizing that bulk synthesis has not yet been achieved. Practical barriers include boron’s conductivity limitations, isotopic complexity, and the lack of scalable preparation methods. Planned follow-up work includes reactivity tests and studying larger boron clusters to test cage stability limits. Commentators and prior researchers praise the apparent breakthrough, yet note that the evidence remains indirect. The group’s strongest claim is that the photoelectron spectrum uniquely fits the buckyball structure, while skeptics argue independent confirmation is still missing and that previous theoretical disagreement may be partly overstated depending on energy normalization. The overall picture is a potentially important but still provisional finding that awaits replication and material-level production.

The text reports a frozen, non-fine-tuned Llama that uses activation-derived readouts to decide when to run a calculator for arithmetic, rather than relying only on prompt parsing, and that this mechanism can pass stronger controls than standard text-driven tool use. In this setup, four operations were targeted—multiplication, division with remainder, gcd, and lcm—with behavior defined as firing on genuine arithmetic prompts and staying silent on hard negatives designed to look tempting. Across 11,736 locked examples with fixed examples, thresholds, and scoring, the method produced large exact-answer improvements while recording zero activations on the constructed adversarial hard-negative suite. The approach was further tested on a filtered slice of the DeepMind Mathematics Dataset, limited to prompts that matched supported two-integer forms and checkable answer formats, so only gcd, division with remainder, and lcm had sufficient sample power. On 3,822 locked examples and 1,233 targets, routed exact rates were high and lifts over the frozen model were +0.810, +0.502, and +0.968 respectively for division with remainder, gcd, and lcm. The mean exact-answer rates after routing were 0.992, 1.000, and 0.980 across those operations. Multiplication was not claimed because the filtered slice lacked enough accepted two-integer multiplication prompts for a statistically powered result. Example prompts showed that normal arithmetic wording should trigger the route, while unrelated arithmetic-like text should not, indicating a distinction between intent and superficial pattern matching. The work frames the gains as evidence that the model is correcting many arithmetic failures rather than simply reproducing what it already knew.

The source post is framed as news that Microsoft and Nvidia are launching the RTX Spark, a Windows-focused system built around a unified memory, CUDA-capable design intended to support more local AI and desktop workloads than cloud-only models. The product is presented as a possible shift toward stronger endpoint intelligence, with messaging emphasizing local execution, integrated memory utilization, and a Microsoft-Nvidia strategy to broaden AI availability beyond pay-per-token cloud usage. In the linked discussion and linked references, the announcement is treated as a notable departure from traditional PC and add-in-card architectures, positioning a high-end SoC-first form factor as a new direction for laptops. The coverage also hints that cost, memory sizing, and software ecosystem choices will determine whether it is adopted outside early enthusiasts, not just raw specs. The reaction captures uncertainty about whether the launch materially changes compute efficiency relative to prior platforms.

The post introduces zeroserve as an open-source, zero-configuration HTTPS server that serves a single tarball site with atomic hot-reload, io_uring-based I/O, and userspace eBPF middleware running on every request in one process. It replaces traditional directive/config files and plugin scripts with one ordered chain of eBPF programs that can inspect and rewrite requests, set response headers, emit JSON, generate token-bucket rate limits, perform AWS SigV4 and OIDC login, or reverse-proxy to backends, with helper APIs for cryptography, encoding, and body parsing. At deploy time, scripts in .c are compiled to eBPF objects and executed in a sandboxed runtime that enforces memory caps, preemption, and call-depth bounds, while metadata maps feed response headers and template placeholders. Benchmarks on Ryzen 1-core runs show zeroserve ahead of nginx and Caddy on small static files and small-body proxying, close on large static files, and slower on large proxied responses where nginx buffering performs better. In a direct comparison with nginx Lua, tuned preemption settings (10 ms) let zeroserve usually outperform, while default 2 ms preemption favors safety over throughput. The server also reports a low idle memory footprint and claims per-core scaling via many shared-binary processes rather than threads, emphasizing its architecture over raw process complexity. Commenters broadly engage with these claims, praising the idea while noting that many teams still prefer declarative configuration and dynamic ecosystem maturity over an eBPF-in-C model; concerns also target trust, hardening proof, and practical fit for workloads beyond static or small responses.