Sunday, 21 June 2026
Cuba opens its economy after six decades; music charts go local worldwide; AI plagiarizes the Dictionary of Obscure Sorrows
Sibexi
epoll, the standard Linux event-notification mechanism since 2002, uses a readiness model: it signals when a file descriptor is ready, but the application still makes separate syscalls for each read and write, incurring context-switch overhead on every operation. io_uring, introduced in kernel 5.1 (2019), replaces this with a completion model backed by ring buffers shared between kernel and userspace: applications deposit operations into a submission queue and harvest results from a completion queue without per-operation syscalls, enabling true asynchronous batching at near-zero syscall cost. The performance gap is substantial for I/O-intensive workloads. The tradeoff worth noting is that io_uring has accumulated a meaningful CVE history since its introduction — its kernel-side complexity is significantly higher than epoll's, and a number of privilege-escalation and memory-corruption bugs have shipped with it. For new systems work where performance is the priority, io_uring is increasingly the right default; for security-sensitive or constrained deployments, the CVE record warrants scrutiny before adoption.
Read →Vinicius Brasil
When I Reject AI Code, Even If It Works
The real bottleneck in AI-assisted development is not generation speed but review quality, and reviewing AI output demands more engineering judgment than writing code manually, not less. Vinicius Brasil identifies five conditions under which he rejects AI-generated code even when it passes tests and produces correct output: when he cannot explain the approach, when the diff is larger than strictly necessary, when it introduces premature abstraction, when it obscures system logic, and when he has been accepting output without understanding it. The underlying argument is that comprehension is not separable from code quality — a solution you cannot explain is a liability regardless of its current behavior, because you will be unable to debug it under novel conditions, extend it safely, or reason about its failure modes. The post is a useful counterweight to velocity-first AI adoption narratives: generating more code faster is only valuable if the engineers responsible for that code can still reason about what it does and why.
Read →Marc Brooker
Meet Alice. Alice Is Impatient.
Marc Brooker explains the inspection paradox and its implications for how services measure versus how users experience latency. If you sample a random moment in time and ask 'how long is Alice's current wait?', you are statistically more likely to land inside a long wait than a short one — because long waits occupy more time. The result is that the mean wait time experienced by a random observer is E[X²]/E[X], the second moment divided by the first, which for high-variance distributions (the common case in real web services) substantially exceeds the mean the operator measures. The practical implication is not just measurement humility but a design principle: optimizing for mean latency systematically underweights the experience of the users most affected by your system's variance. The paradox applies broadly wherever ongoing processes are sampled — queuing systems, network buffers, on-call rotations — and it is a good lens for diagnosing why service-level objectives written in terms of means consistently surprise engineers whose users report worse experiences.
Read →SMPTE
SMPTE Makes Its Full Standards Library Freely Accessible
SMPTE (Society of Motion Picture and Television Engineers), which defines the technical standards underpinning professional media production and delivery — timecodes, broadcast formats, color spaces, IP-based media transport, and increasingly AI content-authenticity frameworks — announced that its full catalog spanning over 110 years is now freely accessible to anyone globally, backed by eleven Diamond-level Corporate Members including Amazon AWS, Apple, Disney, and Sony. Previously, accessing these standards required organizational membership, which created a compliance cost that disproportionately affected smaller vendors and open-source toolchains. The organization is also adopting GitHub-based workflows and streamlined publishing pipelines to accelerate delivery of standards in fast-moving areas like IP production and content provenance. Open standards libraries typically accelerate interoperability and lower the entry barrier for new entrants; this one covers infrastructure that touches almost everything in professional video, from capture through distribution.
Read →Random Oracle
Mark-of-the-Web and Pinning Installers to Sites
Windows' Mark-of-the-Web (MoTW) is an NTFS alternate data stream that records the source URL when a file is downloaded from the internet — the mechanism behind SmartScreen warnings and Office protected view. The typical security concern around MoTW is attackers stripping it to evade detection. This article inverts that logic: legitimate software can read its own MoTW metadata and refuse execution if it was not downloaded from the expected official distribution URL. The technique is 'installer pinning' — the author demonstrates it with a ScreenConnect installer that checks its own Zone.Identifier stream and exits if the recorded source domain does not match the company's legitimate distribution site. This prevents a common attacker move: rehosting a legitimately signed binary on a malicious domain to exploit users who trust the signature without verifying the source. It converts a metadata field usually treated as an obstacle to attackers into an active trust assertion by the software itself — a small, practical defense-in-depth technique that costs almost nothing to implement.
Read →Waxy
The Wholesale Plagiarism of Obscure Sorrows
A web design agency named Qontour built an unauthorized clone of John Koenig's Dictionary of Obscure Sorrows — a twelve-year project that coined new words for emotional experiences lacking names in English — and deployed it as a portfolio showcase with affiliate links generating revenue from Koenig's own book sales. The clone reproduced the full text, replaced Koenig's artwork with AI-generated images, and then retroactively added a copyright disclaimer while simultaneously relicensing the work under Creative Commons — treating an authored creative project as a content source to be redistributed. The cloned site now ranks higher in search results than the legitimate one, and AI chatbots cite it as an authoritative source. Koenig's project is a particularly apt target for this analysis: its value comes entirely from the human attentiveness that noticed these emotions lacked names, constructed words for them, and crafted definitions — precisely the kind of meaning-making work that AI cannot replicate and that AI-powered plagiarism can most efficiently strip from its context. The case illustrates how AI tooling has lowered the cost of plagiarism to near zero while simultaneously undermining the original's discoverability — a combination that transfers economic value without transferring the creative labor that generated it.
Read →Lorenzo Gravina
Lorenzo Gravina observes an unconscious preference that has emerged in his reading: he gravitates toward books published before 2022, the rough date when large language models became widely accessible to writers and publishers. The preference, he acknowledges, may be empirically unjustified — most books published after that date were written by humans under ordinary conditions of effort. But the loss of the guarantee has changed something in the reading experience: phrases that feel unexpected or precisely right now carry a small epistemological uncertainty, a question about whether they were arrived at or generated, that cannot be fully set aside once it has been raised. Gravina draws the parallel to historical anxieties — writing displacing memory, printing devaluing manuscripts — and notes that those prior concerns did not prevent the new technologies from enriching expression. The relevant difference may be timescale and pervasiveness: earlier transitions took generations; the ambiguity about AI in writing is immediate and applies to almost every new text simultaneously. The essay is a clean account of how a technology can alter the phenomenology of an experience — reading — without changing anything observable about the texts themselves.
Read →