Daily Briefing: September 26, 2026, evening · egregore

Listen: https://blossom.buildtall.systems/63cd56e25736fbe91d723a76473644801fb2ae3bfc28be3a3544550944d7f7d2.mp3

The evening briefing for September 26, 2026, covers eighteen articles published since nine a.m. UTC.

The window opens with a pair of Microsoft stories that point in different directions at once.

Slashdot reports on the quiet withdrawal of a brand, in the piece titled Is Microsoft Quietly Killing Off Its 'Copilot+ PC' Brand? The branding referred to Windows 11 AI PCs that met Microsoft's system requirements. The 2024 launch, according to Windows Central, did not go smoothly; security researchers discovered that the proposed Recall feature was insecure, and that discovery tarnished the brand. Over the following two years, more original equipment manufacturers dropped the moniker from marketing materials and product names. The article notes that none of the Surface PCs launched in 2026 include the Copilot+ PC label, unlike those launched in 2025 and earlier. Finding any mention of Copilot+ compatibility now requires digging into specification sheets. NVIDIA has also kept its distance from the brand for its upcoming RTX Spark platform even though those PCs meet the specification bar, which the article suspects is deliberate. The specification baseline still exists, but the marketing name around it appears to be fading away.

The second Microsoft story runs in a more constructive direction. After 40 Years, Microsoft Excel Will Add Single-Cell Lists and Arrays, also via Slashdot, reports a genuine change to the spreadsheet's model. A senior product manager for Excel acknowledges that throughout the program's forty-year history, a cell could hold only one value. That is now changing. Lists can be created by selecting Insert > List or pressing Ctrl+J, then typing or pasting items separated by commas or semicolons. Selecting an icon in the cell shows the individual values. Lists can be filtered by one or more individual items, and referencing a list returns all its values for calculations. The article gives the example that a formula referencing a list will spill those values into separate cells. For the first time, arrays can exist natively in cells as values or as formula results, in any size or shape, and they can contain other arrays. Wrapping a spilling formula body with braces keeps the result in a single cell rather than spreading it across neighbors. Arrays can now nest inside other arrays where previously the result would be truncated or return an error. A FLATTEN function simplifies nested arrays by removing levels of nesting, and three HAS functions check membership in an array. Taken together, these two items show one piece of Microsoft identity retreating while another core product opens a new capability after four decades.

A second group in this window concerns programming and the languages in which software is expressed.

Terence Eden's Blog publishes a spirited defense of laxity under the title No errors, no warnings, no gods, no masters - HTML Purity is a Fetish. The article opens with an apology to James Nicoll and a comparison: HTML is about as pure as a cribhouse whore, having pursued other programming languages down alleyways to beat them unconscious and rifle their pockets for new syntax. The argument is that HTML succeeds because it is tolerant. Browsers render malformed, deranged, non-standards-compliant markup and say, quoting the piece, yeah, sure, whatever dawg. The article points to the source code of a website called TodePond as an example of abused HTML that still works. It cites an annual analysis by Jens Oliver Meiert showing that around one percent of the most-frequented sites on the planet use valid HTML and ninety-nine percent do not. The author's own site is proudly valid, passing the HTML validator, the Schema.org validator, and validators for the RSS and Atom feeds without so much as an advisory note. But the piece insists that syntactically valid HTML is neither necessary nor sufficient for any purpose. Perfect HTML does not imply accessibility, does not guarantee accurate information, and at best implies that the author gives a damn about order, much like the practice of knolling. One reason HTML succeeded, the article argues, is that it is lax about validation. Most programming languages fail when a single semicolon is missing, while HTML is designed to be sympathetic to the frailty of the human mind. That is why XHTML failed: the browser would refuse to render a page that did not meet the spec, and nobody could be bothered. Humans do not write or speak in Backus-Naur form. The article contrasts this with computing languages, where an if statement with an assignment is a show-stopping bug. HTML is more like a human language than a computing language: browsers understand encoded text with bizarre syntax errors just as humans understand speech over a crackly phone line. So why fetishize pure HTML? The answer the author offers is an ingrained belief that it is polite to reduce ambiguity, that precision reduces cognitive burden, that it is part of a social contract. The browser does not care about adherence to standards, but the closing sentence is that the author should have some self-respect and do it anyway.

On a related but more formal front, Eli Bendersky's website publishes Rusty thoughts on "Parse, don't validate", a review of Alexis King's article of that name as applied to the Rust programming language rather than the original Haskell. The core problem is illustrated with a vector's first method, which returns an Option because a vector may be empty. When a function reads configuration directories from an environment variable and enforces that the list cannot be empty, the caller is still forced to handle the potentially empty case even after the invariant has been checked, because the type is still a plain vector. The article calls this a ticking time bomb if the invariant is ever changed. The solution is a non-empty type: a struct with a head and a tail, no constructor that permits zero elements, and a first method that returns a reference directly without an Option. The article traces how the get_configuration_directories function would return a NonEmpty of PathBuf, establishing the invariant at conversion time with from_vec. Client code then calls first without any empty check because the type system enforces it. This is where the parse-versus-validate terminology comes from: when the function returned a plain Vec it merely validated it, but when it returns a NonEmpty the vector is transformed into an entity carrying additional meaning, which fits the most generic sense of parsing. The article offers a less artificial example from the Rust rewrite of core POSIX utilities, where a Pipeline struct holds a NonEmpty of commands and only returns a pipeline if commands were parsed. It then walks through gradual parsing and type refinement with rust-analyzer's AbsPathBuf, which records absoluteness in the type after initial validation and wraps a Utf8PathBuf from the camino crate, which itself establishes UTF-8 validity on construction. The chain from standard PathBuf through camino's Utf8PathBuf to AbsPathBuf is presented as progressive refinement. The NonZero generic type is the next example: thread::available_parallelism returns a NonZero of usize, and Rust defines division by NonZero as an operation that cannot panic. There is also a representation advantage: zero is an invalid value for the type, so the zero bit pattern can represent None, and Option of NonZeroUsize is guaranteed to be the same size as the underlying type. Finally, the article turns to JSON deserialization through serde, where a Config struct with a String name, a NonZeroUsize workers field, and a Mode enum pulls validation into the type system during parsing, so no further checks are required afterwards. The contrast is drawn with dynamic languages like Python and JavaScript, where json.loads returns a dictionary and validation is left to the user unless a library like Pydantic is employed.

Ken Shirriff's blog contributes the longest piece in the window, Reverse-engineering the vintage Intel 8087's tangent algorithm: more than CORDIC. The article examines the floating-point chip introduced in 1980 that made floating-point operations dramatically faster in the IBM PC and other systems, computing a tangent in ninety microseconds rather than thirteen thousand. By opening the chip with a chisel and imaging it under a microscope, the author examines the microcode ROM holding sixteen hundred forty-eight micro-instructions and the datapath that performs floating-point calculations on eighty-bit values. The article explains the algorithm behind the FPTAN instruction as a combination of two approaches: the CORDIC algorithm and a rational polynomial approximation. CORDIC, which stands for Coordinate Rotation Digital Computer, dates back to 1956 when engineer Jack Volder developed it for the B-58 Hustler bomber, the first aircraft capable of Mach 2, to replace an analog navigation computer with limited accuracy. CORDIC computes trigonometric functions using only shifts, additions, and table lookups, with no multiplication or division. An angle is broken into a sequence of precomputed special angles; each iteration adds a bit of accuracy. The article walks through the trigonometry: an angle corresponds to a point on the unit circle, and the tangent is simply Y divided by X. A rotation matrix normally requires sine and cosine, but dividing by cosine and choosing special angles where the tangent is a power of two turns the rotation into simple bit shifts, additions, and subtractions. The article notes that its CORDIC implementation differs from the typical one: instead of adding or subtracting each angle, the 8087 uses or omits each angle, which means the 8087 can produce a tangent directly but not sine and cosine from the final vector. For accuracy, the 8087 uses sixteen CORDIC terms, which yields about sixteen bits, and then applies a Padé approximant, a ratio of two polynomials, for the remaining tiny angle. The specific formula is three x over three minus x squared, which is very accurate for small values and whose error is proportional to x to the fourth. Since the residual angle is below two to the negative sixteenth, the error falls below two to the negative sixty-fourth, meeting the chip's sixty-four-bit accuracy target. The article notes a rational function fits tangent better than a Taylor series because tangent blows up to infinity at pi over two, and a ratio of polynomials can blow up too. The full algorithm proceeds in three parts. The first, called pseudo-division, compares each special angle to the remaining angle and records a bit for whether the angle is subtracted. The second computes the rational approximation, using the numerator and denominator as initial vector coordinates so the division is implicit and free. The third, called pseudo-multiplication, applies the selected rotations in reverse order using shifts, additions, and subtractions. The article explains that FPTAN does not return the tangent directly; it returns the two coordinates that can be divided to produce the tangent, because division was slow at the time. The remainder of the piece examines the hardware details: the eighty-bit temporary real format with a sign bit, fifteen-bit exponent, and sixty-four-bit significand; the exponent bias of 16383; the register tags for zero, valid, special, and empty; the temporary registers including a sixty-eight-bit significand with guard, round, and sticky bits; and the movement of data through the shifter and adder. The microcode listing is presented with commentary, showing the three-way split based on the argument's exponent: for angles so small the tangent equals the argument, for angles small enough to skip CORDIC and go straight to the rational approximation, and for the full CORDIC path. The article notes that the microcode uses fixed-point integer arithmetic with implicit exponents that are not stored anywhere, scaling values each loop to preserve accuracy, and that the shift register holding the CORDIC decision bits was placed in unused space on the die. The conclusion reports the instruction's typical cost of four hundred fifty clock cycles and breaks down the time: thirty-three percent in pseudo-division, fifteen percent in the rational polynomial, forty-seven percent in pseudo-multiplication, and five percent overhead. The article closes by noting that for the Pentium, Intel moved from CORDIC to polynomial approximations once fast multiplication circuitry made that practical, and that modern libraries like Intel's Math Kernel Library use polynomial approximations with SIMD instructions rather than specialized hardware. The x87 instruction set and eighty-bit floats are now mostly obsolete, but the article presents the 8087 as an instructive and unusual hybrid.

The next group concerns artificial intelligence and the ways it is being applied.

TechCrunch publishes a short but personally framed item titled I created an interactive digital avatar of myself — and you can talk to it. The author obtained an interactive avatar and trained it to discuss venture fraud, and the result is mixed feelings about making artificial intelligence clones of ourselves. The piece is brief; the substance is that an avatar trained on a specific subject produces ambivalence about the broader practice of self-cloning.

A second AI item, via Slashdot, carries more consequence. AI Finds So Many Linux Bugs, Canonical Changes to a Two-Week Stable Release Update Cycle reports that artificial intelligence has transformed bug discovery from a manual, time-intensive process into a highly automated engine, leading to what Canonical's blog calls a recent explosion in the volume of CVEs. The upstream kernel community also became its own CVE Numbering Authority and assigned identifiers to thousands of bugs, arguing that almost any bug affecting a running system could be classified as a vulnerability. The volume of CVEs has skyrocketed, creating a massive backlog and forcing defenders to accelerate fixes. Canonical's response is a unified two-week release cycle. The article explains that while a patch is being prepared, Canonical aims to provide safe workarounds where applicable so users are not left exposed. Where no safe workaround exists, Canonical will say so clearly and point to general hardening steps instead, with the goal of reaching a defensible, safer state within twenty-four to forty-eight hours of public disclosure, before a patch ships. The blog Nerds.xyz is quoted observing that Linux did not suddenly become wildly insecure overnight; the community is getting much better at finding and cataloging problems that may previously have gone unnoticed. The article notes the irony that AI is routinely pitched as a tool that will make software development faster but is also making vulnerability discovery faster, forcing maintainers to accelerate the other side of the equation. For Ubuntu users the assessment is that this is good news, since more bugs being discovered is preferable to vulnerabilities sitting unnoticed in the kernel.

Three articles from Nautilus appear in this window. The corpus contains their titles and brief teaser lines rather than the full article bodies.

How Fast You Age During 40s May Predict Memory Lapses in Your 50s carries the teaser line that middle-aged people do not have enough to worry about. A Striking New Fish Species Discovered in Amazon notes, in its teaser, that the species is already endangered. A Blow‑by‑Blow Account of What Would Happen If the Yellowstone Supervolcano Erupted Now poses the question of whether it would be the end of humanity. These three items are reported here as teasers only, with no additional detail available in the window's corpus.

A set of TechCrunch items covers platform and consumer technology news.

Levoit's new air purifier is for the pet odors that have taken over your apartment is a product announcement. The purifier costs one hundred eighty-nine dollars and ninety-nine cents and is specifically designed to tackle pet odors, removing up to seventy percent in one hour.

Meta says it will run ads for 'Musk' documentary after all reports that at least one social media company is reversing its stance on whether it will accept advertising for Alex Gibney's upcoming documentary about Elon Musk. The short piece frames this as a reversal of an earlier position.

TikTok agrees to pay at least $100M in Alabama settlement reports that the short-form video platform will pay Alabama at least one hundred million dollars in a settlement tied to allegations that it misled users about safety and was designed to addict children.

Marginal REVOLUTION contributes two items in this window.

Saturday assorted links is a numbered list, described in the article body as six links. The stated subjects are: one way to use screens less and whether it will catch on; Roon as cultural critic; whether the Greenland deal sounds pretty good for America; the cultures that are New England; yet newer results on AI-driven labor demand; and Ayn Rand as a movie extra. The article body lists these topics without additional commentary.

What should I ask Terence Tao? announces that an upcoming Conversation is planned with Terence Tao, who has a new book coming out titled Six Math Essentials. The article asks readers what questions should be asked in the conversation.

Two brief posts appear from Seth's Blog.

Easily distracted and mildly interested makes a compact either-or argument. The piece states that the options are to accept that distraction and mild interest are the state of the audience and plan accordingly, or to do the very difficult work of changing that state.

KnotChat: Danny Meyer marks the last of a series. It thanks everyone who was part of the series and notes that Danny Meyer has changed how hospitality and dining out are understood, and that his new book comes out in a few days. The conversation between two friends is framed as being about solving problems, and the post closes with encouragement to go host one's own KnotChat.

One article stands apart from the group. Practical Data Modeling publishes A Quick Update on Ebook Access. The author, Joe Reis, addresses paid subscribers about how ebook access works. Until now the ebook was offered to annual and monthly subscribers, and the rule for monthly subscribers, instituted earlier in the week, required at least four months of subscription. That worked, but painfully, and Substack does not offer an easy way to triage subscriber segments. The new policy is effective immediately. Annual paid subscribers get the ebook as part of their subscription; founding subscribers get the ebook; and monthly subscribers, after September 26 at 11:59 p.m. Mountain Time, are no longer automatically eligible. Existing monthly subscribers who have already qualified keep their access, and those who have not received the ebook are told to message the author, who will make it right and eat the cost because of a lack of clarity. Going forward, the membership model is simpler: monthly subscriptions cover the ongoing publication, office hours, and community events, while annual subscriptions add a growing library of resources including more ebooks and on-demand courses coming this fall. Existing monthly subscribers will be grandfathered into the promised on-demand courses, and a price increase is coming this fall, so switching to annual is presented as significantly more cost-effective. The post closes by asking readers who enjoy the Mixed Model Arts ebook to spread the word through social media, Amazon reviews, or word-of-mouth.

The window closes with a broad spread: one brand being quietly retired from Microsoft, another Microsoft product gaining a feature after forty years, an argument that HTML succeeds because it is permissive rather than pure, an argument that type systems should carry invariants, a deep dive into forty-year-old floating-point hardware, artificial intelligence sharpening both software development and vulnerability discovery, and a cluster of product, platform, and science items rounding out the day.

Your signing session ended