HackerNews Digest

Google Search now rewrites organic result links to google.com/goto?url=… instead of exposing the destination URL directly in the HTML. The url= parameter uses a proprietary encoding that cannot be decoded offline; the real target URL is provided only in the HTTP Location header of the /goto response, which must be read (e.g., via a HEAD request) without following the redirect. This change, observed consistently from late August 2026 for logged‑out and private sessions, replaces the older google.com/url?q= wrapper and forces each result to require an additional request to Google to obtain the final URL. The shift is part of Google’s broader anti‑scraping measures aimed at increasing the cost and traceability of automated SERP harvesting by AI crawlers and SEO scrapers. In response, Autom has updated its Google Search pipeline to resolve /goto links by reading the Location header and return the final URLs in its API, preserving existing integration behavior.

The page announces the 2026 Clay Research Awards. Recipients are grouped into three teams: Tuomas Orponen (University of Jyväskylä), Pablo Shmerkin (University of British Columbia), Hong Wang (IHES and NYU), and Joshua Zahl (Nankai University); Robert Burklund, Jeremy Hahn, Ishan Levy, and Tomer Schlank; and Yu Deng together with Zaher Hani. The award citation highlights the “remarkable” contributions of Orponen, Shmerkin, Wang, and Zahl. The page also includes four images with alt‑text captions: C. Fukushima and J. Westerweel of the Technical University of Delft; a photo of mathematician Martin Bridson; Anna Skorobogatova; and a generic image representing the Clay Research Award.

The author ran a Google Ads campaign for the Android puzzle app Dayzle at CA$40 per day, targeting installs. After removing the target CPA, the campaign spent CA$80 daily and reported 21 installs, but the app’s admin panel logged only one. Raw analytics showed 21 new devices, 20 of which ran an outdated app version that Play Store no longer serves—indicating the installs came from an external source despite being credited to Google Play. These devices opened the app once, recorded zero screen time, and never returned, spanning 28 phone models across 19 US states. Over two weeks, 56 billed installs occurred: 33 matched this bot‑like pattern, 7 originated from untargeted countries, and 13 were genuine users who completed 92 games. The author attributes the fraudulent installs to a bot farm that watches the ad video, then installs a locally cached APK, exploiting Google’s install‑based optimization loop. He has filed an invalid‑traffic claim, switched the campaign goal to “won a puzzle,” and advises developers to audit install metrics for similar bot activity.

The article documents a reverse‑engineering effort of the Apple Neural Engine (ANE) used in M1 chips, focusing on its internal architecture rather than trying to make it a general‑purpose accelerator. The ANE comprises 16 compute cores, each with 128 FP16 (or 256 INT8) MAC lanes, giving 2 048 parallel MACs that perform fixed‑point multiply‑accumulate with a 32‑bit accumulator saturating at 2¹⁵. Dot‑product, convolution, and attention operations are distinguished only by data‑mapping and scheduling, while the post‑MAC activation is applied point‑wise via a 33‑entry piecewise‑linear lookup table (mode 2) with linear interpolation and a scaling factor R that controls knot spacing.

The driver does not issue high‑level opcodes; instead it loads a pre‑compiled command stream of fixed‑size task descriptors (TDs) into a task queue (TQ). Each TQ holds status, priority, two descriptor slots, and a 32‑entry BAR table for virtual‑address base relocation. Submission uses TM_ADDR, TM_INFO, and TM_PUSH registers, resembling a GPU push‑buffer. The hardware walks the descriptor stream, executes DMA moves, kernel configuration, and PE operations without dynamic address calculations, indicating a highly static, data‑flow‑oriented execution model. The analysis concludes that ANE’s design reflects early CNN workloads and that later integration of its cores into the GPU marks the end of the standalone NPU.

Lake sturgeon in the Great Lakes may live 200–400 years, far longer than the previously estimated 150 years. Researchers led by Ed Baker (Michigan DNR) analyzed over 40 years of capture‑recapture data, tracking individual growth between captures and modeling time to reach adult size. A male caught in 1990 grew only 7 cm over 34 years (≈0.21 cm yr⁻¹), yielding century‑scale age estimates. This suggests many extant individuals could have been alive when Europeans first explored the region. The species, once abundant, lost ~99 % of its historic population due to overfishing and habitat loss; it is now listed as at‑risk in Ontario and endangered across the Great Lakes. Conservation plans, currently a century‑long effort, may need revision to accommodate the sturgeon’s extreme longevity. Indigenous perspectives view the fish as “grandfather” elders, reflecting its ancient lineage (>150 Myr) and cultural significance. Ongoing recovery includes hatchery releases and training programs with First Nations communities.

The paper “A Design Space Exploration of Async/Await” analyzes straight‑line asynchrony across modern languages (Python, Rust, Swift, C#, JavaScript, Tokio, Smol, Trio). Using a simple pseudocode example—logging in a background task and awaiting it—the authors show that seven runtimes produce four distinct outputs, and no two produce the same result for all three program variants. They attribute this variance to design dimensions that affect observable semantics rather than performance. Nine dimensions are identified and grouped into three lifetime phases: Start of Life (e.g., eager vs. lazy task creation), End of Life (e.g., dynamic vs. indefinite extent, destruction policy such as cancellation or awaiting), and Cancellation (e.g., propagation direction, cancellation awareness). Two dimensions—Dynamic Extent and Destruction—explain why Swift prints “AC” while Trio prints “ABC”. The authors formalize the space with a core calculus and small‑step semantics, highlighting rule choices that cause divergent executions. The work demonstrates that async/await designs are far from uniform, each reflecting trade‑offs in ergonomics, memory, and semantics.

• Jamais vu: feeling that a familiar word, face, or place seems new; opposite of déjà vu; >50 % of participants experience it after ~30 repetitions in lab; prevalence unknown, thought rarer than déjà vu.
• Prosopagnosia (face blindness): difficulty recognizing faces, congenital or acquired after brain injury; estimates 1–2.5 % (≈1 in 100–1 in 40) depending on testing.
• Aphantasia: lack of visual mental imagery; ~1 % report no images, ~4 % have faint or no images, a minority have hyper‑ vivid imagery.
• Clinodactyly: permanent forward bend of the middle joint of a finger, usually the little finger; occurs in <1 % of people; mild cases are asymptomatic but may progress with growth.
• Accessory thumb‑index tendon: when the thumb flexes into the palm, the index fingertip bends simultaneously; prevalence averages ~20 % but reported ranges span 5–60 %.
• Oculocardiac reflex: pressure on the eyeball or traction on an eye muscle can induce bradycardia; observed in 14–90 % of eye‑muscle surgeries, especially in children; not a safe method for self‑induced heart‑rate reduction.

• CSS provides two media queries for foldable devices: @media (horizontal-viewport-segments: <count>) targets side‑by‑side screens (book mode); @media (vertical-viewport-segments: <count>) targets stacked screens (laptop mode).
• A set of CSS environment variables (viewport-segment-width, viewport-segment-height, viewport-segment-top, viewport-segment-left, viewport-segment-bottom, viewport-segment-right) return the dimensions and position of each display segment using a 0‑based (x, y) coordinate grid.
• These variables let developers create responsive layouts that adapt to any foldable geometry without hard‑coding device‑specific sizes. Example: using grid-template-columns: env(viewport-segment-width 0 0) 1fr; inside a horizontal‑segments media query creates a two‑column grid where the first column fills the left screen.
• Adoption should be guided by user analytics; forecasts predict >21 M foldable devices shipped in 2023.
• Current browser support: Microsoft Edge stable implements both media queries and environment variables; Chrome stable lacks support, requiring the experimental flag (chrome://flags) or the Surface Duo emulator.
• The article serves as a concise overview; a deeper implementation guide is available on Smashing Magazine.