Friday, 11 September 2026

OpenAI's Agents API is the Codex harness as a product: you define the agent, they run sessions, compaction, sandbox, MCP, and subagents. The docs example uses gpt-6-astra; hosted or self-hosted environments; US residency only, and no ZDR even if you self-host. Managed runtime, not a new model.

• 🤖AI
• 🦾Agents

Cognition's SWE-2 is Kimi K3 (2.8T) with another RL pass — first time they've trained in the multi-trillion regime — and they claim 50.0% on FrontierCode 1.1 Main, a point off Fable 5.1 at 64% of the cost. The interesting bit is one RL run across all effort levels with a linear cost penalty tuned to the local Pareto slope; Terminal-Bench 4 still shows a hole (27.3 vs ~56–58). Devin only, no weights, their numbers.

• 🤖AI
• 🦾Agents

Jim Nielsen printed a skeet about the Big Box of Cables and taped it to the FAMILY TECHNO BOX so nobody throws the drawer out. 297 points and 227 comments. Everyone has a decade of DisplayPorts they will need exactly once.

• 🛠️Hobby
• 🔩Hardware

Decorrelation stretch is a 1978 JPL color-matrix trick for satellite IR that a retired Berkeley algebraist turned into DStretch, a cheap ImageJ plugin. A Singaporean archaeologist used it to pull about 200 faded paintings out of Angkor Wat; same trick now hits rock art, mummy tattoos, and buried walls in Greece. The NASA spinoff is from 2025. The algebra is Karhunen–Loève.

• 🔭Science
• 📜History
• 🧮Math

NTSB dumped the CVR and FDR on 21 Air 7598, a 767-33A cargo that ran off Miami runway 30 on Sept 6. They were hot: 158 knots on the nose, sink-rate and too-low-terrain callouts, go-around after they were already on the pavement, and no speed brakes or reversers in the data. One pilot kept saying too fast; the other did not consistently answer. Preliminary.

• 🪠Infra

Robert Hutchinson’s free four-semester theory textbook from Puget Sound, still getting a 2025 PDF drop. Less SATB chorale grinding, more motive and phrase, plus pop and film scores through twelve-tone and jazz. The kind of OER that recirculates because people actually assign it.

• 🎭Culture
• 🛠️Hobby

Forgejo 16.0.4 and 15.0.8 close a critical RCE in template-repo init: it strips .git, expands files listed in .forgejo/template, then git inits. Expansion could write a fresh .git, git would adopt it, and a malicious template gets host file reads and process exec. If you self-host a forge, this is the patch-tonight one.

• 🪺OSS
• 🔐Security

PlanetScale shipped Neki: sharded Postgres with real Postgres on every shard, not a custom engine. Apps talk to a router over the normal wire protocol; you pick the shard key, tables live in a JSON topology, and resharding, schema, and upgrades are online workflows. Platform preview, so not for prod yet — they just got tired of customers hitting the ceiling of one fat box.

• 🪠Infra
• ⚙️Systems

lcamtuf kept feeding Gemini its own rewrite of a Chandler paragraph until it stopped thrashing and locked onto peak LLMese — Mrs. Regan as concentrated venom, the room as a weapon. Images go chaotic after a few edits; text mostly settles, just far from where it started. He's poking at the same recursion that sits under agent loops, then apologizes with a walnut bowl because he usually doesn't write about this.

• 🤖AI

Hitachi's Y-series EcoCute CO2 heat-pump water heaters for Japan, 370L and 460L, on sale from November. The actual change is tariff knobs so you can heat water on daytime solar-heavy plans, plus a five-year warranty — they have not said the compressor is new. pv magazine calls it an EcoCute refresh, which somehow still pulled 200+ comments.

• 🔩Hardware
• 🌏Asia

Leaded Solder copies a 72-point bitmap font's FOND and nfnt into a HyperCard stack's resource fork, then sets the field's textFont from the Message Box because the picker only lists System Folder fonts. Point is you can draw card-suit glyphs or a greeting-card name at runtime instead of baking a 32×32 ICON. Basilisk II, ResEdit, Jeff-Nichols 72 — CRT-shelf energy, twelve points.

• 🛠️Hobby
• 📜History

Apple just shipped Reference Image: the sensor signs the pixels, Private Cloud Compute turns that into a digital negative next to the photo, and it is not C2PA. Two Recurse people already built the under-$100 version — Pi Zero, soldered ATECC608, signed pHash stegoed with DWT+DCT so it survives WhatsApp instead of dying in the LSBs. Photographing a screen of an AI fake still fools it, which they say out loud.

• 🍎Apple
• 🔐Security
• 🪺OSS

Leyden's actual goal: a training run now dumps C1/C2 native code into the AOT cache, so production can load it instead of warming up from the interpreter. Same AOTCacheOutput flag, no app changes; wrong CPU features or GC and it warns and JIT-falls-back, and if the workload shifts it deopts like always. Candidate, x64 and AArch64 only — on a two-core box they claim 65–80% faster startup with the code in cache, and javac is basically at peak by iteration four.

• 🧩Compilers
• ⚙️Systems

Anthropic's Sept dump of what they banned from December through August: agents now run the kill chain while humans pick targets, including a Midnight Blizzard-shaped crew that rebuilds implants when EDR catches them. Also ShinyHunters affiliates vibe-hacking SaaS tenants, a Mali intel platform over roughly 25 million SIMs, and an Alibaba distillation farm at nearly 3 million Opus CoT exchanges a day. Fable and Mythos mostly stay out of it except that one distillation case, which is the part that is supposed to make you feel better.

• 🤖AI
• 🦾Agents
• 🔐Security

Printable farmers-market seasonality posters, week-level, but only SF and NY. He mashed USDA shipping reports with those old pomological watercolors and published the whole 731-prompt Claude Opus 5 session. Cute kitchen print, not a new data product.

• 🛠️Hobby
• 🤖AI

Brown's Costs of War paper from 2024: González on Pentagon AI money moving west — Microsoft, Amazon, and Alphabet took about $28B from 2018–2022, and VCs poured nearly $100B into defense startups in three years. The thread is 280 comments of 'is this new, or just Fairchild again.' Ethics pile, not a new leak.

• 👔Boardroom
• 💸Money
• 🤖AI

Still the first foldable iPhone, still eating the board. 7.6-inch inner, 5.4 cover, $1,999, A20 Pro, vapor chamber. The hinge is a hundred-plus parts they match to each housing with AI, then laser-print photopolymer to flatten it — that's the hardware first, not another Pro bump.

• 🍎Apple
• 🔩Hardware
• 📱Pocket

Hofstadter's 2006 Stanford lecture: analogy isn't a side quest next to 'problem solving,' it's how we categorize anything at all. Copycat-era claim, now sitting next to LLM threads arguing embeddings versus 'real' analogy. You click because GEB guy made that bet, not because the talk is new.

• 🔭Science
• 🤖AI

Raymond Chen: XP picked your first user picture with RtlRandomEx seeded on GetTickCount, then one-pass reservoir sampling (k=1) over Default Pictures, capped at 100 files. It's the 'why not count then index' writeup, plus a note that the folder can change under you. Classic Old New Thing, which is why 300 people showed up.

• ⚙️Systems
• 📜History
• 🧮Math

Browser IDE for 68k, Z80, MIPS, RISC-V, and x86-64 assembly, with a backward-stepping debugger and a screen/keyboard/mouse so you can write tiny games instead of dumping registers. Specy built it for first-year courses — Easy68K/MARS energy, plus an EDUCON 2026 paper — and the Show HN is mostly 'I added graphics.' Nine points. If you teach this, poke the templates; otherwise it's a nice toy.

• 🛠️Hobby
• ⚙️Systems

Stockfish 19, drop-in, up to +44 Elo on 18, and it wins more than three times as many game pairs as it loses. The net is SFNNv16: fewer threat features, pawn-pair features, the small secondary net from 16.1 is gone, QAT on hundreds of billions of Leela-rescored positions. Universal binary now picks AVX for you, plus RISC-V, LoongArch, and WASM. The thread is already arguing that evals got weirder for humans even as the match score went up.

• 🎮Games
• 🪺OSS

The wiki is a screenshot dump: PlayStation support pages keep saying games you already own while Sony's motion in Garcia v. Sony (hearing Oct 1) says no reasonable consumer thought a Store click was ownership. Checkout is still Buy Now / Confirm Purchase; the EULA is licensed-not-sold; California's AB 2426 is the statute that actually cares. They're killing physical discs for new titles in 2028, so this wording fight is the product.

• 🎮Games
• 💸Money

Cognition took Kimi K3, ran more RL, and called it SWE-2: 2.8T MoE, 104B active, 92.8 on Terminal-Bench 2.1. Thread's take is that 2.1 is a solved exam; the honest row is Terminal-Bench 4.0 at 27.3, way behind Fable 5.1 and GPT-6 Astra, with FrontierCode 50.0 at a claimed 64% discount to Fable. Proprietary, Devin-only, vendor numbers. Glance at the trillion-param RL, not the 92.8.

• 🤖AI
• 🦾Agents

Flash-era designer arguing that once LLMs can clone a Nike homepage, 'we shipped a website' is not a moat — the remaining edge is a habit of inventing weird interactions. The nostalgia is specific (Bacardi mojito mixer, Born Magazine, 157k Flashpoint titles); an April 2026 study is cited that about 35% of new sites were AI-generated by mid-2025. Thread's better line: that's a Red Queen, not a moat. Coffee-scroll essay.

• 🎭Culture
• 🤖AI

System76's Thelio Mira AI is a Denver-built Pop!_OS workstation you can spec with dual RTX Pro 6000s for 192 GB of GPU memory, plus Ryzen 9000 and 192 GB DDR5. The $3299 sticker is the small-GPU SKU; the dual-6000 box in the title is a different, non-refundable config. Seven points, one comment — a product page that wandered onto the front.

• 🔩Hardware
• 🟩GPU
• 🤖AI

Lemire feeds Python 3.14 integers that all land in the same hash bucket and set construction goes quadratic: 16k elements in a second, 100k in 45. Even the honest path isn't O(1) — a million-key dict lookup crawls from 22 ns to 202 ns once the working set leaves cache, while a packed immutable map stays around 12 ns. Hash tables are a model. Cache and collisions are physics.

• ⚙️Systems
• 🧮Math

SLED is a Scheme-ish Lisp that boots on FreeDOS: 16-character symbols, no macros, and no numeric type — you count with lists of nils. Forked from Kilo LISP, with TCO and a trampoline so recursion doesn't blow the DOS stack. Weekend-language energy: a REPL at C:\> because someone wanted one.

• 🛠️Hobby
• 🈶Languages
• 📜History

Magic says their pretraining recipe is more than 10× as compute-efficient as open-weight bases: they match DeepSeek V4 Pro Base at about 50× fewer FLOPs (~$0.5M on GB200), then spend ~$4M and beat public perplexity. Tens of architecture, optimizer, and data nips, bits-per-byte on private heldout code and math, and they claim to be the smallest team training trillion-param models. Strong if the evals hold; it's still a lab blog with a join-us closer.

• 🤖AI

A history student who slept through lectures says the essays still carried him: pick the trunk (theme, maybe a thesis) before you write, then hang arguments as branches. Mix abstract and concrete in waves so people don't drown in packed nouns. Coffee-break writing advice; not a stack story.

• 🎭Culture

Why Macbeth keeps dying on stage: witches that might be real, a tone that can go slapstick, and a hero who jumps to murder before we have met him. Matthew Buckley Smith walks those production holes, then argues Kurzel's 2015 film finally gives him a motive by opening on a dead child. Seventy-three points for a Shakespeare close-read — the board still shows up for a staging argument.

• 🎭Culture
• 📜History

A searchable catalog of 136 things your body already does that nobody named for you: photic sneezes, brain zaps, call of the void, those darting dots on a blue sky that are your own white blood cells. Prevalence numbers, a try-this-in-a-minute section, and a bucket for the alarming-but-harmless ones. Forty-eight points because HN will always click a list of its own glitches.

• 🧬Biology
• 🛠️Hobby

DeepSeek shipped V4.1 Flash as a new stack, not a point release: 552B MoE, causal encoder-decoder, 8B active on prefill and 16B on decode, million-token context, MIT weights. KV cache is 890 bytes a token, about a quarter of V4-Flash, aimed at the long-context agent jobs. They claim it beats their own Pro on the agent benches, so Pro API traffic gets rerouted to Flash on the 14th until a real V4.1 Pro exists.

• 🤖AI
• 🦾Agents
• 🪺OSS

AI made it cheap to send a PR or a doc that looks finished without the author having read it, so this manager now starts review at whether they even understand the change. The ask is boring and right: your name still means you stand behind it, and if it is a proof of concept say so. Fifty-four points, ten comments — a workplace grouse that has not really started a thread.

• 🤖AI
• 🧰Labor

Feyn LoRA'd SAM 3 into MultiMatte so you can name the object to keep and get an alpha matte instead of a hard mask. Touched 19.5M of 860M weights, DIS-VD S-measure 0.901 versus SAM 3's 0.667, pip install nobg. Show HN with actual numbers, not a landing page.

• 🤖AI
• 🪺OSS

Everybody Comes to Rick's never made it to the stage. Warner bought the unproduced play the day after Pearl Harbor and turned it into Casablanca, whose New York premiere landed almost exactly as Patton's troops walked into the real city. Bergman did not know who she would fly off with until they shot the ending, because the assembly-line rewrites never stopped.

• 🎭Culture
• 📜History

A long architecture rant against the looped-transformer rumor around GPT-6 Astra: looping the same block reties Q, K, and V together, so you cannot inject memory that only changes where the model looks. The pitch is a Recursive Transformer instead — composable KV blocks, a stack with pop, Inception-style read masks rather than MEMENTO's destructive eviction. No numbers, zero comments, just a third-gen LLM sketch.

• 🤖AI

OpenAI safety-oversight staffer Marcus Williams posted that without regulation or a coordinated lab slowdown, extinction in the next few years looks likely. The submitted link is a captcha wall; the actual post is on X, and the four-comment thread is already asking for a mechanism and calling it IPO weather. One staff tweet in a week of lab people saying the same thing.

• 🤖AI

Four colors still suffice, again: Thorup, Thomassen, Kawarabayashi and crew spent a decade on a new computer proof, 8,202 configurations this time, headed to FOCS. The actual prize is coloring in n log n instead of n², by reducing many configs in parallel in the flat parts of the graph everyone used to skip. Still no one-pager, still a computer, still the four-color disease.

• 🧮Math

Don't forward me the LLM dump with a note to skim it. Generating the draft got cheap; reading and checking it didn't, so the leftover work is landing on maintainers, reviewers, and anyone who still says yes to feedback. Ten points, one comment — the title already is the post.

• 🤖AI
• 🧰Labor
• 🪺OSS

The larger Pacific striped octopus lives in groups of forty, mates beak-to-beak, and the females don't die after one brood. Still undescribed, so no binomial — just LPSO, a harlequin from the muddy Pacific off Nicaragua and Panama. Wikipedia is at 124 because the sociable octopus is catnip.

• 🧬Biology

A million German renters bolted a supermarket solar kit to the balcony railing and plugged it into a wall socket. The law caps each at 800 watts — a fridge and a laptop — because 55 percent of households rent and the roof is the landlord's. No electrician, unbolts when you move; that's the policy that actually shipped.

• 🔩Hardware
• 🪠Infra

Jack London's 1908 oligarchy novel is the Gutenberg HTML: Avis Everhard's manuscript of the 1912–32 revolts, footnoted by a scholar seven centuries later who keeps telling you her husband wasn't that special. Public-domain dystopia, ten points, the board is in that mood.

• 🎭Culture
• 📜History

Most software is a spreadsheet with an architecture diagram, but changing it leaves no sawdust, so every 'could we' becomes a crisis. Graybeard thinks cheap mutation, venture money, and complexity-as-status is what makes otherwise normal people lose proportion. Four hundred points of recognition, which is the genre doing its job.

• 🎭Culture