🔥 Today’s Focus

Today’s HN is ruled by two storylines. The first is AI capital settling its accounts: Stripe’s $7B+ acquisition of OpenRouter (92 points), Nvidia sharply scaling back its guarantees for OpenAI infrastructure financing (44 points), and Anthropic’s IPO valuation hinging on a $190-200B revenue forecast for 2028 (33 points) — three business stories hit the front page the same day, and the narrative propped up by “unlimited compute credit” over the past two years is coming due. The second is the endgame of browser ad blocking: Firefox for iOS ships native ad blocking (489 points) and “Firefox is the last major browser that still supports uBlock Origin” (Lobsters △192) sweep the front page, with uBlock Origin author gorhill personally stepping in to clarify the Manifest V3 limitations. On the technical side, the rebuttal to yesterday’s RISC-V rant (296 points) tops the charts — dmitry’s “high-end processors” and the embedded engineer’s “10-cent MCUs” talk past each other; and the single highest-scoring post of the day is Asus’s friction-drive bike kit (602 points) — Monday-morning HN mixing hardcore debates with consumer-electronics hype, as always.

🤖 AI & Models

  • Claude: System Prompts — Claude: System Prompts. 476 points / 209 comments (HN). 🔥 Today’s second-highest: Anthropic publicly releases the full system prompts behind claude.ai and mobile for the first time — the prompts carry a record of Fable/Mythos being suspended for 18 days over US Commerce Department export controls, and the model only learns what happened after its training cutoff through this injected text. 💬 simonw turned the successive prompt revisions into a git commit history for easy diffing: the most interesting change between Opus 4.8→5 is that Claude is explicitly told it ranks a tier below Fable/Mythos — at which point eterm floats a conspiracy theory: the supposed “Opus 5 getting dumber” may be exactly what this self-deprecating prompt caused; tosh complains the prompt eats the most valuable real estate in the context window, which is why his smol agent framework skips system prompts entirely.
  • Models Are Getting Dumber on Purpose — Models Are Getting Dumber on Purpose. 212 points / 128 comments (HN). An attribution piece on the “models are regressing” phenomenon — the 128 comments split between “distillation side effects” and “vendors deliberately dumb models down to save money”; no evidence, all position-taking.
  • The AI Credit Resale Economy — The AI Credit Resale Economy. 205 points / 77 comments (HN). People are mass-registering Anthropic/OpenAI accounts to farm free credits and resell them at a discount — the middleman-arbitrage economy has finally reached the token market; 77 comments crunch the arbitrage margins and ban risk.
  • Stripe Clinches over $7B Deal to Buy AI Firm OpenRouter — Stripe Clinches over $7B Deal to Buy AI Firm OpenRouter. 92 points / 70 comments (HN). The payments giant buys an AI model-routing aggregator — OpenRouter’s “one API to call every model” gives Stripe a ready-made distribution channel for AI apps; 70 comments debate whether this is the start of an AI-infrastructure M&A wave.
  • MathCode, Mathematical Coding Agent — MathCode, Mathematical Coding Agent. 48 points / 13 comments (HN). A coding agent aimed at math problems — using competition math as a benchmark is a clever move; 13 comments check its real performance on IMO problems.
  • Nvidia dramatically reduces amount of OpenAI infra financing it may guarantee — Nvidia dramatically reduces amount of OpenAI infra financing it may guarantee. 44 points / 5 comments (HN). Reuters: Nvidia narrows its guarantee scope for OpenAI’s $250B data-center financing — “the shovel seller starts putting credit limits on the miner”; all 5 comments ask what this means for the Stargate project.
  • Anthropic IPO valuation hinges on $190-200B 2028 revenue forecast — Anthropic IPO valuation hinges on $190-200B 2028 revenue forecast. 33 points / 36 comments (HN). Sources say Anthropic’s IPO valuation story rests on a forecast of $190-200B in revenue by 2028 — 36 comments do the math on how much of the cloud market that would require eating; most conclude it’s absurd.
  • Anthropic’s ‘Watermark’ Text Adulteration in Claude Is a Perversion of Writing — Anthropic’s ‘Watermark’ Text Adulteration in Claude Is a Perversion of Writing. 60 points / 37 comments (HN). Gruber opens fire: Claude quietly watermarking text is a perversion of writing — 37 comments argue over whether watermarks actually hurt output quality, and whether doing it silently counts as deception.

🛡️ Browsers & Ad Blocking

  • Firefox for iOS now has a native adblocker — Firefox for iOS now has a native adblocker. 489 points / 207 comments (HN). 🔥 Today’s third-highest: Firefox iOS ships built-in ad blocking, no third-party content blockers needed — on iOS, WebKit forces every browser onto the same engine, so Mozilla can only differentiate at the feature layer. 💬 simcop2387 says the iOS version only rates as uBlock Origin Lite-tier (Manifest V3 rule limits), and gorhill himself shows up ten minutes later to correct him: cosmetic filtering and scriptlet injection are enabled by default in Optimal/Complete modes, not the crippled version people claim — the uBO author personally debunking the thread is the most substantive comment section of the day.
  • Firefox is now the last major browser that still supports uBlock Origin — Firefox is now the last major browser that still supports uBlock Origin. △192 / 54 comments (Lobsters). Lobsters’ highest score today: after Manifest V3, Chrome/Edge can only run a crippled version, making Firefox the only home for full uBO — among the 54 comments, FRIGN’s worry is the most practical: Mozilla’s Google money will eventually dry up, but neither the “a lightly-maintained Firefox is good enough” optimists nor the “the browser feature arms race never stops” pessimists convinced each other.

🔒 Privacy & Trust

  • Tell HN: Cloudflare silently injects its analytics when you switch nameservers — Tell HN: Cloudflare silently injects its analytics when you switch nameservers. 219 points / 55 comments (HN). After switching nameservers to Cloudflare, a pure-HTML no-JS site gets beacon.min.js automatically injected — on by default, off only if you manually disable it; 219 points suggests plenty of people got burned. 💬 kevincox’s anger is the most representative: “I explicitly turned off all analytics and still got injected. Caching is not permission to modify my website”; altairprime digs up Cloudflare’s announcement 11 months ago that Web Analytics would be enabled by default — it scored only 2 points on HN back then — a default-on buried in newslop, and the backlash is deserved.

🔬 Chips & Hardware

  • A 3rd World Embedded Engineer Responds to “RISC-V They Should Have Known Better” — A 3rd World Embedded Engineer Responds to “RISC-V They Should Have Known Better”. 296 points / 156 comments (HN); △34 / 8 comments (Lobsters). Yesterday’s dmitry RISC-V rant gets a direct rebuttal today: from an embedded perspective, no licensing fees, customizability, and coverage from 10-cent MCUs to Linux-class SoCs is where RISC-V’s value lies — but the consensus across 156 comments is that the two authors talk past each other on the definition of “high-end”. 💬 phire’s take is the sharpest: dmitry’s high-end means laptop/phone/server-class application processors, while Subero thinks “bigger than a cheap MCU” is high-end — cores like VexRISC-V sit exactly in RISC-V’s sweet spot and start struggling once you push up into high-end application markets; madushan1000 lands another jab: x86’s architecture is genuinely bad, yet high-performance chips still rule — the performance gap is just a matter of who’s paying to close it.
  • St Lucie Nuclear Reactor Unit 1 manually shutdown, 3 control rods drop into core — St Lucie Nuclear Reactor Unit 1 manually shutdown, 3 control rods drop into core. 142 points / 105 comments (HN). Control rods unexpectedly dropping into the core triggers a manual shutdown — 105 comments explain that control rods dropping means scram, and scram means the safety mechanisms are working; nuclear fear and nuclear science go roughly 50/50 in the thread.
  • Chestnut – eGPU dock with open-source firmware — Chestnut – eGPU dock with open-source firmware. 128 points / 35 comments (HN). comma.ai’s eGPU dock with fully open-source firmware, $249 bare — 35 comments run the numbers on external-GPU latency loss and value for money.
  • Plastic mechanical computer from 1963: The Digi-Comp 1 [video] — Plastic mechanical computer from 1963: The Digi-Comp 1 [video]. 43 points / 9 comments (HN). An archaeology video of a purely mechanical toy computer — plastic sliders and springs implementing binary arithmetic; 9 comments fondly recall the first “computer” they assembled as kids.
  • Asus Bike Booster — Asus Bike Booster. 602 points / 423 comments (HN). 🔥 Today’s highest score: Asus turns a friction-drive motor into a quick-release kit that clips onto the rear wheel to electrify any bike — 602 points shows how deep HN’s obsession with “electrify without buying a new bike” runs. 💬 bri3d pours cold water on friction drives: fast tire wear, useless in rain and mud, and roughly 20% less efficient than other drive types; molsongolden recommends Bimotal as a twice-as-expensive but wear-free comparison — the comments range from price ($2k buys a whole e-bike) to NYC bike theft.

💻 Programming Languages & Systems

  • I thought I was building a C replacement. I was wrong — I thought I was building a C replacement. I was wrong. △56 / 31 comments (Lobsters). The C3 language author’s self-examination: he thought he was replacing C, then realized that “C replacement” in the public imagination means “replacing what C is used for today” — while C3 and Odin target general application programming, only Zig is seriously taking over C’s job. 💬 author clerno shows up with a TLDR: Zig aims at what C is used for today, Odin/C3 aim at general application development — and the saying “Zig is more C than C” finally makes sense; gulbanana adds a historical parallel: Go also once called itself “a better C”, but actually replaced Java, and ended up as the ops-script killer thanks to single-file cross-compilation.
  • Protobuf has LSP support. You’re welcome — Protobuf has LSP support. You’re welcome. 85 points / 54 comments (HN). Buf adds an LSP for Protobuf — 54 comments ask when it lands in editor defaults and why this took so long.
  • The Case Against Formal Verification, 50 Years Later — The Case Against Formal Verification, 50 Years Later. 63 points / 53 comments (HN). Re-reading De Millo/Lipton/Perlis’s famous 1979 paper — formal verification as a “social process” rather than mathematical proof; 53 comments argue over whether verification tooling in the LLM era counts as vindication.
  • Clamiga: Common Lisp for the Amiga — Clamiga: Common Lisp for the Amiga. 71 points / 7 comments (HN). Porting Common Lisp to a 1985 Amiga — all 7 comments ask the same question: why would anyone do this, and what does it look like running?
  • SIMD in the 90s: Programming Intel’s Pentium MMX — SIMD in the 90s: Programming Intel’s Pentium MMX. 47 points / 20 comments (HN). A retro tutorial on the MMX instruction set — 20 comments reminisce about MMX’s “fake 64-bit” controversy; a good entry point into SIMD history.
  • Protecting the Rust standard library from accidental breakage — Protecting the Rust standard library from accidental breakage. △32 / 4 comments (Lobsters). A proposal to bring semantic-versioning discipline and regression tests to the Rust standard library — all 4 comments say “should have been done long ago”.
  • A quick look at zero-knowledge proofs — A quick look at zero-knowledge proofs. 13 points / discussion ongoing (HN). bernsteinbear’s ZKP explainer, starting from the Schnorr protocol — 13 points is cold for today’s charts, but the quality is there.

🛠️ Database & Engineering Practices

  • Does anyone run Postgres without PgBouncer? — Does anyone run Postgres without PgBouncer?. 135 points / 85 comments (HN). Brandur’s existential question: is a connection pooler really necessary — the 85 comments split into “bare connections are fine at low traffic” and “production dies without PgBouncer”, with the middle ground being how many people PgBouncer’s config complexity scared off.
  • What’s missing to have reproducible builds on PyPI — What’s missing to have reproducible builds on PyPI. △17 / discussion ongoing (Lobsters). snarky.ca inventories the remaining obstacles to reproducible builds on PyPI — from timestamps in wheels to build-environment drift, Python packaging still has a stretch of road ahead.
  • A Tour of Magit’s Status Interface — A Tour of Magit’s Status Interface. △21 / discussion ongoing (Lobsters). A design teardown of Magit’s status interface through Emacs eyes — this explains why Magit is called the best Git frontend.
  • Monitoring SystemD services with Healthchecks.io — Monitoring SystemD services with Healthchecks.io. △6 / 1 comment (Lobsters). A practical guide to heartbeat monitoring for systemd services — “the service is alive” and “the service is doing its job” are two different things; heartbeat checks close that blind spot.
  • Software Engineering fundamentals matter more than ever — Software Engineering fundamentals matter more than ever. △8 / discussion ongoing (Lobsters). Reasserting engineering fundamentals amid the vibe-coding noise — standard Lobsters fare, but this one argues that AI makes fundamentals more valuable, not less.
  • Power Relations — Power Relations. △7 / discussion ongoing (Lobsters). Corbin’s distributed-systems essay — looking at control and trust in system design through the lens of power relations; an unusual angle, worth a read.

🎲 Light & Fun

  • The weekend is 100 years old — The weekend is 100 years old. 167 points / 100 comments (HN). The centenary of the two-day weekend — 100 comments argue over “the weekend was invented” and whether the four-day week will be the next century’s legacy.
  • Every Fucking Website — Every Fucking Website. △41 / 24 comments (Lobsters). A satire of AI-generated website sameness: soft palettes, rounded cards, status dots, 10px footnote text — the consensus among 24 comments is that AI websites’ surface polish already passes first-glance inspection. 💬 nemin names the uncanny part: these sites look professionally impeccable, and once you know they’re zero-effort products it feels awful; benj has it worse — even the em dash is AI-contaminated now, and he flinches at every dash he sees.
  • A True Telnet BBS on a Casio Calculator — A True Telnet BBS on a Casio Calculator. 80 points / 9 comments (HN). A real Telnet BBS running on a Casio calculator — 9 comments ask about keyboard input latency; a new toy for retro-tinkering hobbyists.
  • Anton Chekhov played at love most of his life — Anton Chekhov played at love most of his life. 54 points / 10 comments (HN). WUSTL’s study of Chekhov’s letters: he rehearsed “winning and losing” metaphors over and over in his love letters — a rare literary post for Monday-morning HN; all 10 comments marvel that great writers are also just people.
  • Archie G. Norcross’ Maine Forest Fire Maps (1918–22) — Archie G. Norcross’ Maine Forest Fire Maps (1918–22). 24 points / 3 comments (HN). Century-old hand-drawn fire maps collected by The Public Domain Review — a visual feast for cartography-history buffs; all 3 comments praise the color schemes.
  • Stop Turning every purchase into a tip request — Stop Turning every purchase into a tip request. 25 points / 19 comments (HN). A complaint post about tip inflation — 19 comments compare tipping cultures across countries and how annoying that default 20/25/30% option on the iPad is.

📐 Mathematics & Science

  • A SAT Attack on Tarski’s High School Algebra Problem — A SAT Attack on Tarski’s High School Algebra Problem. 79 points / 32 comments (HN). Attacking Tarski’s high-school algebra problem with SAT solvers — 32 comments verify solver scale and proof lengths; a new battle report from the automated-reasoning community.
  • A faster way to calculate the day-of-the-week — A faster way to calculate the day-of-the-week. △14 / 2 comments (Lobsters). Optimizing date-to-weekday conversion down to assembly level — 2 comments compare against Zeller’s congruence; the programmer’s eternal math toy.
  • How Is Compression Prediction? — How Is Compression Prediction?. △15 / 4 comments (Lobsters). An information-theory experiment using compression ratios to predict “compressibility” — 4 comments discuss the sources of prediction error.
  • Low-Tech Ceramic Water Filter — Low-Tech Ceramic Water Filter. 59 points / 13 comments (HN). A ceramic water filter fired from clay and sawdust — 13 comments discuss bacteria retention rates and DIY feasibility; a classic Low-Tech Lab project.

📝 Summary

Today’s mood is “settling day”: AI business news (Stripe acquisition, Nvidia guarantee cut, Anthropic IPO valuation) and browser ad blocking (Firefox iOS 489 points + uBlock Lobsters △192) each take half of HN’s top scores, and the community’s AI conversation shifts from “how capable are the models” to “where does the money come from, who pays”. Meanwhile gorhill personally debunking the uBO Lite rumors and simonw’s git-tracking of the Claude prompts show that transparency and evidence are becoming the community’s new currency in the AI era. Top 3 must-reads: the Claude system prompts in full (476 points — the official text of a model’s self-knowledge, including the injected record of the Fable/Mythos export-control suspension), the RISC-V rebuttal (296 points — a direct face-off with yesterday’s rant; the fight over the definition of “high-end” carries more information than the original post), and Cloudflare’s silent analytics injection (219 points — how a default-on tracking script burns user trust). Cross-platform resonance: the Stripe/OpenRouter acquisition and Nvidia’s shrinking financing guarantees landed the same day — the capital cycle around AI infrastructure is starting to tighten; Lobsters’ vibe-coding skepticism (Every Fucking Website, SWE fundamentals, C replacement) mirrors HN’s commercial carnival, and the two communities’ stance gap on the same wave remains stark.