Top of Hacker News

In today's GitHub Trending: OpenBB, a self-hosted research terminal aimed at the $2,000-a-month Bloomberg seat; mikeroyal/Self-Hosting-Guide, an index of run-it-yourself software; Meshery, a manager for the Kubernetes you already operate; and freeCodeCamp's 448k-star learn-to-code curriculum.

The pitch is self-reliance. The replies read as an audit: a Tell HN thread documents OpenBB emailing people it scraped from its own stargazers, and the loudest voices under the self-hosting guide call it a list of tools, not the howto it promises.

The CPX42 cloud tier jumped from €25.49 to €69.49 — a 2.7x increase that landed the morning it took effect. Existing machines are grandfathered; rescale one and you pay the new rate.

The thread's consensus isn't outrage at Hetzner so much as recognition: a cloud-infrastructure founder and several engineers traced it straight to the global RAM and storage shortage that's been squeezing everyone building hardware this year.

Running npm install would have been enough. The package's prepare script auto-runs a file that, buried under walls of commented-out tests, fetches and executes whatever a remote server sends back.

He never cloned it locally. He spun up a throwaway VPS, pointed a read-only AI agent at the tree, and it flagged app/test/index.js almost immediately. The thread's verdict on who's behind it was fast and grim: DPRK.

The numbers are real: 200 million endpoints created on the public relays in the last 30 days, with about 95% of traffic flowing directly device-to-device instead of through a cloud hop.

Two of Iroh's own developers showed up to field questions, including the hard one a 'really pretty old' networking veteran asked: how does a flat key-based address space scale without every node tracking every other?

You write your app once against Apple's Foundation Models protocol, then choose per session: the small local model, or Claude in the cloud. Requests go straight from the app to Anthropic; Apple isn't in the path.

The thread caught the sleight of hand fast. One reader who benchmarks open models noted Apple's framework caps local models near 3B parameters, so calling a frontier cloud model through an 'on-device' API is, charitably, hybrid.

Roku's whole appeal was being service-agnostic hardware. Now a content company owns the platform, the home screen, and the ad inventory, and longtime users are already describing the creepy cross-device popups that hinted at where this goes.

A throwaway detail that landed: Roku's Walmart security cameras are rebranded Wyze units. The advice in the thread was to migrate off both the cameras and the boxes before the integration tightens.

Fin's pitch is an AI agent that resolves support tickets across chat, email, phone, WhatsApp and Slack, powered by a proprietary model called Apex that it claims beats frontier models on resolution rate.

Two people who built in the space called the category a grift: the bots don't solve the reason a human contacts support in the first place, and customers would rather wait for a person than loop with a wrong answer.

Thompson traces a pattern: Mythos withheld as too dangerous in April, Fable released with guardrails in June and promptly jailbroken, then both gated under ITAR export controls. Safety, he argues, conveniently aligns with talent, mission, and business.

The sharpest rebuttal in the thread: you can't be on your way to 'power over everything' and get distilled into free Chinese models within months. Pick one. The bottleneck is compute and data, not the model.

The essay's exhibit A is a Founders Fund 'mafia' video where Sam Altman, Palmer Luckey and others play deception games — reality TV as a laundering technology that turns someone you'd keep at arm's length into a recurring guest.

HN's pushback reframes the whole complaint: you're describing business and management people who learned to perform, not engineers who lost their way. The split is whether 'nerd' was ever a job or always a temperament.

The HackerOne form and security email go dark July 1 to August 3, and the 8.22.0 release slips two weeks to September 2. Daniel Stenberg's reasoning is plain: the team is burnt out and needs rest.

The thread half-celebrates, half-worries about one loophole: paid support clients can still pressure for a fix if someone discloses publicly during the freeze — and curl, for all its ubiquity, has no second on-call human.

It's an n=1 case report from a three-person group in São Paulo, in a journal some researchers list as predatory. The patient, an octogenarian, got 5 grams then 3 grams; improvements were real but transient.

A McGill/MILA postdoc supplied the context the headline omits: case studies are among the easiest journal outputs to fabricate, and this publisher once ran an AI-generated anatomical figure of a rat that became a scientific punchline.

The compound has already passed safety evaluations for other diseases, which the release frames as a fast path to human trials. The first reply set the tone: 'Great news! If you are a mouse.'

The real reversal is the science underneath. A developer quoted chemist Derek Lowe's verdict that amyloid-directed therapies 'truly, truly do not appear to be the answer' — and a son caring for a parent with the disease explained why amyloid removal hasn't stopped decline.

The linked page is raw data labeled 'Industrial Production: Battery,' indexed to 2017 — no value, no chemistry breakdown. As one reader noted, breaking your own record every year is what a rising line does.

The context that recolored everything: 2025 cell-production capacity at roughly USA 70 GWh, Europe 252, China 1,755. The US line is genuinely up and, in global terms, still tiny — and the materials are mined overseas anyway.

The essay asks readers to examine their assumptions, then, critics say, makes whoppers of its own: that governments do nothing, that elites stay perfectly unified, that everyone else stays quiet.

The sharpest rebuttal modeled it as two countries — an AI island and everywhere else — and pointed out the island still needs to trade for things it can't produce, which keeps the rest of the economy very much alive.

Power it on and it broadcasts an open Wi-Fi network with a captive portal that hands you the library — no internet required. Storage caps it at roughly six ebooks, which the author frames as a feature: each drop reflects whoever made it.

The pushback was semantic and sharp: the titles are Call of the Wild and Huck Finn, not anything actually illegal to own in the US. The build is the star; the framing got the side-eye.

A 20W emergency hand-crank charger feeds a Pi 5 through a custom capacitor board that buffers about 20 seconds of power. A 350M model hits ~49 tokens/sec; idle draw is 4W, and an answer-plus-speech burst spikes toward 5A.

You can feel the inference load in your wrist: when the model and text-to-speech run together, the crank physically stiffens. Half the thread wanted to buy one; the other half pictured humans on hamster wheels for the AI overlords.

The author's claim: if a Rust function isn't marked unsafe, it's impossible to use incorrectly in a memory sense, so a memory bug there is always a soundness hole. C and C++ shift that burden to the caller, so the same class of mistake never becomes a CVE.

The thread split hard. One camp: CVE counts are so useless they ignore anyone citing them. Another, channeling Stroustrup: most of these bugs are C-style C++, and modern C++ plus hardened libraries closes the gap.

An engineer's quick eval put it plainly: about 7x slower and 4x the cost of calling one frontier model directly, which lands Fusion in a 'use it only when you really need it' bucket.

The deeper objection came from someone who shipped this exact thing months ago: asking one model to judge another mostly measures how closely the answer resembles what the judge would have said. It's temperature, dressed up as a panel.

0.15 adds multiple bibliographies in one document, automatic MathML in HTML export, an experimental bundle export, variable fonts and spot colors. For people generating PDFs programmatically, it's been saving real money.

The one recurring wish: a double-click desktop editor so non-developers can use a .typ file without setting up VSCode and plugins. And LLMs still struggle to write Typst, which complicates the otherwise slick Pandoc-to-Typst book pipelines.

The author, who also wrote the popular gptel package, frames Emacs as having a discoverability problem he's chipping at 'one demo at a time.' The built-ins are genuinely useful and genuinely buried.

The top reply reframes the whole premise: the issue was never finding features, it's keeping a large config stable across package updates. The fix people actually want is a good out-of-the-box experience, not another blog post.

The book covers the 80286, EGA hardware, the engine architecture and the CGA port, with source on GitHub and a free high-res PDF. The author calls it a return to debugging C and assembly as a teenager.

The context one reader added is the real lesson: dedicated console hardware rendered sprites for free, so a 'more powerful' PC had to invent tricks to fake what a SNES did in silicon. Several others noted the site looks a lot like Fabien Sanglard's.

The closest anyone gets: GLM and Qwen run agentically on self-hosted boxes and feel near Haiku, but the thread's experienced voices agree none of them reach Opus 4.8 or GPT 5.5 for real work.

The most-upvoted reply isn't an answer, it's a complaint about the genre: 'I use Qwen and get great results' tells you nothing without the quantization, llama parameters, context window, GPU and VRAM behind it.