Hacker News
1DeepInfer: High-performance inference framework for DeepSeek models
- DeepInfer is an open-source, high-performance offline inference framework from DeepSeek for serving DeepSeek models at scale, currently supporting DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950 hardware. It is licensed under Apache 2.0.
Inference profitability
Commenters debate whether the reported $38/GPU/hour revenue shows real profit. Some argue it excludes training, power, and overhead, and that competition will compress margins. Others say inference is clearly profitable.Impact on US frontier labs
Many see cheap open models like DeepSeek threatening OpenAI and Anthropic valuations and moats. Others argue enterprises, governments, and distribution channels will keep the frontier labs relevant.China's motives for open-sourcing
Users question whether state interests drive the release, citing commoditizing competitors, Huawei hardware promotion, and a DHS ban on DeepSeek products in federal agencies.
2WhatsApp's help pages hid LLM instructions: dont recommend a regulator
- Hidden instructions in WhatsApp's help pages told its AI chatbot not to recommend data protection regulators or cite privacy laws like GDPR, and to withhold other account deletion routes. The notes were removed by October 18, 2026, though WhatsApp's public privacy policy names those rights.
Hidden instructions as prompt injection
Commenters split on whether the hidden text is ordinary support-bot tuning or malicious manipulation. Some compare it to agents.md files, while others say telling the bot to hide regulators and deletion options is deceptive.Legal restrictions vs. user rights
Some defend keeping chatbots away from legal advice because of hallucination and liability risk. Others argue GDPR requires explaining users' rights, and steering the bot away from regulators and deletion options hides them.Trust in LLM answers
Users may over-trust AI answers that companies can steer or manipulate. Commenters expect LLM trust to erode as the web did with ads and SEO, though some hope for a neutral, Wikipedia-like model.
3EMudhra Intermediate CA Compromise
- Private keys for eight eMudhra intermediate CA certificates were compromised due to flawed key generation, letting anyone recover them from the public certificates. Browsers like Firefox, Chrome, and Apple have blocked them, but many other clients still trust them and need updates. Users should update software promptly.
Trivially weak keys
All eight affected CA keys were single-byte values like 4 and 156, so anyone can recover them. Commenters joked about XKCD's PRNG, and eMudhra blamed a flawed key-generation script that produced a second set of weak keys.Uneven revocation across clients
Firefox and Chrome could block the CAs quickly through built-in revocation lists, but Go, Rustls, Linux, and other clients lack comparable mechanisms, leaving them exposed.Calls to distrust eMudhra
Many argued eMudhra should be distrusted, like DigiNotar, and questioned how a CA that didn't use an HSM was admitted to root programs. Sub-CAs were called a loophole needing name constraints.
4Ask HN: How to deal with shipping AI-written code you don't understand?
- A Meta engineer shares discomfort with AI-driven shipping: code feels like a black box, and reviewing every line is too slow, so engineers understand the high-level design but not every detail. They ask whether others feel the same and how they cope.
Design-first workflow with LLMs
Commenters describe keeping control by designing abstractions and interfaces up front (UML, Python sketches, plans), then having the LLM implement them, since LLMs tend to produce poor abstractions and unmaintainable structure.Testing as the new verification
Many argue heavy automated testing, property-based tests, and CI gates can replace line-by-line review, though others warn tests miss design flaws, security issues, and unintended behavior.Losing understanding of code
Critics say writing code keeps engineers grounded in the codebase, so reviewing or delegating everything erodes tacit knowledge and accountability, while defenders say this mirrors how principal engineers and managers already work.
5Starcraft: Ashfall
- A Terran Expeditionary Force strategy game set on the Ashen Long Bridge, where players lead marines, firebats, and siege tanks to hold the line against a Zerg swarm with reinforcements. The battle won't start because the game requires portrait orientation.
Addictive idle gameplay
Players find the game very addictive despite minimal input, comparing it to Cookie Clicker. Critics call it mindless, while others defend repetitive, low-intensity play as emotional regulation.Blizzard IP and copyright
Commenters note the game uses StarCraft's IP, original art, and voice lines, and question whether it's fair use. Some argue copyright should change for AI-era derivative works, while others call it a trademark issue.Mobile ad games comparison
Commenters say the real mobile ad games bear little resemblance to their videos and are timer-based, pay-to-win builder games. This fan-made version is praised for delivering the arcade experience the ads promised.
6DeepSeek banned from U.S. federal civilian agencies under new DHS order
Security risks of Chinese open-weight models
Commenters debate whether self-hosted open weights are risky. Some warn of hidden triggers or backdoors that can't be detected by inspecting weights, while others say local hosting limits exposure and US models carry similar risks.Regulatory capture and protectionism
Many see the ban as favoring OpenAI and Anthropic, possibly due to their lobbying, and worry it will reduce open-weight releases and price competition. Others view it as legitimate national security policy.Vague scope of the order
Users question whether "developed using DeepSeek" covers open-source libraries or AI-assisted code. They argue the ambiguous wording could make compliance impossible and enforcement arbitrary.
7Voxtral 3
- Voxtral 3 is a new family of speech models and a voice agent SDK for real-time, multi-step conversations in 50+ languages. It offers accurate transcription, expressive speech synthesis, and tool integration for workflow automation, with EU-hosted or self-deployed options that keep data under the organization's control.
Open weights and local STT/TTS
Commenters lament that the model is closed and note local speech models lag behind. They suggest alternatives like Qwen ASR, Pocket-TTS, Kokoro, Parakeet, Speaches, and FUTO's models, and question Mistral's open releases.Voice agent social engineering risks
Users predict voice agents wired to company intranets will be manipulated by callers with fake authority or malicious URLs. Some argue AI agents are easier to exploit repeatedly than humans, and suggest sandboxing their APIs.Subtitle timestamp correction tools
Users seek local tools to fix existing SRT timing. Recommendations include alass, subsync, srt-resync, and Bazarr, plus LLM-built forced alignment with CTC models, described as state of the art.
8SimCity, but with Permits
- Permit Pending is a browser-based city simulation where you play as an elected mayor who must get projects through zoning boards, hearings, appeals, lawsuits, and elections. Every improvement sparks opposition, and the economy, residents, and local paper all respond to your choices. There is no win screen; the goal is to build the city you want while staying in office.
Democracy causing housing gridlock
Commenters argue that a century of rules, zoning, and community input to keep everyone happy makes building nearly impossible. Expats self-organizing in Costa Rica show how slow consensus can be, though many note that the alternatives are worse.Veto points and urban renewal
Debate over whether today's permitting stems from Robert Moses-era abuses or now mainly protects wealthy incumbents. Critics say community process blocks needed housing and transit, while others defend the rules as safeguards.Overwhelming game design
Players found the game overwhelming, with many buttons and constant petitions from the start. Some see the chaos as deliberate, mirroring a planner's paperwork, while others say the UI and pacing need work, especially on mobile.
9Apple broke an essential presentation feature of iPads
- The author, a teacher who projects iPad handouts and notes via AirPlay, complains that iPadOS 27 broke multi-window presenting: tapping any other app flips the external display to show the whole iPad screen, exposing private notes and poll results to students. They urge Apple to keep presenter-mode apps in control of the display and encourage others to file feedback.
Presentation mode and AirPlay control
Users say the change, where the most recently tapped app takes over the external display, breaks switching from slides to a live demo. Proposals include letting the last explicitly cast app keep control or offering a share-screen choice.Declining software quality at Apple
Users say recent releases, including macOS and iPadOS 27, are buggy and under-tested. Blame falls on the yearly release cycle, cut QA, feature creep, and reliance on automated tests. Many plan to stay on current versions.iPad window management and computing limits
Commenters find iPadOS window management clumsy, especially without a mouse, and say iPads lack the flexibility of a real computer. Others praise iPads for presentations and teaching.
10OpenBSD 8.0
- OpenBSD 8.0, the 61st release, came out October 20, 2026. Highlights include a larger amd64 kernel address space, hibernate support on arm64, RISC-V vector and interrupt controller drivers, Intel Arc graphics support, expanded VMM/VMD guest features, many wireless and hardware driver fixes, and the retirement of loongson support.
Unix-like feel and tooling
Users praise OpenBSD as the most Unix-like system, citing its minimal default "ps" output, the pf firewall, excellent man pages, and pledge/unveil security. Several have run it for decades on old laptops.Painless upgrade process
Commenters report that sysupgrade, sysmerge, and pkg_add -u make major release upgrades smooth, though some note that Linux and other OSes are just as easy. One user needed a monitor attached for the upgrade to work.Satirical release artwork
The 8.0 release image sparked debate. Some read it as AI satire and found it funny, others disliked its self-destructive tone, and people compared it to Simpsons and Blinkenlights references.