Top 10 Hacker News posts, summarized
HN discussion
(485 points, 207 comments)
The article content could not be loaded due to a JavaScript error message indicating browser extension, network, or settings issues. The title indicates Firefox for iOS has added a native adblocker, but the article body itself is inaccessible.
Commenters note Firefox's native adblocker has notable exceptions: it does not block ads on search engine results pages (Google, Bing, DuckDuckGo) or sponsored content on Firefox's own Home/New Tab pages. Firefox Focus already offered system-wide adblocking via iOS's Content Blocker API since the late 2010s. Users criticize the continued lack of extension support on iOS (which Orion provides), and question whether the new blocker supports cosmetic filtering or only DNS-level blocking. Many consider uBlock Origin Lite for Safari the best iOS adblocker, while others prefer paid options like Wipr 2. Additional complaints include Firefox iOS's autocomplete behavior and the absence of Gecko engine (Apple requires WebKit). Some users report the feature not yet available to them.
HN discussion
(472 points, 208 comments)
Anthropic publishes the system prompts used for Claude's web interface (claude.ai) and mobile apps, which provide the model with current information like the date and encourage specific behaviors such as using Markdown for code snippets. These prompts are periodically updated to improve responses, though updates do not apply to the Claude API. The documentation shows versioned entries for each model, with changes bolded between versions; starting with the Claude 4.6 generation, each model ID represents a single fixed snapshot with one entry.
Commenters note the system prompts have grown from roughly 300 words in early versions to over 3,000 words in recent ones, with Simon Willison maintaining a git history tracking changes. The Opus 5 prompt reveals a safeguards routing mechanism where Fable 5 queries may be redirected to Opus 5 due to export controls, including a verbatim blog quote for the model to reference. Several users criticize the omission of tool definitions and Claude Code system prompts, which are not published but can be extracted via proxy. Other observations include the absence of safety guardrails in the earliest prompts, hardcoded knowledge like the 2024 U.S. election results, and skepticism about claimed prompt size reductions. Users also question why prompts aren't baked into the model to save token costs, why harnesses don't allow customizable system prompts, and whether the "brief and concise" instruction matches actual behavior.
HN discussion
(293 points, 153 comments)
Embedded engineer Armstrong Subero responds to Dmitry Grinberg's critique of RISC-V by arguing that Grinberg overlooks the architecture's most consequential advantage: accessibility for engineers outside the US and Europe. Writing from Trinidad and Tobago, Subero details how shipping costs, import restrictions, and vendor discrimination make a ten-cent RISC-V chip (CH32V003) fundamentally different from a one-dollar alternative—it determines whether thirty students each get hardware or watch a single demo. Subero demonstrates RISC-V's scalability across three devices on his desk: the 10¢ CH32V003 (RV32EC, no multiplier/divider), the dual-core CH32H417 (400 MHz, USB 3.2, Ethernet, runs a web browser and facial recognition), and the Baochip (VexRISC-V with MMU, runs seL4, Xous, and Linux). All share the same base ISA, register model, calling convention, and toolchain, enabling skill transfer across the entire stack for under $100 total. Subero contrasts this with ARM's licensing-enforced product segmentation—Cortex-M (no MMU) vs. Cortex-A (MMU, Linux)—which requires entirely different core families, royalty negotiations, and debug probes costing hundreds of dollars. He argues that RISC-V's "fragmentation" via optional extensions is precisely what enables scalability: small cores don't carry large cores' baggage, and privilege features are checkboxes in the privileged spec, not license gates. RISC-V wins the cheap microcontroller space not through ISA elegance but through price, availability, and an open ecosystem that puts embedded development within reach of the global majority.
Commenters are split on Subero's shipping-cost argument: several note that Nigeria and Bangladesh have robust electronics supply chains with low last-mile costs, and one observes that when shipping dominates at $60–200, the difference between a 10¢ and $1 chip becomes a rounding error. Others argue Subero talks past Grinberg's core critique—that RISC-V's fragmentation and optional extensions hinder binary distribution and high-end performance against ARM64, not its embedded suitability. A common thread acknowledges both perspectives as valid: Grinberg highlights architectural compromises ("could be better"), while Subero demonstrates practical empowerment ("already much better"). The licensing freedom to add an MMU to a soft core (as with Baochip) without royalty negotiations is widely cited as RISC-V's genuine structural advantage. One commenter dismisses the article as "AI slop," while another provides an archive link (archive.is/9FaRC) for the rate-limited original. Several express enthusiasm for Subero's Rovari education platform and the democratizing potential of a $100 full-stack RISC-V development path.
HN discussion
(211 points, 128 comments)
The article argues that frontier AI labs are deliberately trading world knowledge for reasoning capability, resulting in models that excel at math and code benchmarks but fail at factual recall. Reasoning compresses efficiently into small parameter counts (procedures like problem decomposition and self-checking), while facts require ~2 bits per parameter and rot quickly (APIs change, prices shift). Current leaders like GLM-5.2 (40B active params) score 99.2% on AIME 2026, while Gemini 2.5 Pro leads SimpleQA factual recall at only 53%, and small models hallucinate 80-82% of the time. The proposed architecture moves facts out of model weights into a runtime "harness" (retrieval, tools, web search), enabling frontier-quality reasoning on consumer GPUs (20-40B params at 4-bit). This makes hallucination addressable: cited sources are checkable and fixable, unlike errors baked into weights. The model becomes a stable reasoning engine handed current world state at inference, decoupling the expensive training artifact from daily-changing facts.
Commenters raised several substantive critiques. Multiple users noted the article's benchmarks are outdated (SimpleQA last updated Sept 2025; Gemini 2.5 Pro is 16 months old) and questioned whether the reasoning/facts separation holds for domains like history where reasoning requires factual grounding. The "harness" approach faces practical hurdles: Google search API costs $2.50/1K queries, current retrieval is limited to chunked embedding search, and moving facts to context windows doesn't guarantee accurate output (RAG can still misread or stitch sources incorrectly). Several commenters disputed the hallucination claim, arguing hallucinations are runtime artifacts not "wrong facts in weights," and that internet sources contain errors users can't fix. Others questioned the consumer GPU assumption (most lack 24GB VRAM) and whether the coding/agent focus reflects broader AI use cases. A recurring theme was skepticism about the article's provenance, with multiple users claiming AI-generation detection tools flag it as 100% synthetic.
HN discussion
(205 points, 76 comments)
The article investigates the commercialized resale market for AI inference credits, where brokers purchase unused credits from startups (including Y Combinator Startup School allocations) and resell them at discounts through direct outreach, dedicated marketplaces (AI Credits, AICreditMart), and "bulk discount" routers (CheapCredits, Tokvana, Neokens). Brokers operate as proxies that route requests through pooled API keys rather than transferring keys directly, with one broker offering $100k/day in spend. The author found credits traded across Telegram channels, Reddit, and private founder groups, estimating tens of millions of dollars in credits available. Discounts of 40% or more are claimed via bulk pricing, though the author suspects alternative supply sources such as stolen keys, carded accounts, or automated trial signups. The market has sufficient liquidity for tokens to function as a pseudo-currency, inviting abuse that will likely prompt provider crackdowns.
Commenters debate the supply sources: some view reselling unused startup credits as a ToS violation but genuine arbitrage, while others argue 98% discounts indicate stolen API keys, credit card fraud, or automated trial farming, with speculation that buyers may receive substituted models (e.g., DeepSeek responses sold as Anthropic). Security risks are highlighted—proxies can intercept and manipulate tool calls, enabling code execution or data exfiltration. The Chinese market is described as far larger. Several note the pattern mirrors historic grey markets for digital goods (airline miles, loyalty points), with organized operations including botnet rings and carding. Stripe is criticized for insufficient fraud prevention. A macroeconomic perspective frames the trade as inevitable given tokens' commodity nature, digital transferability, and global jurisdictional arbitrage. Finally, the underlying economics are questioned: heavy subsidization (e.g., $200 subscriptions generating $20k+ in API costs) is deemed unsustainable, suggesting a coming correction.
HN discussion
(217 points, 55 comments)
The author discovered that Cloudflare silently injects a JavaScript analytics beacon (from `static.cloudflareinsights.com/beacon.min.js`) into their HTML-only, JavaScript-free website (textlog.cc) after switching nameservers to Cloudflare. The injected script includes a token and version identifier (`"version":"2024.11.0"`) and appears to be Cloudflare's Real User Monitoring (RUM) / Web Analytics functionality. The author characterizes this as "hostile code" injection into a site Cloudflare does not host, raising questions about transparency, consent, and potential legal implications under the Computer Fraud and Abuse Act. The post references a Cloudflare blog post titled "The RUM Diaries: Enabling Web Analytics by Default," suggesting this may be a recent default behavior change.
Commenters clarified that this injection only occurs when Cloudflare is used as a reverse proxy (orange-clouded/proxied records), not when used for DNS-only (grey-clouded) resolution. Several users confirmed the behavior is tied to the "Web Analytics" feature, which may now be enabled by default for new domains or after certain changes. Multiple commenters reported difficulty locating the setting to disable the injection, with one noting they had to enable analytics first to access the off switch. The discussion highlighted that as a man-in-the-middle proxy terminating TLS, Cloudflare can modify HTML responses—including injecting scripts—and thus also has visibility into all cleartext traffic. Content Security Policy (CSP) with `script-src 'self'` was cited as a client-side mitigation to block the injected beacon. Some users expressed frustration over Cloudflare changing defaults without clear notification, burning goodwill, while others noted this is fundamental to how CDN/proxy services operate. Alternative providers and self-hosted solutions were suggested for those seeking to avoid such surprises.
HN discussion
(166 points, 98 comments)
The article traces the 100-year history of the two-day weekend, beginning with the Soviet Union's failed 1929-1940 experiment eliminating shared days off — which devastated morale and family life without boosting productivity. Henry Ford institutionalized the five-day week in 1926 to combat assembly-line exhaustion and create car-buying leisure time, while Boots the Chemist proved its commercial viability in the UK in 1933. The weekend became widespread post-WWII, evolving into a cultural coordination device anchored by rituals like Saturday football and Sunday church. Today, the weekend is fracturing along class lines: "time owners" (knowledge workers) increasingly treat Friday as an unofficial half-day while "time slaves" (shift, gig, and care workers) remain clock-bound. Smartphones blur work-life boundaries, four-day-week pilots show promising productivity gains but face political resistance, and social atomization weakens the weekend's communal function. The article argues the weekend endures precisely because it remains a shared, imperfect resistance to market time.
Commenters emphasize the week and weekend as recent social constructs rather than natural rhythms, citing historical research on how industrialization and Protestant culture cemented the seven-day cycle. A strong thread explores the growing bifurcation between workers with schedule autonomy and those without, with several asking practically how to escape "the industrial clock" given housing, family, and healthcare constraints. The four-day week generates both enthusiasm and skepticism, while one extensive comment frames labor concessions as historically won through collective struggle and now being eroded by manufactured hyper-individualism. Others note the Anglo-centric recency of many "traditions," propose radical calendar reforms, or warn that technological surveillance may render the weekend a temporary historical anomaly.
HN discussion
(141 points, 104 comments)
Unit 1 at the St. Lucie Nuclear Power Plant in Florida was manually shut down on August 13, 2026, at 9:47 a.m. EDT after three control rods dropped into the reactor core while operating at 100% power. The Nuclear Regulatory Commission classified the event as a non-emergency. The reactor trip was uncomplicated, with all systems responding normally; the plant was stabilized in Mode 3 (Hot Standby), and decay heat is being removed via steam discharge to the main condenser using turbine bypass valves and main feedwater. Unit 2 was unaffected. According to NextEra Energy, Unit 1 has since returned to 100% power after operators resolved the equipment issue.
Commenters discussed the technical nature of dropped control rods in pressurized water reactors, noting that rods are held by electromagnets and drop on power loss as a fail-safe; an unplanned drop forces a shutdown because the reactor becomes subcritical, but this is a practiced, generally non-hazardous event. A 2024 incident at the same plant was attributed to a procedural issue combined with electrical failure of the control rod grippers. Several users criticized the article for omitting the plant's location (Florida) and for repetitive phrasing. Others emphasized the need for relatable risk context to help the public distinguish between routine safety events and serious accidents, while some added humor referencing Chernobyl/Fukushima and the "3.6 roentgen" meme.
HN discussion
(88 points, 69 comments)
Unable to fetch article: HTTP 403
The discussion centers on skepticism regarding OpenRouter’s $7 billion valuation—up from $1.3 billion months prior—for what many view as a simple API proxy, with commenters noting the price exceeds the market caps of established companies like Lyft and Dolby. Critics question the technical necessity of a model router and worry about service degradation, citing poor support and the potential loss of free model access (e.g., DeepSeek). However, several commenters articulate a coherent strategic thesis for Stripe: the acquisition secures massive AI payment volume (OpenAI recently switched to Adyen) and positions Stripe as the financial rail for the emerging "token economy." By owning the routing layer, Stripe gains distribution leverage, data visibility into prompts and usage patterns, and the infrastructure to abstract variable LLM costs into manageable metered billing for future AI applications, effectively aiming to become the "AWS for tokens."
A secondary thread debates the competitive landscape, comparing OpenRouter’s traffic favorably against rivals like fal.ai and framing its moat as switching costs derived from embedded logging and cost-optimization tooling—likening its position to Slack versus AWS Bedrock’s Microsoft Teams. There is notable cynicism regarding antitrust scrutiny, with users observing the deal faces less resistance than a hypothetical Stripe-PayPal merger would have. Developers are already seeking alternatives or requesting reseller/OAuth capabilities to capture margin on the usage they drive, reflecting an expectation that the acquisition will prioritize Stripe’s platform economics over the neutral, developer-first ethos OpenRouter previously cultivated.
HN discussion
(84 points, 53 comments)
Buf has released a production-grade Language Server Protocol (LSP) server for Protobuf, providing modern IDE support including go-to-definition, code completion, reference finding, and semantic syntax highlighting for editors like VSCode (via the Buf extension) and Neovim (via the Buf CLI and nvim-lspconfig). The LSP is powered by Buf's Protobuf compiler frontend, protocompile, which is faster and more flexible than protoc and already used in parts of Google's codebase. Buf built a new query-driven frontend with a custom AST and intermediate representation to enable incremental compilation, precise diagnostics (such as correctly flagging duplicate `repeated` modifiers), and memory-efficient handling of large workspaces. Future plans include automatic import fixes, tighter `buf.yaml` integration, custom option completion, automatic field/enum number suggestions, and dedicated Protovalidate support with CEL syntax highlighting.
Commenters expressed skepticism about Protobuf's relevance in the LLM era, with some arguing that JSON-over-HTTP outperforms gRPC/Protobuf for Google Cloud services and criticizing Protobuf's tooling complexity, build-step overhead, and implementation quality. Others noted that Protobuf's compatibility constraints (e.g., discouraging field renaming/reordering) limit LSP utility for refactoring, though LLMs may mitigate migration challenges. A few highlighted Buf's from-scratch parser reimplementation (protocompile) as a technical necessity for error recovery but warned of potential implementation drift. An alternative schema language for game development was mentioned, and one commenter praised Buf for making Protobuf "barely usable" on legacy projects, while another found the announcement's "You're welcome" tone amusingly presumptuous.
Generated with hn-summaries