Top 10 Hacker News posts, summarized
HN discussion
(286 points, 122 comments)
Cursor removed dollar-cost visibility from the Usage page and CSV export for self-serve plans (Individual and Teams), replacing them with token counts. A Cursor engineer explained the change was intentional: included plan usage was displayed as dollar amounts that exceeded the subscription cost, causing confusion. Enterprise plans retain dollar views. On-demand usage beyond included limits still shows costs on the Spending page and in CSV exports (marked "Included" for plan-covered usage). However, the API endpoint for usage events now returns zeroed cost fields retroactively, breaking per-request cost tracking for model efficiency analysis and budgeting. A second Cursor employee acknowledged a bug in the CSV export was fixed but confirmed the dollar graph removal was deliberate.
HN commenters overwhelmingly criticize the change as user-hostile and a transparency regression. Many report switching to alternatives like Claude Code, Codex, or plain editors, citing better agentic workflows and trust. A detailed token-efficiency comparison shows order-of-magnitude differences across harnesses for the same tasks, underscoring why per-request cost data matters. Commenters accuse Cursor of "enshittification," hiding costs to obscure value perception, and breaking trust—especially after a recent enterprise renewal. Some note the change coincides with Cursor's acquisition by an Elon Musk-affiliated entity, fueling skepticism about long-term intentions.
HN discussion
(235 points, 159 comments)
A GitHub issue reports that ripgrep 15.2.0 built for `x86_64-unknown-linux-musl` occasionally crashes with SIGSEGV during very-large directory searches. The crash occurs in musl's `mallocng` allocator during a `calloc` call from `opendir`, triggered by an integrity assertion failure on heap metadata. The issue is reproducible by generating a ~20 GiB tree of 1.8 million files and running ripgrep in a loop searching for a non-existent string; on a 24-core system with the tree cached in RAM, the segfault typically appears within a minute. The backtrace shows the crash originates in the `ignore` crate's parallel directory walker (`ignore::walk::Worker`), which spawns multiple threads that concurrently call `opendir`/`readdir`. The reporter provided a core dump and a debug-symbol build for analysis. The expected behavior is no segfault.
Commenters identified the root cause as a kernel bug (linked in the analysis repository) rather than a ripgrep or musl defect, with one noting the kernel patch discussion explicitly references the ripgrep report and characterizes the associated analysis as "AI-generated." Several discussants questioned why the bug manifests only with musl and not glibc, with speculation around musl's thread stack size and its `mallocng` allocator's poor performance under high thread contention. One commenter reported switching to `mimalloc` improved multithreaded allocation performance by 20x. Another criticized the workload pattern—running ripgrep at scale on large cluster filesystems—as fundamentally mismatched to metadata-heavy filesystem architectures, warning it can destabilize shared storage systems.
HN discussion
(246 points, 133 comments)
Canada has signed the UN Convention against Cybercrime, a treaty originally proposed by Russia in 2017 to displace the Budapest Convention. After opposing the UN resolution in 2019 and actively negotiating for human rights safeguards, Canada declined to sign at the October 2024 Hanoi ceremony alongside the U.S., Japan, and several European nations. Nine months later, the government signed without public consultation or explanation. The treaty extends beyond cybercrime: its procedural powers apply to electronic evidence for any criminal offence, and its cooperation obligations cover any "serious crime" punishable by four or more years' imprisonment—definitions that in some signatory states include political dissent, journalism, and same-sex relationships. Critics including the EFF, Human Rights Watch, and nearly two dozen Canadian organizations warn the convention functions as a global surveillance pact, mandating real-time interception powers while omitting judicial authorization requirements, permitting gag orders, and lacking a political offence exception. The treaty also risks enabling transnational repression against diaspora communities and criminalizing security research. Canada already possesses the Budapest Convention and bilateral treaties for cooperation with preferred partners, meaning the convention's marginal value lies primarily with high-risk authoritarian states.
Commenters noted that Australia, the EU, and the UK also signed, but emphasized that signature alone creates no binding obligations without ratification. A dominant theme was the treaty's role in advancing an "identity-managed internet": several linked it to age verification mandates, KYC expansion, and the broader trajectory of state surveillance, arguing that effective cross-border prosecution requires verified digital identities. Others expressed cynicism about democratic governments' motives, viewing the treaty as a vehicle for domestic data collection rather than genuine crime fighting. The absence of U.S. participation was highlighted, alongside Qatar's reservation exempting child protection articles. Michael Geist's longstanding coverage of privacy issues was praised, while some dismissed the convention as performative given the perceived impotence of states against cybercrime. A recurring frustration was the lack of proactive digital rights legislation to counter the steady erosion of online anonymity and due process.
HN discussion
(288 points, 89 comments)
The article argues that Google has systematically undermined RSS adoption through a pattern of "Embrace, Extend, and Extinguish." It documents multiple instances where Google integrated RSS support into products—Chrome's RSS button, FeedBurner, Google Reader, Google Alerts, a Chrome extension, and Google News—only to later remove or degrade that support without clear justification. The 2013 shutdown of Google Reader, despite a loyal user base, is presented as the pivotal moment that caused many users to abandon RSS entirely. The article also notes Google's 2021 announcement of renewed Chrome RSS support that never materialized, concluding that Google's behavior has damaged user confidence in RSS and threatens the openness of the web.
Commenters express strong resentment toward Google, with many citing Google Reader's shutdown as the moment the "internet as I knew it" ended. Skepticism surrounds Google's claimed "declining usage" justification, particularly since Google+ was being pushed simultaneously. While some argue RSS was never mainstream due to monetization challenges and technical complexity, others note RSS remains technically alive through WordPress, tools like NetNewsWire and RSSHub, and YouTube's channel feeds. A recurring theme is the distribution problem: the "hobbyist web" has collapsed because Google search no longer surfaces independent blogs, and algorithmic platforms (TikTok, Reddit) capture attention. Several commenters share technical workarounds, while others debate whether Facebook or Google bears more responsibility for RSS's decline.
HN discussion
(211 points, 103 comments)
RamenHaus is a visual showcase website featuring photography of ramen dishes from around the world. The displayed entry highlights a Shoyu ramen from Tombo Ramen in Tokyo, Japan, photographed in 2019. The site presents individual ramen bowls with minimal context—listing the style, shop name, location, and year—organized in an index format with navigation between entries. The presentation emphasizes high-quality food photography as the primary content.
Commenters largely praised the photography quality, with several noting the images were "drool inducing" and "spectacular." Discussion branched into personal ramen recommendations across London, Austin, Bangkok, Osaka, and Tokyo, with users citing specific shops like Kanada-ya, Tatsu-ya, Bankara, and Ittenbari. Some debated ramen's value proposition in Western cities versus Japan, while others shared nostalgic tech anecdotes like the #ramendan Twitter challenge. Technical feedback noted the rotating image animation appeared as a plastic replica on certain browsers due to reduced-motion settings, and one user criticized a Rome listing as unappetizing. Several expressed interest in contributing photos from their own cities.
HN discussion
(243 points, 66 comments)
Unable to fetch article: No content extracted (possible paywall or JS-heavy site)
The discussion is overwhelmingly skeptical of Flint's value proposition, with commenters questioning the need for a new JSON-based visualization specification "for the AI era." The core argument is that LLMs already excel at generating code for mature, well-documented libraries like Plotly, ECharts, Vega-Lite, and ggplot2, making an intermediate abstraction layer redundant. Critics argue that teaching an LLM a new, proprietary DSL via system prompts is less token-efficient and more error-prone than leveraging the model's existing training on standard libraries, especially given LLMs' documented struggles with strict JSON syntax. Several users note that Flint’s multi-backend switching capability solves a problem that rarely exists in practice, as developers typically standardize on a single library.
A minority of commenters acknowledge a niche use case for simplified, constrained schemas to improve reliability for basic charting tasks, but most view the project as "reinventing the wheel" without clear benchmarks demonstrating superiority over Vega-Lite or direct code generation. The sentiment is summarized by the suggestion that effort would be better spent contributing to existing ecosystems rather than fragmenting the landscape with a new syntax that lacks tooling (linters, LSP) and appears to offer marginal benefits over established grammars of graphics.
HN discussion
(177 points, 109 comments)
The author recounts the trajectory of "Jim," a charismatic but unfocused individual who leveraged personal connections and Silicon Valley's network effects to ascend from a coding bootcamp student and bartender to a Y Combinator-backed founder and Head of Engineering within three years. Despite early promise, Jim's career was marked by financial recklessness, repeated startup failures, and a descent into drug-fueled excess, culminating in a nervous breakdown, the loss of his fiancée, and financial ruin. The author uses Jim's story to illustrate a broader "meat grinder" dynamic in Silicon Valley: a pipeline that rewards performative founder personas and social capital over substance, chews through thousands of similar individuals, and celebrates the few survivors while the majority are discarded. The author contrasts Jim's chaotic, parabolic arc with their own steady, linear progression, concluding that consistent, grounded effort ultimately outperforms the high-variance spectacle the ecosystem incentivizes.
Commenters expressed skepticism about the plausibility of Jim's rapid rise to Head of Engineering without formal credentials, while others validated the depiction of a subset of Silicon Valley culture driven by performative entrepreneurship, easy VC money, and fraternity-like social dynamics. Several highlighted the normalization of drug use and "founder parties" as an underacknowledged problem, and debated whether lifestyle markers like home brewing or pets genuinely indicate irresponsibility. A recurring theme was the distinction between founders motivated by building versus those performing the identity of a founder, with comparisons drawn to figures like Elizabeth Holmes and Adam Neumann. Critics also pointed to VC incentive structures that fund charismatic but inexperienced operators, and noted the survivorship bias in public narratives that glorify the spectacle while obscuring the quiet, competent companies doing sustained work.
HN discussion
(180 points, 80 comments)
NetBSD 11.0 has been released after a delayed schedule, with the project opting for transparency over further delays regarding three open security issues. The release includes installation images split by media type (CD-ROM under 700MB, full DVD images, and USB .img files), with pre-configured U-Boot images for ARM devices. The acknowledged security issues affect hdaudio (local privilege escalation with a workaround), ipfilter (remote null pointer dereference, not enabled by default), and pf (use-after-free in fragment reassembly, deprecated and not enabled by default). All fixes will be committed to the stable branch for the planned 11.1 release within two months.
Commenters discussed NetBSD's viability as a daily driver and questioned the continued CD-ROM image support given USB ubiquity. Several users praised the project's transparency about open security issues rather than delaying the release. A user reported a smooth upgrade using sysupgrade(8) with cgd(4) encryption. Notable technical highlights from the release announcement include npf(7) firewall improvements with layer 2 and user/group filtering, and a new MICROVM kernel for x86 that boots in approximately 10ms. Questions were raised about Wine compatibility for running Windows SDR software, and broader curiosity about the current development momentum and user base across the BSD variants.
HN discussion
(168 points, 79 comments)
The article promotes a volume of "The Art of 64-Bit Assembly" focused on advanced MASM programming under Windows. It positions the book as bridging the gap between high-level language concepts (C++, Python, Rust) and their low-level implementation, covering vtables and method dispatch, Windows Structured Exception Handling (SEH), thunks, closures, iterators, coroutines, generators, fibers, concurrency primitives, Unicode string handling, and domain-specific macro languages. Each topic is rebuilt from scratch in assembly with explicit instruction-level decisions. The book targets readers who already know assembly and want to understand the "hard parts" without relying on runtime abstractions or plausible but incomplete explanations.
Commenters debate the relevance of hand-written assembly in the LLM era, with some arguing LLMs can handle simple assembly tasks while others note the book's specific Windows/MASM focus (vtables, SEH) represents undocumented tribal knowledge LLMs currently lack. Several criticize the marketing copy as AI-generated and the title as overly narrow (x64, Windows, MASM only). Technical discussion includes MASM vs. NASM/GAS macro capabilities, vtable ABI questions, and whether writing a custom compiler is preferable to raw assembly for performance work. A few note the author's multi-decade history with the book series, while others request Linux/PowerISA equivalents. Meta-commentary observes the thread focuses on tooling preferences and AI speculation rather than the book's technical content.
HN discussion
(97 points, 34 comments)
A soundness bug in the Lean 4 kernel (issue #14576) was discovered and fixed in late July 2024. On July 25, Ramana Kumar published an AI-assisted, `sorry`-free "disproof" of the Collatz conjecture that exploited a flaw in the kernel's handling of nested inductive types with phantom parameters. On July 28, Kiran Gopinathan reduced the exploit to a minimal proof of `False` and reported the issue; a fix was merged within an hour. The bug allowed ill-typed arguments to escape type checking when the kernel eliminated a nested occurrence under an inductive type whose parameters were phantom (absent from constructor fields). Critically, the bug was only reachable via metaprogramming by sending declarations directly to the kernel; the frontend elaborator caught the ill-typed terms, so the flaw is an implementation bug, not a hole in Lean's meta-theory. The independent kernel `nanoda` initially failed to detect the exploit because it had a separate, unrelated bug (missing a check on projection-node type names) that was fixed a week earlier. The Lean FRO has since added regression tests, hardened kernel invariants, improved parameter checking, and employed an AI cybersecurity agent that found several additional metaprogramming-reachable bugs, all caught by `nanoda`. Verification efforts like `lean4lean` would have caught the bug when completing the inductive-types consistency proof. The post argues against restricting metaprogramming, emphasizing that the kernel must remain the sole trusted component.
Commenters offered varied perspectives: some viewed the incident as an inevitable consequence of AI stress-testing formal systems ("If AI is water, Lean is the pipe..."), while others warned against trusting LLM-generated proofs even when they pass verification. A recurring theme was that soundness bugs are not unique to Lean—simpler type checkers like Rust's also have them—and that verified results provide an extraordinarily strong but not absolute guarantee, with the key strengths being a minimized trusted code base and rapid, serious response to issues. Skeptics dismissed Lean as overhyped and suggested alternatives like Coq, Isabelle, or Metamath for AI-driven formalization. Technical questions arose about whether a Collatz disproof must be constructive (a counterexample) and whether any kernel bug has ever allowed proving a new true statement without also enabling a direct proof of `False`. One commenter proposed a bounty on proving `False` as a trust-enhancing measure, and another invoked Knuth's famous caution about proved-but-untested code.
Generated with hn-summaries