Top 10 Hacker News posts, summarized
HN discussion
(894 points, 543 comments)
A decade-long Backblaze customer reports that the service has silently stopped backing up OneDrive and Dropbox folders, as well as .git directories, without notifying users. This change, buried in release notes, shifts Backblaze from its original promise of backing up "all user data" to a policy of excluding files from "directly connected storage." The author argues this breaks the core trust of a backup service, as cloud storage files have different retention policies and are less secure, and that the lack of transparency is unacceptable. The user feels betrayed, noting that their 383GB OneDrive folder is no longer being backed up despite their continued payment.
HN users are highly critical of Backblaze's actions, viewing it as a significant breach of trust and a sign of "enshittification." Many commenters argue this is a common issue with consumer backup services that use vague excuses to avoid backing up certain data or to quietly renege on "unlimited" promises. A key concern is the lack of transparency; users find it unacceptable that such a major policy change was hidden in release notes and not clearly marked in the client's exclusion list. Some commenters suggest the move is a push towards their B2 cloud storage service, while others recommend switching to self-hosted solutions like Restic with providers like Backblaze B2 or Hetzner for more control and reliability.
HN discussion
(1035 points, 261 comments)
DaVinci Resolve's new Photo page integrates professional Hollywood-grade color grading tools into still photography, offering features like primary corrections, curves, Power Windows, and node-based workflows beyond traditional layer-based editing. It supports RAW files up to 32K resolution, includes AI tools (Magic Mask, Relight, Face Refinement), GPU acceleration for speed, and collaborative cloud workflows. Photographers access familiar adjustments (white balance, exposure) alongside advanced color controls, while colorists can apply their expertise to stills. The page also manages image libraries, tethered shooting, exports, and uses Resolve FX for cinematic effects, all preserving EXIF metadata.
HN users express strong excitement, seeing the Photo page as a viable competitor to Adobe Lightroom and Capture One, particularly for Linux users seeking alternatives to Adobe. Key reactions include praise for its value proposition (free tier vs. expensive alternatives), advanced color tools surpassing traditional photo editors, and potential to disrupt the market. Discussion highlights include: enthusiasm for resolving Adobe fatigue, appreciation for GPU acceleration and RAW handling, speculation about Blackmagic Design's business model (hardware subsidizing software), and requests for standalone apps. Some concerns exist about Linux support (CUDA dependency) and integration stability, while users note it addresses long-standing gaps in photo editing innovation compared to video workflows.
HN discussion
(808 points, 458 comments)
Google is expanding its spam policies to explicitly prohibit "back button hijacking," defining it as a deceptive practice that violates "malicious practices." This occurs when a site interferes with the browser's back button, preventing users from returning to the previous page as expected (e.g., redirecting to unwanted pages or blocking normal navigation). Google cites user frustration, broken expectations, and a rise in such behavior as reasons for this policy change, which will lead to manual spam actions or automated demotions in search results starting June 15, 2026. Site owners are urged to audit and remove any scripts, libraries, or ad platform code causing this hijacking.
Hacker News comments express broad frustration about the delay in addressing back button hijacking, naming Microsoft, Reddit, and Amazon as notable offenders. Skepticism about enforcement effectiveness is prominent, with users predicting a "cat and mouse game" where sites find workarounds like intermediate landing pages. Many demand browser-side fixes instead of relying solely on penalties, criticizing Google for hypocrisy given other poor user experiences like cookie walls and ad clutter. Third-party advertising platforms are seen as a common source of the problem, with calls to prevent them from manipulating browser history. There's also significant frustration related to other manipulative practices like scroll hijacking and paywalls.
HN discussion
(473 points, 406 comments)
The article introduces `jj`, the command-line interface for Jujutsu, a distributed version control system (DVCS) positioned as both simpler and more powerful than Git. It claims Jujutsu synthesizes the best features of Git and Mercurial into a cleaner, more cohesive toolset, offering advanced capabilities that are difficult with Git. A key advantage is its Git-compatible backend, allowing users to try it without requiring collaborators to switch and without losing their history if they revert to Git.
The Hacker News discussion is polarized. While some users express excitement about trying `jj`, several critical points emerge. A common complaint is that `jj` automatically stages file edits, forcing users to adopt new workflows like creating empty commits to avoid accidentally rewriting history. Other users question its necessity, drawing parallels to the "standards" XKCD comic and arguing the transitional cost isn't justified without a significant, undeniable advantage over Git. There are also practical concerns, such as workflow incompatibilities with GitHub's code review and the difficulty in conveying a compelling reason to switch from Git's entrenched ecosystem.
HN discussion
(424 points, 178 comments)
A California resident contacted Flock Safety's privacy contact to request the deletion of their personal information and vehicle data under the California Consumer Privacy Act (CCPA). Flock Safety refused the request, stating that they are a service provider for their customers (who own the data) and cannot process deletion requests directly. The company further explained that its LPR technology only captures publicly visible vehicle information, not sensitive personal data, and that data is retained for 30 days by default. The author believes Flock's response is legally inaccurate under the CCPA.
HN commenters were largely critical of Flock's position, with many arguing that its "we are just the processor" defense undermines the CCPA. Several commenters drew parallels to other service providers like cloud hosts or email marketing platforms, suggesting this creates a loophole where individuals must contact multiple parties to exercise their privacy rights. A minority of comments countered that Flock's role is similar to that of a tool for law enforcement, and that users should direct their requests to the actual data controllers (the municipalities). Some commenters were resigned to the situation, while others proposed more direct forms of opposition.
HN discussion
(461 points, 135 comments)
Chicago-based music fan Aadam Jacobs has digitized over 2,500 of his 10,000+ concert recordings from the 1980s onward, partnering with the Internet Archive to preserve them. The collection includes rare performances from artists like Nirvana (1989), Sonic Youth, R.E.M., and others, with volunteers improving the audio quality from original mediocre recordings. Jacobs and volunteers use cassette decks to digitize the tapes, then clean and catalog them, even identifying obscure tracks. The archive highlights the value of preserving such historically significant live music.
HN commenters expressed nostalgia for bootleg culture, noting how rare live recordings were once traded physically and sold in stores. Some shared personal experiences with bands encouraging recordings, like Ween, while others emphasized the importance of preserving such archives despite copyright concerns. A user mentioned DMCA takedowns as a threat, while another highlighted the rise of live music streaming. The discussion also explored technical aspects like AI-enhanced audio cleanup and interest in decentralized preservation (IPFS/P2P). Many praised the archive’s cultural significance and encouraged uploading personal bootlegs.
HN discussion
(375 points, 167 comments)
The article details an experiment where a Lean-verified implementation of zlib (lean-zip) was subjected to extensive fuzzing by a Claude agent. Over 105 million executions, the verified application code itself showed zero memory vulnerabilities. However, two bugs were found: a heap buffer overflow in the unverified Lean 4 runtime (specifically in `lean_alloc_sarray`) and a denial-of-service vulnerability in lean-zip's unverified archive parser. The author concludes that formal verification produced a remarkably robust codebase for the parts that were proven correct, but the discoveries highlight that verification's effectiveness is limited by the scope of what is proven and the correctness of the trusted computing base.
HN comments heavily criticize the article's title as clickbait, noting that no bugs were found in the *proven* code itself. A central theme is the distinction between the verified code and the untrusted runtime or unverified components, with one commenter calling the runtime bug a "big" discovery. The discussion also emphasizes the inherent challenges of formal verification, focusing on the difficulty of creating correct specifications and the problem of "spec completeness." Multiple commenters quote Donald Knuth's famous warning about proving code correct without testing it, and others reference the Halting Problem to argue that perfect bug elimination is theoretically impossible.
HN discussion
(316 points, 202 comments)
Claude Code Routines are autonomous, scheduled tasks that can run on three types of triggers: recurring schedules, API calls, or GitHub events. These routines run as full Claude Code cloud sessions with access to repositories, connectors, and environments, allowing them to perform tasks like backlog maintenance, alert triage, code review, deployment verification, documentation updates, and library porting. The feature offers multiple creation methods (web interface, CLI, or Desktop app), with each routine counting against the user's daily run allowance. Routines can be configured with specific repositories, environment variables, and connectors, and they run without requiring approval during execution.
The HN discussion reveals several concerns about Claude Code Routines, particularly regarding Anthropic's recent reduction in usage limits and whether these compute-intensive features effectively require the most expensive subscription plan. Users questioned the timing of releasing features that likely increase compute usage while simultaneously reducing limits. There was also speculation about Anthropic potentially creating a full replacement for OpenClaw (likely OpenHands), with some users noting the lack of moat compared to third-party alternatives. Additionally, several users raised questions about Terms of Service violations when integrating these routines with third-party applications, and some expressed frustration that organizational-level controls are missing. Interestingly, one commenter mentioned they've already been using similar scheduled features, though they experienced some bugs.
HN discussion
(112 points, 175 comments)
California's proposed bill A.B. 2047 would mandate censorship software ("censorware") on all 3D printers to block firearm component prints, criminalizing the use of open-source firmware or disabling these algorithms. The legislation would lock users into manufacturer ecosystems, enable anti-consumer practices like planned obsolescence and resale bans, and impose significant bureaucratic burdens on manufacturers and the state Department of Justice. Despite targeting rare "ghost guns" already banned under existing law, the bill threatens innovation, consumer choice, privacy, and could expand beyond guns to censor other content like copyrighted material or political speech.
Hacker News comments overwhelmingly criticize the bill as technologically implausible and misdirected, comparing it to absurd regulations like restricting saws or pens. Critics argue it fails to address the root cause of gun violence (e.g., ammunition access) and would be easily circumvented, while burdening innovators and small manufacturers. Many commenters suggest the bill is driven by lobbying from large printer manufacturers seeking to enforce DRM-like control or by gun-control advocates with poor technical understanding. Concerns include the precedent for global surveillance via "state-certified algorithms" and the expansion of censorship beyond firearms, alongside skepticism about enforcement against open-source or unlicensed hardware.
HN discussion
(215 points, 41 comments)
The article introduces Introspective Diffusion Language Models (I-DLM), which address the quality gap between diffusion language models (DLMs) and autoregressive (AR) models by fixing "introspective consistency" failures. I-DLM uses introspective strided decoding (ISD) to verify previously generated tokens while generating new ones in a single forward pass. Empirically, I-DLM-8B matches its AR counterpart's quality, outperforms LLaDA-2.1-mini (16B) by +26 on AIME-24 and +15 on LiveCodeBench-v6 with half the parameters, and achieves 2.9-4.1x higher throughput at high concurrency. Gated LoRA enables bit-for-bit lossless acceleration, and I-DLM integrates with SGLang for production deployment without custom infrastructure.
Commenters highlight I-DLM's practical innovations: thepasch notes it cleverly converts AR models (e.g., Qwen) into diffusers, enabling faster generation with identical outputs via LoRA adapters. Simianwords questions whether diffusion models can generate/introspect in sequential blocks, while mlmonkey points out I-DLM’s token-by-block approach challenges traditional diffusion’s "all-at-once" paradigm. Practical concerns include deployment (ilaksh asks about SGLang vs. vLLM integration) and model availability (scotty79 shares the I-DLM-32B link). Some express skepticism about speedup claims (e.g., memory vs. compute bottlenecks) and the paradigm’s novelty, while others emphasize excitement about bridging DLM and AR capabilities.
Generated with hn-summaries