Hacker News Evening Brief: 2026-06-18
Tonight’s batch spans nuclear energy’s European resurgence, a 10,000-repo GitHub malware campaign, AMD quietly stripping memory encryption from consumer chips, a new VCS from Epic Games built for binary assets, and an e-paper monitor that hits 60Hz. Dive in.
Security & Privacy
I found 10k GitHub repositories distributing Trojan malware
Summary: A researcher discovered over 10,000 GitHub repositories distributing Trojan malware through cloned repos that inject links to malicious zip archives into READMEs. The repos are not forks — they independently clone legitimate projects, list the original author as a contributor, then push suspicious commits every few hours to game search rankings. The campaign particularly exploits Bing search results, where malicious copies sometimes outrank originals. A detection script was written to identify the pattern through matching names, recent suspicious commits, and archive links.
HN Discussion: One commenter argued the repos target AI coding agents rather than humans — agents searching for dependencies may pull from whichever repo appears first in results. Users reported GitHub’s unresponsiveness to takedown requests, with one citing a blatantly pirated repo still active after two years of reports. The Bing search quality issue was highlighted as a contributor, with one user abandoning Bing after landing on a phishing page through its results.
AMD silently removes memory encryption from consumer Ryzen CPUs
Summary: AMD removed memory encryption support from consumer Ryzen CPUs through an AGESA firmware update, without notifying users that a security feature had vanished. The capability protects against physical memory attacks such as cold-boot key extraction and RAMbleed side channels. AMD engineers went silent when pressed about the rationale. The removal matters for full-disk encryption users: if a laptop is seized while running, the disk decryption key resides in plaintext in now-unprotected RAM.
HN Discussion: Some users dismissed the concern, arguing that physical hardware access already implies compromise regardless of memory encryption. Others pointed out the feature was never formally marketed for consumer CPUs, making the removal technically not a broken promise — though silently stripping a security capability remains alarming. One commenter noted memory encryption also guards against ECC-error side channels like RAMbleed, broadening the threat model beyond physical access.
Geopolitics & War
Swiss parliament lifts ban on new nuclear power plants
Summary: The Swiss parliament voted to lift the ban on constructing new nuclear power plants after a marathon debate during the 2026 summer session. The ban was originally introduced after the Fukushima disaster. The decision still faces a public referendum, with left-leaning parties and Greens strongly opposed. The policy shift aligns with a broader European trend where countries are reconsidering nuclear for energy security and decarbonization, particularly to address Switzerland’s seasonal hydro deficit in winter.
HN Discussion: Commenters highlighted Switzerland’s seasonal energy gap: abundant hydro in spring and summer but deficits in winter that nuclear could bridge. Some questioned the economics, citing Vogtle in Georgia and Ontario’s $500B nuclear plans as examples of extremely expensive builds. Others were optimistic about small modular reactors and predicted a startup boom once core SMR challenges are solved.
Business & Industry
TerraPower in Deal with Meta for Eight Natrium 345 MW Advanced Nuclear Plants
Summary: TerraPower and Meta announced an agreement to develop up to eight Natrium advanced nuclear reactors, each rated at 345 MW, to supply carbon-free power for data centers. The Natrium design uses a liquid sodium-cooled fast reactor paired with molten salt energy storage that can boost output to 500 MW during peak demand. The deal goes beyond a standard power purchase agreement, with Meta providing upfront funding, though financial terms were not disclosed. TerraPower’s first demonstration reactor is under construction in Wyoming with a target of 2030–2031.
HN Discussion: Skeptics questioned whether TerraPower can deliver eight production reactors on any reasonable timeline given that their first demo reactor remains unfinished. Commenters questioned Meta’s massive AI infrastructure spending, noting the company lacks a clear standalone AI product despite billions invested. The asserted 24/7/365 reliability claim for an unproven design was criticized as premature.
AI & Tech Policy
Launch HN: TesterArmy (YC P26) – Agents that test web and mobile apps
Summary: TesterArmy, a YC P26 startup, launched an AI agent platform that automatically runs browser-based tests across web and mobile app journeys, then delivers bug reports with screenshots and recordings. The product requires no SDK, no test scripts, and no infrastructure — users paste a URL or upload an app binary to begin. Integrations include Slack, GitHub, CI/CD, Vercel, Expo, Discord, and webhooks. The platform targets teams that find traditional end-to-end tests slow to set up and expensive to maintain.
HN Discussion: Commenters questioned cost-effectiveness compared to LLM-written deterministic E2E tests, asking about token costs and result stability given AI non-determinism. One user argued that if AI coding assistants write the code, they should also write their own verification tests rather than requiring a separate external tool. Comparisons were made to Revyl, another agent-driven testing product, with users seeing this category as the future of QA.
Has W Social switched to closed source?
Summary: Elena Rossini reports that W Social, a European Bluesky fork positioned as the EU’s alternative to X, appears to have quietly abandoned its open-source commitments. The platform secured endorsements from high-profile EU politicians and blanket press coverage despite transparency concerns. Rossini contrasts W Social with Eurosky, a competing non-profit ATProto network that builds in the open with full transparency but received zero press attention. The article raises concerns about W Social’s identity verification system, user data being used for AI model training, and the gap between public image and operational reality.
HN Discussion: Commenters called W Social “extremely shady” from its first HN advertisement, with one user claiming six accounts under different names despite the verification system. Users questioned why EU politicians and national media promoted a for-profit venture while ignoring the transparent Eurosky alternative. The consensus was that EU institutions prioritize data sovereignty optics and supporting EU businesses over genuine open-source principles.
DeepSeek Introduces Vision
Summary: DeepSeek added vision capabilities to its chat platform, enabling the model to understand and describe the contents of uploaded images. The feature performs visual understanding rather than image generation or modification. The DeepSeek chat application still lacks native text-to-speech and speech-to-text, despite competitors offering these as standard. Users have reported the model increasingly responding in Chinese, including during reasoning steps, raising questions about recent silent updates.
HN Discussion: Commenters noted the absence of ASR and TTS features as surprising for a major AI platform in 2026. Several users reported the model responding in Chinese more frequently, including within reasoning chains, prompting speculation about a silent model update. One commenter expressed broader frustration with model regressions, wishing for the ability to pin earlier versions of AI models.
We built a persistent agent memory layer on Elasticsearch with 0.89 recall
Summary: Elastic built a persistent, multi-tenant agent memory layer using three Elasticsearch indices, achieving 0.89 recall at K=10 across 168 test questions. The architecture combines hybrid retrieval (BM25 plus vector search) with reciprocal rank fusion, a reranker, supersession logic, and temporal decay for memory lifecycle management. Per-user isolation is enforced through document-level security to prevent tenant leaks. The implementation is fully open-source and targets scenarios where agents need persistent cross-session context.
HN Discussion: Critics called Elasticsearch massive overkill for agent memory, arguing that SQLite, LanceDB, or any vector database would be more appropriate. The blog post’s writing style was criticized as heavily AI-generated, prompting one commenter to post a human-readable summary on Pastebin. Commenters questioned whether there is any evidence that these complex memory architectures actually improve agent performance in production.
Migrate from OpenClaw
Summary: Nous Research published a migration guide for moving from OpenClaw to their Hermes Agent platform, including a hermes claw migrate command that imports OpenClaw configurations and maps config keys to Hermes equivalents. Multi-provider setups collapse into a single OAuth portal offering access to 300+ models. Hermes Agent positions itself as a more polished, less technical alternative for running AI agent workloads.
HN Discussion: Commenters compared the two platforms: Hermes as more polished for less technical users, OpenClaw as deeper with more capabilities including multi-agent setups, cron jobs, and extensive skill systems. A user raised the plagiarism controversy around Nous Research, linking to a prior HN discussion about alleged training data practices. Critics questioned the value proposition of either platform, asking why not simply run Codex or Claude Code in a Docker container. One user reported Hermes consumed over 10k tokens of context just listing default skills.
Tech Tools & Projects
Ubiquiti: Enterprise NAS, Built on ZFS
Summary: Ubiquiti announced an Enterprise NAS built on ZFS, featuring dual 25 Gigabit SFP28 ports, redundant power supplies, and 12 drive bays. Priced at $3,999, the product continues Ubiquiti’s pattern of one-time hardware purchases without recurring subscription fees. It positions itself against enterprise storage vendors by combining ZFS data integrity — snapshots, checksums, and merkle-tree-based delta backups — with Ubiquiti’s network ecosystem integration.
HN Discussion: Commenters praised the lack of recurring licensing fees, contrasting this with the industry-wide shift toward MRR storage products. Performance questions were raised about whether spinning HDD arrays can saturate dual 25Gbps links, with one user reporting difficulty achieving maximum throughput on a comparable TrueNAS ZFS setup with L2ARC optimizations. ZFS enthusiasts welcomed the filesystem choice over alternatives lacking equivalent fault tolerance.
Emacs 31 is around the corner: The changes I’m daily driving
Summary: Emacs 31 is in pre-release with significant built-in improvements that reduce reliance on external packages. The author tracks new features through an “Emacs Solo” configuration using zero external packages, marking each Emacs 31 addition with breadcrumb comments. The post details specific upcoming changes including built-in LSP improvements, native configuration options, and enhanced editing modes that previously required third-party packages. Names and defaults may still shift before final release.
HN Discussion: A user asked why anyone would use Emacs in 2026 when most jobs require working inside containers with VSCode, prompting a lively defense from long-time users. Veterans cited Emacs keybindings as universally supported via GNU readline across shells, browsers, and network equipment, making the muscle memory transferable. Several commenters noted they briefly left Emacs for VSCode during the AI integration gap but returned once tools like Claude became usable inside Emacs.
Modos Color Monitor Pushes E-Paper Displays Further
Summary: Modos, a two-person startup, is developing the Modos Flow: a 13.3-inch color e-paper monitor with 3,200 × 2,400 native resolution, touch input, and a 60Hz refresh rate. The display targets users seeking low-eye-strain, outdoor-visible, and low-power computing. IEEE Spectrum reports this as a significant advance over previous e-paper monitors, which were limited to monochrome and lower refresh rates. The product represents growing momentum in alternative display technologies alongside RLCD panels from Daylight Computer and e-ink devices from Boox.
HN Discussion: Commenters expressed excitement about the alternative display space, citing Daylight Computer’s RLCD, Boox devices, and now Modos as signs of genuine momentum. The 60Hz refresh rate and 3,200 × 2,400 resolution on e-paper were called ambitious, with users wishing the team success. An ideal device form factor was described as an ultralight Android tablet with keyboard case and outdoor-readable display, sufficient for note-taking and SSH.
.gitignore Isn’t the Only Way to Ignore Files in Git
Summary: The article explains three Git ignore mechanisms: the familiar .gitignore, the per-repository .git/info/exclude, and the global ~/.config/git/ignore. The .git/info/exclude file is useful for ignoring personal files in a shared repo without modifying the committed .gitignore. The global ignore applies across all repositories on a machine, ideal for OS-specific files like .DS_Store or IDE artifacts. The author recommends using ~/.config/git/ignore over the older ~/.gitignore_global approach, leveraging the XDG config directory pattern.
HN Discussion: Developers praised the global git ignore as an underused feature that prevents cluttering every project’s .gitignore with personal IDE, OS, and AI tool files. One user recommended the “attic” pattern: a globally ignored directory name that can be created in any project to hold scratch files. A counterpoint was raised that .DS_Store should go in .gitignore rather than the global exclude, since not every Mac user will configure their global ignore.
Image Toolbox (T8RIN)
Summary: Image Toolbox is an open-source Android app for advanced image manipulation, offering crop, draw, filters, OCR, format conversion, and compression in a single package. The tool covers workflows from basic cropping to batch format conversion and text extraction, all processed on-device without cloud services. The project is actively maintained on GitHub by developer T8RIN and has built a reputation as an essential FOSS Android utility.
HN Discussion: Users praised Image Toolbox as an essential FOSS Android app, particularly valuing its compression and format conversion capabilities. The open-source nature and active development were frequently cited as reasons it appears on must-have app lists.
Dwarf Fortress in the Browser
Summary: Remote-DF runs a Dwarf Fortress instance remotely and streams the game’s ASCII interface to a browser client, enabling play without local installation. The implementation uses SSH as the transport protocol for the remote connection, rendering the console-style display in a web window. Videos shared by the author demonstrate the setup running the classic game in a browser.
HN Discussion: Commenters questioned the choice of SSH for low-latency remote rendering, suggesting WireGuard, WebRTC, or QUIC as more appropriate alternatives. Users criticized the massive auto-generated README as unreadable, noting they tune out when encountering “vibe-coded” documentation of that style. A practical question was raised about why not simply use a remote desktop tool like RustDesk to stream the game.
Lore – Open source version control system designed for scalability
Summary: Lore is an open-source, content-addressed version control system maintained by Epic Games, designed for projects combining code with large binary assets. It targets game development workflows where Git struggles with 3D models, textures, audio files, and other non-text assets that multiple artists need to collaborate on. The system offers centralized architecture with on-demand scalability, free branching, verifiable tamper-evident history, and a full-surface API across C/C++, C#, Rust, Go, Python, and JavaScript. Lore is positioned as a direct competitor to Perforce.
HN Discussion: Game developers welcomed Lore as a much-needed Perforce challenger, noting Perforce’s complexity in branching and administration as key pain points. A user asked about versioning personal files — music projects, photos, videos, markdown notes — as a monorepo. Commenters clarified Lore is specifically designed for Unreal and game development, not intended to replace Git for general software. Git’s verbose push output was cited as evidence that even basic VCS UX remains unresolved.
Web & Infrastructure
A website that lists websites to submit your website to
Summary: Submission.Directory is a curated directory of 51 sites where developers can submit their website, startup, or product for visibility and backlinks. Categories include launch platforms (Product Hunt, Hacker News), startup directories, SaaS directories, AI tool directories, maker communities, and design galleries. Each listed site includes domain rating, link type (dofollow, nofollow, profile, guest post), and whether it is free or paid. The project targets SEO practitioners and indie makers looking to earn quality backlinks and improve ranking across search engines and AI answer engines.
HN Discussion: Commenters drew parallels to 1990s submission services like Submit It, noting the challenge has shifted from search engine crawling to audience-targeted discovery. Several users joked about the recursive nature of a directory listing directories, referencing Wikipedia’s list-of-lists-of-lists. Some questioned whether drive-by link submissions for SEO juice violate community norms on many of the listed sites.
Vinyl Cache and Varnish Cache
Summary: Vinyl Cache is the community fork of Varnish Cache, the HTTP accelerator widely deployed in content delivery infrastructure. The project documents a substantial history of security advisories including request smuggling, HTTP/2 desync attacks, and rapid reset vulnerabilities across Varnish versions. Vinyl Cache 9.0.1 is the current community release, while Varnish Cache 8.0.2 and 6.0.18 remain maintained upstream. The fork appears to have revitalized development, with features stalled for years now actively progressing.
HN Discussion: Commenters noted the fork has positive side effects: development that was blocked to protect Varnish Software’s commercial offering is now moving forward. Questions arose about whether Fastly still uses Vinyl/Varnish Cache in production and who other major deployers are. The diplomatic tone of the article was noted, with users wondering about the actual relationship between Varnish Software and the FOSS project.
Academic & Research
Hospitals and universities repurposing drugs at 90% lower cost
Summary: A King’s College London study highlights how hospitals and universities are repurposing off-patent drugs for new indications at up to 90% lower cost than developing new medicines. Drug repurposing leverages existing safety data and manufacturing infrastructure, dramatically reducing clinical trial timelines and development costs. The approach is particularly valuable for rare diseases where pharmaceutical companies lack financial incentive to develop new treatments. The research documents multiple case studies where repurposed drugs achieved clinical success at a fraction of conventional drug development expense.
HN Discussion: Commenters shared personal experiences with drug repurposing for rare diseases, including Huntington’s Disease research funded by nonprofits like Cures Within Reach. The esketamine example drew criticism as the inverse problem: a company modifying off-patent ketamine into an enantiomer to re-patent it, despite evidence it may be less effective. Users pointed to systemic incentive failures where patent modifications are more profitable than genuine repurposing.
Advanced Compilers: The Self-Guided Online Course
Summary: Cornell’s CS 6120 is a PhD-level advanced compilers course by Adrian Sampson, available as a free self-guided online curriculum. The course covers intermediate representations, data flow analysis, classic optimizations, parallelization, JIT compilation, and garbage collection. Implementation tasks use LLVM and a custom educational IR, with open-ended assignments designed to solidify theoretical understanding through real code. The self-guided version includes video lessons, written notes, and a structured timeline interleaving paper readings with implementation tasks.
HN Discussion: A compiler engineer criticized the dynamic compilers section for focusing on trace compilation, arguing it is a dead end in favor of type feedback, speculation, and tiering. Some questioned whether topics like dead code elimination and SSA form truly belong in an “advanced” course rather than an introductory one. Discussion touched on ML-based compiler optimizations, including Rust’s compiler reportedly using machine learning to detect unsafe code patterns.
Unity vs. Floating Point
Summary: Unity’s Mathf functions cast floats to double, call the double-precision math implementation, then convert back to float — adding unnecessary overhead for trigonometry, rounding, and comparison operations. The .NET Core 2.0 introduction of System.MathF in 2017 provided native single-precision alternatives, but Unity has not updated Mathf nearly a decade later. The author disassembles the generated code to show exactly where the hidden double-precision round-trip occurs. Only Mathf.Abs avoids the penalty; all other math functions suffer from it.
HN Discussion: A commenter expressed a wish for JavaScript or an open-source array language to support single-precision floats, noting the value of the assembly-level breakdown in the post. The discussion highlighted how game engine code paths can carry decade-old technical debt that silently degrades performance for all Unity developers.
History & Science
Seven Perfect Shuffles Randomize a Deck of Cards. But How Many Sloppy Ones?
Summary: A decades-old proof established that seven riffle shuffles suffice to randomize a 52-card deck, but only if the deck is cut with near-perfect precision. A new proof relaxes that assumption, showing approximately 14 imperfect (“sloppy”) riffle shuffles achieve full randomization. The original GSR shuffling model cuts at a binomially distributed point, meaning roughly one-fifth of the time the split is at least as asymmetric as 21-31 cards. The result provides tighter bounds on mixing times for Markov chains used in PRNG analysis.
HN Discussion: One commenter pointed out that seven perfectly interleaved riffle shuffles actually return a deck to its original order, making the headline technically wrong for a specific famous result. A mathematician clarified that the GSR model’s “more or less in half” hides significant variance. Discussion touched on whether the proof has practical consequences for evaluating PRNG algorithms.
How Alberta Eradicated Rats
Summary: Alberta, Canada has been rat-free for over 70 years through a systematic eradication program begun in the 1950s. The province established a rat control zone along its eastern border, deploying pest control officers to intercept introductions from incoming shipments. The campaign used warfarin-treated bait alongside community education — including a memorable demonstration where a pest control officer ate warfarin-treated oats before an audience to prove safety. For context, New York City’s rat population is estimated at three million, roughly one per three residents.
HN Discussion: Readers praised the article for detailing how eradication actually worked rather than treating Alberta’s status as a given. Commenters noted that while Alberta is rat-free, mice remain persistent, illustrating the difficulty of eliminating smaller rodent species. New Zealand’s “Predator Free 2050” program was referenced as a comparable eradication effort targeting rats, possums, and stoats.
System Administration
Microsoft new Outlook takes 10 seconds to do what Outlook Classic does instantly
Summary: Microsoft’s new Outlook, built on WebView2, takes up to 10 seconds to perform operations that Outlook Classic handles instantly. The web-based architecture loads unnecessary data, renders everything on every window, and has a fundamentally wrong load order. Even basic Windows utilities are affected: Notepad now takes 3-4 seconds to load on Windows 11 and includes in-app purchases for AI writing features. Modern NVMe SSDs with 20+ Gbps throughput are now insufficient to make basic applications feel snappy.
HN Discussion: Commenters contrasted the new Outlook with Fastmail’s web client, which matches or exceeds Outlook Classic’s speed, proving the web architecture isn’t the root cause. Users lamented that SSDs eliminated old latency only for software bloat to recreate it. One commenter wished Microsoft would get “the performance religion” the way they previously got “the security religion” internally.
I hate compilers
Summary: The post details the challenges of achieving reproducible C/C++ builds when vendoring compiled WASM binaries into other projects. Non-determinism creeps in through compiler-embedded timestamps like __DATE__ and __TIME__ macros, build-path leakage, and toolchain version drift. The author argues that low-level engineering is harder than high-level work because it is tightly coupled to specific build environments and toolchain behavior. The article is hosted behind an Anubis proof-of-work anti-bot system, which itself became a point of contention.
HN Discussion: A commenter pointed out that using __DATE__ and __TIME__ to stamp builds is a deliberate feature — if you don’t want non-deterministic output, don’t use them. antirez criticized the Anubis anti-bot system for forcing every browser to compute energy-hungry hashes, creating the same energy waste the anti-scraping stance claims to oppose. Discussion touched on the gap between reproducible-build difficulty in low-level versus high-level environments.
Sogen – High-performance Windows and Linux userspace emulator
Summary: Sogen is a high-performance userspace emulator capable of running Windows applications on Linux and vice versa without kernel-level modifications. Videos shared on social media show the author running Call of Duty Modern Warfare 2/3 through Sogen, with multiplayer reportedly functional. The project targets syscall-level emulation, translating OS calls between platforms at the user-space boundary. The landing page provides minimal technical detail, prompting community requests for architecture documentation.
HN Discussion: Users noted the landing page explains almost nothing about how the emulator works, with confusion over whether it cross-emulates OSes or merely sandboxes same-OS applications. The demonstrated ability to run modern Windows games including multiplayer on Linux via userspace emulation was met with surprise. A developer asked whether an equivalent userspace emulator exists for Android native binaries.
Other
The Harajuku Moment
Summary: Tim Ferriss recounts a personal turning point in Tokyo that catalyzed significant behavioral change. The essay argues that people rarely change from accumulating more how-to advice; rather, change requires an emotional reckoning where inaction becomes more painful than action. The narrative uses a specific anecdote from the Harajuku district as the setting where motivation shifted from intellectual understanding to committed action. The piece connects to the motivation-to-change literature: readiness builds over time until a crystallizing event surfaces the desire for change.
HN Discussion: One commenter clarified the piece is about personal growth rather than Harajuku’s cultural fashion scene, noting over 95% of people in the area are now tourists. Another connected the thesis to motivation-to-change literature, suggesting desire to change builds unconsciously before reaching awareness. A reader noted that exposure to verbose AI content has trained them to skim, causing impatience with long-form narrative hooks.
Emacs, how it all started (for me)
Summary: A long-time Emacs user since 2008 chronicles the chaotic path of discovering Emacs through Microsoft Word, Notepad, and the French programming site Le Site du Zéro. Experimentation with many programming languages naturally led to Emacs as a “good-enough” solution for diverse language ecosystems. The article is framed as personal lore rather than a tutorial, contributing to the April 2026 Emacs Carnival community writing topic. Despite strong opinions on programming languages, the author remains agnostic in the editor war, having never tried Vi or Vim.
HN Discussion: A commenter described the typical pre-2015 journey of learning about legendary hackers who used Emacs or Vim, then choosing one and expanding proficiency. A newer user expressed frustration with claims that Emacs is “self-discovering,” saying they’ve never successfully used its built-in help system and resort to Google. A vim-to-nvim migrant noted that Emacs felt overwhelming compared to Neovim’s approachability.
Show HN: Gerrymandle - Daily puzzle game where you redraw electoral districts
Summary: Gerrymandle is a daily browser puzzle where players redraw electoral district boundaries to win elections for their assigned party. Players click adjacent tiles to form connected districts that must be contiguous, equally sized, and cover all houses on the map. The game teaches gerrymandering techniques like packing and cracking through hands-on puzzle solving, with a new map each day and hints available for stuck players. The project turns a complex political concept into an accessible daily activity.
HN Discussion: Commenters praised the concept as ideal for high school civics classes, calling it a creative way to build gerrymandering awareness. The tie-breaking rule — nobody wins a tied district — was noted as unrealistic but acceptable for conveying the concept. Several users reported initial confusion with the rules, suggesting introductory levels of increasing difficulty would improve onboarding.
What was nice about the UI of Windows 2000
Summary: A retrospective on Windows 2000’s UI design praises its visual clarity: solid background colors, labelled desktop icons, and a taskbar with functional depth. The Start button demonstrated physical affordance — pressing it visually sank the button, providing immediate feedback that it was activated. The UI used crisp font rendering and skeuomorphic indicators that helped users discover functionality without training. The article argues the software industry once worked toward shared visual paradigms and platform consistency, unlike modern fragmented design.
HN Discussion: A commenter shared a conversation where a non-technical friend accepted that smartphone UIs require randomly trying everything on screen — a failure of modern discoverability. Users noted the shift away from leading users through interfaces, as designers assume everyone knows how bottom-corner menus work. The article was criticized for omitting Windows 2000’s “Personalized Menus” feature, which hid infrequently-used programs and was widely considered the one must-disable setting.