← Oct 10Sunday, October 11, 2026 · all daysOct 12 → · today »
1.Build your own decision model(nishtahir.com)
454 points by softwaredoug 3 days ago | 102 comments | permalink
tl;dr: Instead of generating tokens autoregressively, you can build a fast "decision model" by running a single forward pass on an LLM (e.g., Qwen3-1.7B) and masking logits to only allowed answer tokens (A-E), picking the highest probability. Benchmarked on CommonsenseQA, this yields ~59% accuracy (62% after finetuning), but the raw probabilities are badly miscalibrated—overconfident in the 0.9+ range. Temperature scaling (T≈3.8) flattens the distribution so confidence scores actually match empirical accuracy.
HN Discussion:
  • ↓This is just classifiers rebranded with hype, which is frustrating to ML practitioners
  • ↑Enthusiastic builders sharing their own implementations and variations inspired by the approach
  • ~Lighter-weight or browser-based alternatives offer better performance for local execution
  • ↓Skepticism about the hype — asking for real use cases built with this technique
  • ↓Post-hoc temperature tuning for calibration feels methodologically questionable
2.2D Vehicles(patkerr.co.uk)
621 points by Michelangelo11 6 days ago | 118 comments | permalink
tl;dr: In 1996, the author wrote a 2D rigid-body physics simulation in GFA BASIC on an Atari ST that became the vehicle system in the original Grand Theft Auto. For its 30th anniversary, they've recreated it in JavaScript ("Motion Lab") with playable Car, Ship, and Brick modes. The post details the design: a proper rigid-body core with torques (uncommon in games then) layered with a deliberately simplified, non-physical tyre model that applies velocity-proportional resistance along and across each wheel's heading to produce convincing car-like handling.
HN Discussion:
  • ↑Appreciation for the historical significance and nostalgia of GTA 1/2 car physics
  • •Related references to other foundational 2D driving/physics models and games
  • ↓Criticism that the blog post's diagrams are poor and lacks explanation of design reasoning
  • ↑Praise for the prototyping approach and value of simple 'toy' physics
  • ↑Personal struggles replicating similar 2D vehicle physics, highlighting the achievement
3.A city-building game in which the city would prefer you didn't(housing.over.pizza)
542 points by JumpCrisscross 3 days ago | 195 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Shares similar satirical game ideas about bureaucracy and development obstacles
  • ↑Praises the game as fun gallows humor reflecting real building difficulties
  • •Discusses real SF housing policy and potential solutions like SB 79
  • ↑Suggests SF actually wants restrictive housing policies by design
  • ↑Extends the critique to other contexts like UK's HS2 infrastructure obstruction
4.The Lightbulb Computer(lightbulbcomputer.com)
581 points by oskarth 4 days ago | 95 comments | permalink
tl;dr: A former Apple designer prototypes the "Lightbulb Computer," a projector-plus-camera device in an Edison bulb form factor that projects interactive interfaces onto real-world surfaces—recipes on counters, maps on tables, controls near smart home devices—and responds to voice and pointing gestures. The author pitches it as a social, ambient alternative to both screen-based and face-worn (AR/VR) computing, arguing projector and computer vision tech is nearly ready for a consumer version, with privacy handled via on-device processing and hardware camera guards.
HN Discussion:
  • ↑Creator sharing technical details and emphasizing it's a research prototype about interaction design
  • ↑Enthusiastic endorsement with suggestions for UX improvements like timestamped voice triggers
  • ↓Concerned about pervasive surveillance and always-present ambient computing in private spaces
  • ~Skeptical that a real product would preserve privacy and avoid cloud dependency
  • ↑Supports the vision of escaping screens as attention black holes via ambient computing
5.Talorys – A self-hosted personal AI agent on Cloudflare's free tier(github.com)
323 points by rociiu 3 days ago | 150 comments | permalink
tl;dr: Talorys is an open-source single-user AI assistant that deploys to your own Cloudflare account via `npx create-talorys@latest`, using Pages, Workers, SQLite-backed Durable Objects, and Workers AI (GLM-4.7-flash) — all on the free tier with no external servers or telemetry. It offers chat, persistent memory, tasks/notes/projects, and scheduled automations via Durable Object alarms, with non-AI features continuing to work if the daily Workers AI quota is exhausted. Data stays in your Cloudflare account, and the installer is idempotent with built-in update, backup, and password-reset commands.
HN Discussion:
  • ↓Calling this 'self-hosted' is misleading since it entirely depends on Cloudflare's infrastructure
  • ↑The open-source nature means it's trivial to swap out Cloudflare components, so complaints miss the hacker spirit
  • ~The 'self-hosted' label likely just means no third-party dependency beyond user's own Cloudflare account
  • ↓Cloudflare's free AI tier has billing issues and poor support, making reliance on it risky
  • ~Constructive technical feedback — the memory feature needs semantic recall to actually work well
6.Grieving the loss of details(purplesyringa.moe)
490 points by signa11 7 days ago | 410 comments | permalink
tl;dr: A self-described "low-level coder" who thrives on deep understanding of machine details (assembly, cycle-counting, performance optimization) laments that the industry's shift toward LLM-driven development and high-level abstraction has devalued their expertise. They can't use LLMs effectively because their mind requires reinventing concepts from first principles, and they find being redirected to an AI instead of consulted as an expert deeply alienating. What started as career anxiety has escalated to considering disability, as hobbyist retrocomputing communities are also increasingly dominated by LLM enthusiasts.
HN Discussion:
  • ↑Shares the author's sense of grief, alienation, and loss of craft due to LLMs
  • ~Empathizes but notes LLM adoption is not universal; some coders still work the old way
  • ↓Rejects the article's premise that LLMs produce better low-level code than skilled humans
  • ↓Software is now infrastructure; artisanal preferences don't entitle one to a livelihood
  • ~Tooling gap: we need better ways to think deeply about code alongside LLMs
7.Cloudflare acquires Deno(deno.com)
1370 points by ilreb 4 days ago | 713 comments | permalink
tl;dr: Cloudflare is acquiring the entire Deno team, which will shift focus from the standalone Deno runtime and Deno Deploy to a shared platform built around Cloudflare Workers and Durable Objects. Deno will receive bug fixes and security updates for one more year before development ends (remaining open source for community continuation), while Deno Deploy shuts down in six months with migration support to Workers. JSR and rusty_v8 work will continue under Cloudflare, with the team's new focus being "celld," a distributed application model targeting AI agent workloads.
HN Discussion:
  • ↑Deno is effectively dead and the announcement framing is misleading
  • ↑Sadness and mourning for Deno as a beloved runtime being shut down
  • ~Deno lost its way earlier by prioritizing npm compatibility over its original vision
  • •The celld/workerd merger is the more interesting technical story being overlooked
  • •This is part of a broader concerning trend of developer tooling consolidation via acquisitions
8.Triple-A Minesweeper(minesweeper.mikelacher.com)
1417 points by robin_reala 4 days ago | 284 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Praises the parody and suggests additional tropes like dialogue or shader compilation to lampoon
  • ↑Adds to the satire by listing more modern AAA annoyances missing from the parody
  • ↑Interprets the work as critiquing excessive handholding in modern AAA games
  • •Shares related parody works like Minesweeper movie or AAA Mario
  • •Questions technical aspects such as whether voices were AI-generated
9.I would like the value of my home to rise, while my property taxes fall(conversableeconomist.com)
275 points by colinprince 3 days ago | 683 comments | permalink
tl;dr: Several US states have recently slashed property taxes on owner-occupied homes, shifting the burden onto commercial property, state funds, and sales/income taxes—effectively converting a local service-financing tool into redistributive taxation. Homeowners, despite seeing major post-COVID wealth gains, are driving this backlash because property taxes function as wealth taxes requiring cash from asset-rich, income-poor owners (especially older voters). The ironic result: lower property taxes will push home prices even higher, further enriching current owners while worsening affordability for future buyers and weakening local government funding for schools and police.
HN Discussion:
  • ↑Shifting taxes from homeowners to commercial/rental properties unfairly burdens renters who are less wealthy
  • ↓Property taxes unfairly punish owner-occupiers whose unrealized home value gains just raise cost of living
  • ↑We should tax wealth hoarding more to free up underused homes for younger families
  • ~Land value taxation (possibly with income-based caps) is a better solution than current property tax
  • •Property tax mechanics are misunderstood; budgets drive rates, not valuations directly
10.REA Reverse – Engineer Anything(rea.tools)
768 points by modinfo 4 days ago | 327 comments | permalink
tl;dr: REA is a tool that connects to coding agents (via `npx rea-agents@latest setup`) to let them reverse-engineer software by inspecting binaries, scripts, and runtime behavior. The agent can read disassembly, decompiled code, and loaded scripts, then explain or rebuild functionality in plain English. Demos include recovering Chrome's dinosaur game acceleration rule (0.001 per update, max 13) and Windows Calculator's percentage logic (which differs between +/− and ×/÷ operations).
HN Discussion:
  • ↑Quality of REA's output (e.g., Touhou decomp) is impressive compared to typical AI decompilation efforts
  • ~Frontier models already handle reverse engineering well without specialized tooling like REA
  • •This tool/trend is driving the wave of AI-generated clones and game mashups we're seeing online
  • ↑AI reverse engineering heralds a future of 'liquid software' where programs are manifested on demand
  • ↓REA has significant technical limitations (e.g., jadx-based Android reversing doesn't scale)
11.Bitwarden Dual License Model(community.bitwarden.com)
397 points by Cider9986 3 days ago | 300 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Change is understandable and acceptable as long as source remains available and self-hosting works
  • ↑Bitwarden's software quality is poor; third-party clients and alternatives are preferable
  • ↑This licensing shift was inevitable once they accepted VC funding
  • •Indifferent to licensing drama as long as the product continues to work well
  • ↑Concerned about security implications, build verification, and trust in forks or self-hosting
12.Apple/macOS removed from official Unix registry(opengroup.org)
291 points by john_alan 3 days ago | 243 comments | permalink
tl;dr: Summary not available. (The provided article content is generic marketing copy from The Open Group about UNIX certification benefits — it contains no information about Apple/macOS being removed from the UNIX registry, despite the headline.)
HN Discussion:
  • ↓The article's claim is factually wrong — macOS is still listed under UNIX 03 certification
  • ↓The headline is misleading or clickbait and lacks sufficient information to assess the claim
  • ↑If true, the removal just acknowledges reality since UNIX certification is no longer relevant compared to Linux
  • ↑UNIX certification was always odd/meaningless since it applied to configurations nobody actually used
  • ~macOS is becoming less developer-friendly, pushing creative users to alternatives like NixOS
13.`123456' password used in Danish CPR data breach(cphpost.dk)
407 points by baal80spam 3 days ago | 216 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Systemic failure requires accountability across all levels, not just scapegoating individuals
  • ↑Shock and dismay at continued use of trivially weak passwords in important systems
  • ~Blame lies with the broader system design, including unmonitored database access and misuse of CPR numbers for authentication
  • ↑Third parties with registry access should face regular mandatory security audits like PCI compliance
  • •Technical skepticism about whether the weak-password account was actually active/exploitable
14.Sorry, I'm in a meeting(iminafleeting.com)
1032 points by splintersio 4 days ago | 283 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑Shares similar real-world experiences of using fake meetings to protect focus time
  • ↑Finds the fake meeting scripts humorous and accurate to real meetings
  • ↑Compares it to classic 'boss mode' features in old software/games
  • ↓Criticizes the audio for being unrealistic—too clear, no overlaps, no diverse accents
  • •Reports technical issues with the site (broken audio, non-working regions)
15.Show HN: WallHop – 12ft.io is gone, so I built a replacement(wallhop.io)
271 points by TheWaffleHax 3 days ago | 120 comments | permalink
tl;dr: WallHop is a replacement for the defunct 12ft.io paywall bypass tool. Users can skip paywalls by prefixing URLs with a slash, using an iPhone Shortcut, a bookmarklet, or raw/API endpoints. The creator recommends subscribing to news sites you read regularly, pitching WallHop as a tool for one-off articles.
HN Discussion:
  • ↑Enthusiastic endorsement of the tool as convenient and worth bookmarking
  • •Curiosity about why archive.is remains the most reliable fallback method
  • ↓Reports that the tool fails on certain sites like Reuters and The Hindu
  • ↓Questions about the legality and ethics, suggesting better journalism funding models
  • ~Debate over whether the project should be open source given its GPL-3 origins
16.Our $445M Series D(oxide.computer)
697 points by ahlCVA 4 days ago | 329 comments | permalink
tl;dr: Oxide Computer raised a $445M Series D led by Eclipse, with participation from existing investors (USIT, Riot, Jane Street) and new backers Atreides Management and strategic investor AMD. The company recently became profitable (paying income tax on operations) but needs the capital to fund inventory and manufacturing for a backlog that significantly exceeds supply. The round builds on their Series C goal of de-risking the company to ensure longevity and independence while scaling production.
HN Discussion:
  • ↑Oxide is an inspiring, admirable company doing things the right way
  • ↑Praise for Oxide's excellent communications and messaging style
  • ~Mild criticism of Oxide's AI-focused social media messaging
  • •Questioning why they chose equity over debt financing for inventory
  • ~Confusion or skepticism about Oxide's product offering and technical stack (no GPU support)
17.Eye of Sauron: Long-Range Hidden Spy Camera Detection (2024)(usenix.org)
338 points by ortusdux 6 days ago | 83 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Related detection techniques like non-linear junction detectors already exist and are commercially available
  • ~The approach is clever but seems like obvious engineering that should have existed earlier
  • ↑The technique is genius and offers hope in the privacy cat-and-mouse game
  • ↓Practical concerns about RF interference from LTE/5G limiting the method's effectiveness
  • ~Simpler alternatives like IR camera checks or buffer-based mitigations exist
18.Programming Isn't Special(blog.glyph.im)
276 points by ingve 5 days ago | 265 comments | permalink
tl;dr: Programmers wrongly accept AI coding tools because they consider programming merely functional rather than creative, but the author argues programming is art—even mundane code involves aesthetic decisions, and great innovations (like the author's Deferred abstraction) emerge from the contemplation of tedious work. Using AI to automate away mundane programming eliminates the practice opportunities needed to develop craft and achieve occasional moments of greatness, just as it would in writing, music, or visual art. Programmers should resist AI encroachment the way other creative professionals do.
HN Discussion:
  • ↓Most programmers aren't interested in originality or craftsmanship, undermining the art analogy
  • ↑Programming genuinely is art with real aesthetic value worth preserving
  • ↓Programming is mostly tedious plumbing/business glue, so AI automation is welcome relief
  • ~Programming is a medium that can be either art or engineering depending on the practitioner
  • ~Code differs from traditional art because the program, not the code, is what users consume
19.Telegram Desktop vulnerability allowed any user's file to be stolen(beaksec.github.io)
461 points by g-b-r 4 days ago | 317 comments | permalink
tl;dr: Telegram Desktop (through 7.2.8) had a vulnerability where clicked `tg://` links were passed between instances over a local socket using a semicolon-delimited format that wasn't escaped, allowing command injection. Attackers could chain this with the internal `interpret:` URI scheme—originally used for publishing releases—which reads arbitrary files and sends them to a chat without authorization checks, enabling exfiltration of session files and full account takeover. Fixed in 7.2.9 (CVE-2026-107181) by removing `interpret://` and properly escaping the socket protocol.
HN Discussion:
  • ↑Complex input formats inevitably become exploitable bytecode-like interpreters
  • ↑Operating systems should sandbox apps by default to prevent arbitrary file access
  • ↑Telegram has broader trust issues like silently re-enabling disabled settings
  • ↓This is a routine clientside bug with little broader lesson to learn
  • ↓Writeup quality was poor due to apparent AI-generated phrasing
20.Sharing AI progress in mathematics(openai.com)
1353 points by OfficialTurkey 7 days ago | 1530 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Personal reflection on AI solving problems researchers spent decades working on
  • ↑Excitement about significance of Unique Games Conjecture proof for complexity theory
  • ↑Highlighting breadth of impact across major open problems including Millennium Prize problems
  • ~Existential concern about AI surpassing humans in meaningful intellectual work
  • ↑Framing the progress as beauty to appreciate rather than competition to fear