Sep 1Wednesday, September 2, 2026 · all daysSep 3 · today »
1.Claude Fable 5.1 and Claude Mythos 5.1(anthropic.com)
1404 points by denysvitali 8 days ago | 1372 comments | permalink
tl;dr: Anthropic released Claude Fable 5.1 and Mythos 5.1, claiming state-of-the-art performance on coding, agentic, and scientific research benchmarks, with Mythos 5.1 restricted to vetted cybersecurity and life sciences users. Notable results include designing protein binders with ~50% hit rates, generating a higher-resolution Venus elevation map, and speeding up open-source bio ML models up to 2.5x. Pricing drops ~25% (up to 45% for agentic workloads) via 75%-cheaper cache reads, and a new Enterprise Frontier Safeguards system offers zero-data-retention privacy while keeping misuse protections.
HN Discussion:
  • Insider praise for improved writing style and upcoming scientific breakthroughs
  • Price reduction reveals weak market demand and benchmark gains are marginal outside one test
  • Anthropic is over-hyping, nerfing models, and hiding thought traces rather than delivering
  • ~Token budgets and predictable quotas matter more than incremental model improvements
  • Frontier releases are boring; cheaper fast models like DeepSeek/GLM are more exciting
2.How accurate have Ed Zitron's AI skeptic predictions been?(danluu.com)
864 points by jatins 8 days ago | 1035 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Article correctly evaluates Ed's literal predictions rather than readers' own projections
  • Ed is directionally right about AI revenue/sustainability concerns despite flawed specifics
  • Critique misinterprets 'dying' too literally, missing Ed's rot-economy framing
  • Ed misuses numbers and cynically distorts technical nuances to fit his narrative
  • Punditry incentives explain Ed's sloppy but audience-aligned commentary
3.FBI Probes Service Selling 153M+ Drivers Licenses(krebsonsecurity.com)
413 points by tatersolid 8 days ago | 276 comments | permalink
tl;dr: A dark web service called Nexus was selling scans of 153M+ US and Canadian drivers licenses, including those of high-ranking officials like Defense Secretary Pete Hegseth. KrebsOnSecurity traced the source to Louisiana-based idscan.net, an identity verification vendor used by Hertz, Planet13 dispensaries, and other major brands, based on timestamps matching victims' car rentals and dispensary visits. The FBI's New Orleans field office has opened an investigation, and the Nexus site went offline shortly after Krebs published the story.
HN Discussion:
  • US should adopt cryptographic identity systems like public key infrastructure instead of ID scans
  • Companies should minimize data retention and delete verification data after use
  • Strict liability and financial penalties would force companies to secure and minimize data
  • Current ID verification processes are security theater that harm regular users while attackers bypass them
  • US lacks European-style regulations due to regulatory capture and corporate influence
4.Introducing Ad Blocker for Firefox on iOS(blog.mozilla.org)
571 points by HieronymusBosch 9 days ago | 185 comments | permalink
tl;dr: Firefox on iOS now includes a built-in Ad Blocker, off by default, that uses Apple's WebKit Content Blocker API and the EasyList filter to block third-party ads and trackers. It won't block first-party ads, search result ads, or Firefox's own sponsored shortcuts, and can be enabled via Settings > Browsing > Ad Blocker. Mozilla built it directly into the browser because iOS doesn't support extensions the way Desktop and Android do.
HN Discussion:
  • Search engine ads should also be blocked since they often contain scams
  • Frustration with slow rollout making the feature unavailable despite announcement
  • ~Mozilla avoids blocking YouTube/search ads due to Google revenue dependence, which is understandable
  • Ad blocking should be default and standard across all Firefox platforms
  • Mozilla is too late with a minimal ad-blocker compared to existing iOS alternatives
5.My local model setup on an M4 Pro Mac Mini(lws.io)
323 points by raybb 8 days ago | 192 comments | permalink
tl;dr: The author runs a local LLM setup on a 48GB M4 Pro Mac mini using Qwen3.6-35B-A3B and Gemma-4-E4B (both 4-bit quantized) via the oMLX inference server, with Tailscale connecting his iPhone and MacBook to access it through clients like Hermes, Apollo, and Raycast. He argues MoE models are key for consumer hardware since only active parameters (3B of 35B) matter for inference speed, achieving ~34 tok/s generation. The motivation: avoiding cloud API pricing volatility, privacy risks, and geopolitical restrictions, while handling ~80% of daily tasks locally at zero marginal cost.
HN Discussion:
  • ~Local LLM setup is still too complex; waiting for turnkey solutions
  • Questions the cost motivation since simple AI tasks are free in the cloud
  • ~Privacy, not cost, should be the primary motivation for local models
  • Sharing performance benchmarks that validate MoE model efficiency claims
  • Prefill latency makes Mac local LLMs inferior to dedicated hardware like DGX Spark
6.Ambient CSS v3 – Blender meets CSS(ambientcss.vercel.app)
298 points by kikkupico 9 days ago | 97 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • Implementation feels buggy, laggy, or inconsistent with confusing behavior across components
  • Accessibility and semantics are poor since native HTML controls are replaced with custom JS widgets
  • Concept and aesthetic are appealing, reminiscent of VST GUIs or skeuomorphic Web 2.0 design
  • ~Effects don't work well on mobile or feel unusable on touch devices
  • The Blender/3D framing is questionable — elevation and textures aren't truly 3D as implied
7.Movie Scene Map – 13,312 films, series, games, anime and manga(moviescenemap.com)
326 points by Flightmussy 9 days ago | 46 comments | permalink
tl;dr: Movie Scene Map is a free interactive atlas of 15,565 real filming locations across 166 countries, covering 9,287 films/series plus 2,153 games, 407 anime, and 365 manga (placed by narrative setting). Data is sourced from Wikidata statements and Wikipedia articles/categories rather than scraping, and the full dataset is downloadable as CC0 GeoJSON/CSV with a read-only MCP endpoint for AI assistants. Coverage is uneven and curated—missing titles can be added upstream via Wikidata edits.
HN Discussion:
  • Enthusiastic praise for the concept, UX, and practical travel utility
  • ~Feature requests for filters, better linking, or richer scene metadata
  • Reports of data inaccuracies or incorrect filming location entries
  • ~Notes about missing coverage and suggestions to expand via crowdsourcing or partnerships
  • Appreciation for non-corporate, specialist internet projects like this
8.True Rate of Unemployment(lisep.org)
301 points by ptrhvns 8 days ago | 339 comments | permalink
tl;dr: The True Rate of Unemployment (TRU), developed by LISEP using BLS data, measures the share of the U.S. labor force that is jobless, involuntarily part-time, or earning less than $26,000 annually (2025 dollars). It aims to give policymakers a more accurate picture of Americans' financial well-being than the official unemployment rate.
HN Discussion:
  • Official statistics are misleading and alternative measures like TRU are needed for accuracy
  • The $26,000 living wage threshold is too low and understates the problem of insufficient wages
  • TRU's methodology is flawed, producing implausible results and conflating low-wage work with unemployment
  • The headline unemployment rate already tracks TRU well, making the new metric redundant
  • ~TRU is better understood as political framing to quantify livable-income difficulty rather than a statistical indictment
9.The ChatGPT/Codex app bundles a full copy of LibreOffice(simonwillison.net)
480 points by timpera 8 days ago | 240 comments | permalink
tl;dr: The OpenAI Codex/ChatGPT desktop app stashes 1.7GB in `~/.cache/codex-primary-runtime`, bundling full copies of Python, Node.js, Poppler, git, and the entire LibreOffice suite. Codex uses these binaries via "skills" defined in its plugins directory, apparently to handle document processing tasks locally.
HN Discussion:
  • OpenAI should give back to LibreOffice through donations or contributions
  • Bundling LibreOffice is reasonable since it's reliable for handling legacy office file formats
  • The Codex app is poorly designed and the bloated dependencies reflect that
  • This may be a license (MPL 2.0) violation given missing attribution
  • Questioning whether the bundling is actually problematic or just normal app behavior
10.The creator of Jujutsu has joined ERSC(ersc.io)
260 points by steveklabnik 8 days ago | 255 comments | permalink
tl;dr: Martin von Zweigbergk, creator of the Jujutsu (jj) version control system, has left Google to become CTO of East River Source Control, a startup building next-gen VCS tooling for the AI era. He'll continue maintaining Jujutsu as open source, but argues the Git-based server/storage layer needs to be rethought—work better suited to a company than an OSS project. ERSC, backed by Amplify Partners, is launching its Storage product in private beta later this month.
HN Discussion:
  • Enthusiastic jj users praising its UX, mental model, and undo capabilities
  • Skeptics questioning jj's value proposition since git already works fine for their workflow
  • Concern that Martin joining a startup could be bad news for the open-source jj project's future
  • Users noting jj's current scaling limitations and need for better backend infrastructure
  • Seeking clarification on what ERSC actually builds (alternative backend for jj)
11.Ask HN: Who is hiring? (September 2026)
236 points by whoishiring 9 days ago | 283 comments | permalink
tl;dr: Monthly Hacker News thread for companies to post job openings, requiring posters to specify location/remote status and be directly involved in hiring (no recruiters or job boards). Third-party search tools are linked for filtering listings, and a companion "Who wants to be hired?" thread is referenced for job seekers.
HN Discussion:
  • AI-focused startups posting remote engineering roles following the thread format
  • European research and hybrid roles offering visa sponsorship and relocation
  • Robotics and physical AI companies recruiting onsite technical talent
  • Open source infrastructure companies hiring engineers across Europe
  • Mission-driven organizations recruiting for security and healthcare access roles
12.Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s(github.com)
226 points by carloslfu 9 days ago | 107 comments | permalink
tl;dr: Slotstream is a single Swift binary that runs the 104GB Qwen3.8-Flash-Next MoE model on Apple Silicon Macs with far less RAM than the model size, streaming experts from SSD into a fixed cache pool shared across layers. On a 48GB M5 Pro it hits ~12 tok/s warm decode with ~32GB peak memory, and exposes Ollama/OpenAI-compatible APIs on port 11434. It works because MLX can't partially materialize mmap'd tensors (top-10 expert gather touches all 512), so it uses pread into managed slots instead.
HN Discussion:
  • ~README needs cleanup and better introduction for new users with zero context
  • Sharing personal experiences running similar models locally on Mac hardware with different tools
  • Enthusiasm for democratizing AI by running large models on modest hardware
  • Skepticism about performance claims on low-memory hardware configurations
  • Questions about practical aspects like context window, energy use, and model capability comparisons
13.Atlas: A World Model for Spatial Intelligence(worldlabs.ai)
261 points by johnsutor 8 days ago | 59 comments | permalink
tl;dr: World Labs unveiled Atlas, a multimodal autoregressive diffusion transformer that natively handles text, images, video, and 3D depth maps within a shared spatial context. It performs camera-controlled video generation (up to 1 minute at 1440p), sparse-view 3D reconstruction (outputting point clouds or Gaussian splats), space-time reframing, and robotics Real-to-Sim workflows, reportedly outperforming specialized models on camera-following and 3D reconstruction benchmarks. Atlas is entering early access with select partners and will power future versions of World Labs' Marble product.
HN Discussion:
  • Excitement about novel applications like game prototyping and 3D reconstruction from sparse images
  • ~Article underplays interesting uses like extracting semantic latent knowledge for robotics
  • Skepticism about consistency and limitations like hallucinations and frozen-time motion
  • Questions about practical capabilities like exact dimensions, characters, and splat output costs
  • Confusion over the vague and overused term 'world model'
14.AnkiDroid: Google Play no longer allowing Open Collective donation link(github.com)
917 points by hexa555 9 days ago | 279 comments | permalink
tl;dr: Google Play is threatening to remove AnkiDroid (10M+ installs) from the store on September 11 unless it removes its Open Collective donation link, arguing donations must go to a "tax-exempt organization." AnkiDroid's fiscal host, Open Source Collective, provided an IRS determination letter confirming 501(c)(6) tax-exempt status, but Google support has repeatedly rejected it without explaining why 501(c)(6) doesn't qualify (the policy doesn't specify 501(c)(3)). The team is removing the donation link under protest and asking for help escalating to someone inside Google.
HN Discussion:
  • App store monopolies are harmful and this incident proves they need regulation
  • Google is technically correct because 501(c)(6) donations aren't tax-deductible to the donor
  • Google's opaque AI-driven support process is broken and hurts developers
  • ~PWAs or alternative distribution methods would sidestep this problem entirely
  • Expressing gratitude for AnkiDroid and intent to donate
15.Play Store blocks AuroraStore, hurting GrapheneOS users(gitlab.com)
547 points by erikvanoosten 9 days ago | 253 comments | permalink
tl;dr: Aurora Store users with anonymous accounts are seeing "Server busy, please try again later" errors on all app installs, apparently due to Google blocking the shared anonymous accounts Aurora relies on to proxy Play Store downloads. The issue reproduces across stable and nightly builds (including 4.8.4 on CalyxOS/Android 16), and persists regardless of VPN, cache clearing, or account refresh—effectively breaking Play Store access for GrapheneOS and other de-Googled users who don't have their own Google accounts.
HN Discussion:
  • GrapheneOS recommends sandboxed Play Store over Aurora, so impact is minimal
  • Aurora is preferred for privacy and UX reasons, so this genuinely hurts de-Googled users
  • The issue affects many de-Googled ecosystems beyond just GrapheneOS (LineageOS, /e/OS, CalyxOS)
  • The title editorializes and the actual cause isn't confirmed to be intentional Google blocking
  • Aurora fills a critical gap since Android lacks official app installation without a Google account
16.I trained a small transformer in 1.5hrs and it beats many LLMs(mvakde.github.io)
651 points by porridgeraisin 9 days ago | 162 comments | permalink
tl;dr: A researcher trained a small transformer from scratch on a 5090 GPU in 1.5 hours for $0.67, scoring 45% on ARC-1 and 7% on ARC-2, matching TRM/HRM without recursion. Key improvements over the prior version include modern architecture (SwiGLU, RMSNorm), NorMuon optimizer, 3D RoPE with per-task embeddings, and supervised training on output tokens only. The author argues recursion isn't necessary, criticizes reliance on synthetic data and augmentations in other ARC approaches, and believes 65% is achievable within the transformer framework.
HN Discussion:
  • Author clarifies the work is a small transformer, not an LLM, demonstrating complex problems can be tackled without LLMs
  • The improvements are 'squeezing the lemon' tactics that should come after achieving near-SoTA with the novel method
  • Requests clarification on the training-on-eval defense for less informed readers
  • Praises author's achievement and suggests it will lead to career opportunities
  • Raises concern about input leakage during pretraining as a flaw in the approach
17.Hang on to Your Firefox(newsonaut.com)
938 points by speckx 8 days ago | 511 comments | permalink
tl;dr: Firefox is the last major browser with an independent engine, and abandoning it over minor complaints (like its presence on X) risks leaving the web dominated by Chrome and its derivatives, with only Safari as a niche holdout. The author argues that critics—particularly on Hacker News and among bloggers switching to Chromium-based alternatives like Vivaldi—are being inconsistent and shortsighted, and that supporting Firefox is essential for preserving browser engine diversity.
HN Discussion:
  • ~Firefox's importance justifies criticism of Mozilla's anti-user decisions, not silence about them
  • Coalition-building means supporting Firefox despite disagreements with Mozilla
  • Firefox is worth using for its superior ad-blocking and unique engine
  • Firefox has real usability problems (RAM, CPU, mobile UX) that make advocacy difficult
  • Other independent engines like Ladybird and Servo offer alternative paths to diversity
18.Apple reveals 'shocking evidence' from ex-employee's MacBook in OpenAI suit(9to5mac.com)
241 points by colinprince 8 days ago | 186 comments | permalink
tl;dr: Apple filed new evidence in its trade secret lawsuit against OpenAI and former engineer Chang Liu, claiming forensic analysis of Liu's MacBook shows he downloaded a confidential Apple circuit schematic and used it in LTspice simulations at OpenAI, allegedly feeding it to an AI agent. Apple also alleges Liu instructed an OpenAI colleague to destroy evidence after learning of Apple's internal investigation. Apple is using the findings to push for expedited discovery and access to a related Mac mini that synced to the laptop via iCloud.
HN Discussion:
  • Summarizes the specific allegations against Liu, reinforcing the article's factual claims
  • Notes the novel legal argument about AI training on trade secrets creating irreversible propagation
  • Raises privacy concerns about iCloud syncing between personal and company devices
  • Compares OpenAI unfavorably to companies like Pepsi that reported IP theft attempts
  • Points out hypocrisy in Apple's trade secret argument given AI companies' fair use stance on training data
19.Tmp.0ut Volume 5(tmpout.sh)
211 points by ghuntley 9 days ago | 42 comments | permalink
tl;dr: Volume 5 of tmp.0ut, an ELF/low-level hacking zine, features 21 articles including an interview with Doug McIlroy, a 57-byte x86-64 Linux ELF analysis, a 440-byte metamorphic ELF-64 virus, deep dives into Linux kernel executable loading, polyglot ELFs, syscall hook detection via side-channels, code virtualization, and fine-grained load-time ASLR. Also included are BGGP6 recap, static kernel patching, and a mixtape.
HN Discussion:
  • Nostalgia for the craft and ingenuity of older computing culture
  • Appreciation for zines as an alternative to modern platforms for deep technical content
  • Enthusiasm for specific content like the McIlroy interview and mixtape
  • ~Corrections and critiques on factual errors and readability issues
  • Technical observation connecting article content to real-world applications like Cloudflare's captcha
20.American Airlines mechanic Azriel “Al” Blackman has died(simpleflying.com)
382 points by NaOH 11 days ago | 163 comments | permalink
tl;dr: Azriel "Al" Blackman, the Guinness World Records holder for the longest career as an airline mechanic, has died at 100, ending an 80-year tenure at American Airlines that began in 1942 when he was 16. Starting on Sikorsky flying boats at LaGuardia's Marine Air Terminal and finishing on the Boeing 777, he was honored in 2017 when American dedicated a 777-200 (N751AN, nicknamed "7BK") to him. He continued working until celebrating his 80th anniversary with the airline in 2022.
HN Discussion:
  • Personal anecdotes about airline mechanics' dedication and the industry's honoring of their service
  • Concern about loss of tacit craftsman knowledge when veterans retire without training apprentices
  • Observation on the name Azriel meaning Angel of Death
  • Reflection on the decline of lifetime employment in favor of consultant/short-term economy
  • Admiration and tribute to Blackman's remarkable 80-year career