Published on September 01, 2026 at 18:01 CEST (UTC+2)
AnkiDroid: Google Play no longer allowing Open Collective donation link (541 points by hexa555)
AnkiDroid (Google Play donation link rejection)
AnkiDroid, the popular open-source flashcard app, is facing removal from Google Play after Google rejected updates because the app links to an Open Collective donation page. Google requires proof of tax-exempt status for donation links, and AnkiDroid is seeking clarification on whether an IRS 501(c)(6) determination satisfies the policy. The project is asking the community to amplify the issue rather than contact Google support directly. Unless resolved by September 11, the app will be removed worldwide, highlighting the fragility of open-source projects relying on platform policies.
Ask HN: Who is hiring? (September 2026) (48 points by whoishiring)
Ask HN: Who is hiring? (September 2026)
This is the monthly Hacker News hiring thread where companies post job openings with location, remote policy, and a description of what they do. The preview shows Stream, a company building real-time video, chat, moderation, and newsfeed APIs in Go, hiring across Amsterdam, Skopje, Boulder, Toronto, and remote. Stream's customer list includes major apps like Strava, Nextdoor, Patreon, and Midjourney, indicating how AI-adjacent products rely on real-time infrastructure.
Io_uring Without Readahead (56 points by porridgeraisin)
io_uring Without Readahead
Fernando Simões investigates a pull request that implements application-level readahead for Turso's io_uring database backend. Because the io_uring backend opens files with O_DIRECT, the kernel's readahead is disabled, so each read page waits for the previous one, creating an I/O bottleneck. Adding a 32-page readahead window submits many reads in flight simultaneously, dramatically improving performance on TPC-H benchmark queries. The post is a deep technical dive into how low-level I/O request merging and concurrency shape database performance.
44% on ARC-AGI-1 in 67 cents (248 points by porridgeraisin)
44% on ARC-AGI-1 in 67 cents
Mithil Vakde trained a small transformer from scratch in 1.5 hours on an NVIDIA 5090 for just 67 cents, achieving 44% on the ARC-AGI-1 public evaluation—matching or beating many large language models. The work is an explicit attempt to explore the limits of sample efficiency in transformers and to reduce research iteration costs. The author positions ARC-AGI as a strong benchmark because it is data-sparse, meta-learning-heavy, and unsaturated for data efficiency. The project is open source and already spans three blog posts, with the previous results attracting attention from prominent AI researchers.
American Airlines' Legendary Mechanic Passes Away at 100 After 80-Year Career (254 points by NaOH)
American Airlines' Legendary Mechanic Passes Away at 100
Azriel "Al" Blackman, an American Airlines mechanic, passed away at age 100 after a career spanning more than 80 years in aviation maintenance. His death was announced by American Airlines and the aircraft maintenance community on social media, with many celebrating his record-breaking tenure. The story is a human-interest tribute to an industry legend whose life paralleled the evolution of commercial aviation.
Ambient CSS v3 – Blender meets CSS (5 points by kikkupico)
Ambient CSS v3 – Blender meets CSS
Ambient CSS is described as a physics-based lighting system for CSS, bringing Blender-like material and lighting concepts into web styling. The project appears to enable ambient occlusion, dynamic shading, and physical lighting effects in CSS without complex JavaScript or WebGL. The post has very low engagement (5 points), suggesting it is a niche experimental project.
Fastpotify (671 points by nreece)
Fastpotify
Fastpotify is a lightweight, native Spotify client built in Rust with egui and librespot, distributed under the MIT license. It starts in under a second, uses only 100–250 MB of RAM, and supports local playback at up to 320 kbps, Spotify Connect controls, library browsing, and podcast playback. Other features include themeable UI, a Winamp-style mini player with spectrum analyzer, a MilkDrop visualizer, and MPRIS media controls on Linux. It targets Linux, macOS, and Windows as an alternative to the heavy official Electron-based client.
EFF to Governor Newsom: Veto California's AB 1709 (55 points by mdp2021)
EFF to Governor Newsom: Veto California's AB 1709
The EFF is urging California Governor Gavin Newsom to veto Assembly Bill 1709, which would effectively ban social media use for people under 16. The EFF argues that while the bill is well-intentioned, it would cut off young people—especially vulnerable and marginalized youth—from essential information and supportive online communities they cannot access offline. The organization frames the law as deeply flawed despite its protective intent.
We Are Rebuilding Monica (27 points by robinhood)
We Are Rebuilding Monica
Monica, an open-source personal CRM, is being rebuilt from scratch, with version 3 expected before the end of 2026. Founder Regis Freyd started the project almost ten years ago because he forgot details about friends and family, and existing professional CRMs were designed for sales rather than personal relationships. The rebuild series, "Building Monica," will document the problems, decisions, and lessons learned during the rewrite. The project remains open source.
Introducing Ad Blocker for Firefox on iOS (73 points by HieronymusBosch)
Introducing Ad Blocker for Firefox on iOS
Mozilla has added a built-in Ad Blocker to Firefox on iOS, using Apple's WebKit Content Blocker technology and the EasyList filter list. It blocks many third-party ads and ad-related trackers before they load, and it is off by default, letting users enable it via Settings > Browsing > Ad Blocker. It won't block first-party ads or ads in search results, and it complements Firefox's existing Enhanced Tracking Protection. This gives iOS users a simpler, no-extension alternative for reducing clutter.
1. The rise of efficiency-first AI research
- The trend: The ARC-AGI result—44% accuracy for 67 cents in 1.5 hours on a consumer GPU—shows a growing movement toward sample efficiency, small models, and low-cost experimentation rather than brute-force scaling.
- Why it matters: It challenges the assumption that large models and massive compute are prerequisites for strong reasoning performance. Efficiency metrics (cost, training time, data used) are becoming as important as benchmark accuracy, democratizing who can meaningfully contribute to AI research.
- Implications: Teams should report cost-per-benchmark-point and energy usage alongside accuracy. Low-cost, tiny-model baselines can serve as strong references for evaluating larger systems, and researchers should seriously explore distilled, data-efficient training pipelines.
2. Open-source sustainability is now an AI/ML ecosystem risk
- The trend: AnkiDroid's Google Play conflict and Monica's full rebuild both illustrate how open-source projects face existential pressure from platform policies, funding instability, and maintainer burnout. Much of the AI/ML stack—PyTorch, Hugging Face, vLLM, Llama.cpp, and countless datasets—depends on this same fragile ecosystem.
- Why it matters: If open-source infrastructure cannot secure sustainable funding or navigate app-store/corporate policies, the AI community loses critical tools, and consolidation around a few corporate platforms accelerates.
- Implications: AI organizations should invest in foundations, donation matching, and governance models that shield core tools from single-company or single-platform risk. Projects should proactively clarify their legal status (e.g., 501(c)(3)/501(c)(6) tax-exempt designation) before platforms demand it.
3. Low-level systems engineering remains a hidden AI bottleneck
- The trend: The io_uring readahead analysis shows how application-level I/O management—when the kernel can't help—produces dramatic database performance improvements. AI workloads constantly scan huge datasets, so storage, memory, and database layers are often the real limit on training and inference throughput.
- Why it matters: Model architecture gets the headlines, but data loading and serving pipelines determine real-world cost and latency. As AI applications grow, the demand for engineers who understand io_uring, O_DIRECT, batching, prefetching, and request merging will outpace demand for pure model researchers.
- Implications: AI teams should profile their data pipelines before buying more GPUs; small I/O optimizations (like readahead windows) can match the ROI of model tuning. Systems skills are a strategic advantage for AI engineers, not a legacy concern.
4. AI hiring is shifting from model training to infrastructure serving
- The trend: The September "Who is hiring" thread shows companies building the real-time plumbing for AI-powered products—Stream's Go-based APIs serve feeds, chat, moderation, and video for AI-native firms like Midjourney. Meanwhile, native clients like Fastpotify demonstrate demand for efficient, local-first frontends to AI-adjacent services.
- Why it matters: The AI job market is broadening beyond ML researchers into systems engineers, Rust/Go developers, and infrastructure specialists who can ship reliable, low-latency products. Frontier model research is concentrated in a few labs; the bulk of AI employment is now in serving, integration, and tooling.
- Implications: Engineers should build skills in languages like Rust and Go, real-time streaming, vector databases, and edge deployment. Companies should structure hiring around productized AI systems—inference serving, moderation, personalization—not just model development.
5. Regulation is becoming a first-class design constraint for AI systems
- The trend: California's AB 1709, opposed by the EFF, represents a wave of legislation targeting social media and algorithmic platforms that affect young people. Future AI systems—especially recommendation engines, conversational agents, and content moderation—will increasingly be forced to comply with age-based, safety-focused rules.
- Why it matters: AI products that ignore regulatory trajectories risk sudden bans, fines, or forced redesigns. The EFF's argument that broad restrictions harm vulnerable users also highlights the tension between safety and access, forcing designers to make nuanced trade-offs rather than binary block/allow decisions.
- Implications: AI developers should build age-awareness, transparency, and harm-mitigation into system architecture from day one, rather than patching compliance afterward. Expect standards for algorithmic impact assessments and design choices that preserve access to beneficial communities while reducing harm.
6. Local-first and privacy-preserving software is converging with AI
- The trend: Fastpotify's native, lightweight design and Firefox's on-device Ad Blocker reflect a broader push toward software that minimizes data collection and runs locally. In AI/ML, this manifests as on-device inference, federated learning, and private model serving.
- Why it matters: Privacy is no longer just a compliance checkbox—it is a product differentiator that reduces latency, cuts server costs, and builds user trust. Regulatory pressure on data collection is intensifying, making local-first approaches more attractive for AI applications.
- Implications: AI product teams should evaluate whether inference can run on-device (via quantized models, WebKit content blockers, or local neural engines) at acceptable quality. Open-source, transparent clients will be well-positioned in markets where users are increasingly privacy-conscious.
7. Benchmarks are deceiving: small models can outperform large ones at specific tasks
- The trend: The ARC-AGI blog explicitly notes that a small transformer matches many LLMs on a benchmark that the author argues is "unsaturated" for data efficiency. This echoes a broader pattern: on meta-learning, puzzle-solving, or skill-acquisition tasks, careful small models can beat generic large ones.
- Why it matters: If benchmark leadership is overvalued, organizations may waste enormous compute chasing leaderboard positions that don't reflect practical capabilities. Data efficiency—how quickly a model learns a new rule across tasks—is a more meaningful goal for AGI progress than raw parameter count.
- Implications: Evaluation should include data-efficiency curves, cost-of-training, and task-transfer metrics, not just single-accuracy numbers. Researchers and practitioners should treat small models as legitimate baselines and competitors, and funders should reward reproducible, cheap experimentation rather than only GPU-guzzling scale.
Analysis generated by deepseek-reasoner