Fastpotify — A Native Spotify Client in Rust (772 points)

Fastpotify is an MIT-licensed Spotify client built with Rust, egui, and librespot. It ships as a native binary with no browser engine, starts in under a second, and uses 100–250 MB of RAM. It supports local playback at up to 320 kbps with gapless playback, Spotify Connect for controlling other devices, library browsing, playlist editing, and theming. The nostalgia play: it includes a Winamp 2 skin mini-player and a MilkDrop visualizer via projectM.

The 772-point score reflects genuine frustration with Spotify’s official client, which is an Electron app that routinely eats 500+ MB of RAM and feels sluggish. Fastpotify isn’t trying to replace Spotify’s backend — it’s a better frontend for the same service, built by someone who clearly cares about desktop performance. The librespot foundation means it reverse-engineers Spotify’s protocol, which puts it in the same legal gray area as every other third-party Spotify client. Spotify hasn’t cracked down on librespot-based projects yet, but that’s a policy decision, not a guarantee.

Source: fastpotify.rocks

Google Play has forced AnkiDroid, the open-source Android flashcard app, to remove its Open Collective donation link from the Play Store listing. The issue, filed as a community help request, documents Google’s enforcement of a policy that prohibits soliciting donations through external payment platforms within app listings. AnkiDroid is a free, open-source project that relies on community donations to fund development.

This is the latest in a pattern of Google tightening the screws on open-source projects that don’t funnel money through Google Play’s billing system. The policy technically applies to all apps, but the practical effect falls hardest on projects that can’t afford to integrate Google Play Billing or that philosophically object to giving Google a 15–30% cut of donations. The HN discussion predictably turned into a broader referendum on Google’s control over the Android ecosystem. The real question isn’t whether Google has the right to enforce its store policies — it does — but whether a platform monopoly should be able to dictate how open-source projects fund themselves.

Source: github.com/ankidroid/Anki-Android#21656

Claude Fable 5.1 and Claude Mythos 5.1 (523 points)

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, which are the same underlying model with different safeguard levels. Fable 5.1 is generally available; Mythos 5.1 is restricted to vetted cybersecurity and life sciences professionals through trusted access programs. The headline claims: 25% cheaper for typical workloads (up to 45% for agentic tasks) via reduced cache read pricing, new Enterprise Frontier Safeguards for zero-data-retention customers, and benchmark improvements across coding, reasoning, and scientific research.

The benchmarks are impressive on paper — 52.6% on Terminal-Bench-Science (vs. 24.7% for Fable 5), 55.8% on Terminal-Bench 4.0, 60.9% on Humanity’s Last Exam. The partner testimonials are the usual parade of Fortune 500 names praising the model’s ability to do their jobs. What’s genuinely interesting is the scientific research angle: Mythos 5.1 designed high-affinity protein binders with a 50% hit rate (vs. 10–15% typical), created a new high-resolution elevation map of Venus from 30-year-old Magellan data, and optimized GPU kernels for seven bioinformatics models with up to 2.5x speedups.

The anti-distillation mechanisms are noteworthy — Anthropic is closing off the technique of manually editing conversation context to extract Claude’s thinking traces. This is a direct response to the open-source community’s systematic distillation of frontier models. The Life Sciences Verification Program, developed with the US government, signals that Anthropic is taking the dual-use biology risk seriously enough to gate access behind identity verification. Whether this is genuine safety theater or meaningful risk mitigation depends on how rigorous the vetting actually is.

Source: anthropic.com

I Trained a Small Transformer in 1.5 Hours and It Beats Many LLMs (439 points)

A developer trained a small transformer from scratch on a single RTX 5090 in 1.5 hours that achieves competitive performance on the ARC-AGI benchmark, matching or beating many much larger LLMs. The model scores 44% on ARC-1 public eval and 7% on ARC-2, matching the performance of TRM/HRM (Test-time Training / Hierarchical Reasoning Model) approaches. The code is open-source on GitHub.

The author’s thesis is that sample efficiency is the most important unsolved problem in AI, and ARC-AGI is the ideal testbed because it has very few training samples (1,000 puzzles), requires meta-learning (each puzzle has different rules), and is trivially easy for humans but hard for current AI. The previous version of this work went viral on X, drawing attention from researchers like Lucas Beyer and Jeremy Howard. The key insight isn’t that small models can beat big ones — it’s that the right training methodology on the right benchmark can expose how much of large model performance is brute-force memorization versus genuine generalization. ARC-AGI continues to be the most honest benchmark in AI.

Source: mvakde.github.io

Play Store Blocks AuroraStore, Hurting GrapheneOS Users (365 points)

Google Play Store is blocking AuroraStore, an alternative Android app store that lets users access Play Store apps without Google services, with a persistent “Server busy, please try again later” error. The issue affects anonymous accounts and persists across VPNs, cache clears, and device restarts. AuroraStore is the primary way GrapheneOS users install apps that aren’t available on F-Droid, making this effectively a blockade against privacy-focused Android users.

This looks like a deliberate server-side change by Google rather than a bug. GrapheneOS users — who explicitly chose to de-Google their phones — depend on AuroraStore as a compatibility layer. If Google is intentionally blocking anonymous access to its app catalog, it’s a significant escalation in the platform wars. The timing coincides with Google’s broader push to lock down Android’s sideloading ecosystem. The HN thread is full of people suggesting workarounds, but the fundamental problem is that Google controls the app catalog and can cut off alternative access at any time.

Source: gitlab.com/AuroraOSS/AuroraStore#1566

American Airlines Mechanic Azriel “Al” Blackman Has Died (340 points)

Al Blackman, an American Airlines mechanic who worked for 80 years and held the Guinness World Record for longest career as an airline mechanic, has died at age 100. He started at American Airlines in 1942 at age 16 and never retired, continuing to work on aircraft well into his 90s.

In an era of disposable careers and 18-month job tenures, Blackman’s 80-year run at a single company is almost incomprehensible. He worked on everything from propeller-driven DC-3s to modern widebody jets, witnessing the entire arc of commercial aviation. The story resonated on HN not because of aviation nostalgia but because of the sheer improbability of someone finding work meaningful enough to do for eight decades. American Airlines named a maintenance facility after him while he was still alive — a rare honor for someone who turned wrenches rather than sat in boardrooms.

Source: simpleflying.com

The Restroom Archive (339 points)

The Restroom Archive is an ongoing repository of publicly accessible restrooms. That’s it — it’s a database of where you can find a bathroom when you need one.

The 339-point score suggests HN’s audience has strong feelings about the availability of public restrooms, which tracks — anyone who’s ever been caught short in an unfamiliar city knows the pain. The site fills a genuine gap: Google Maps has restroom data for some businesses, but there’s no dedicated, community-maintained database of public facilities. Whether this can sustain itself as a project depends on whether enough people contribute data. The concept is sound; the execution challenge is coverage.

Source: restroomarchive.com

Throughline

Today’s front page is dominated by two themes: platform control and AI capability jumps. Google appears three times — squeezing AnkiDroid’s donation links, blocking AuroraStore for GrapheneOS users, and generally reminding everyone that owning the platform means owning the rules. Anthropic’s Fable 5.1 release represents the other pole: frontier AI models getting meaningfully cheaper and more capable, with the interesting wrinkle that the most impressive results (protein design, Venus mapping) come from the restricted Mythos variant, not the generally available one. The tiny ARC-AGI transformer is the counterpoint to both — proof that clever methodology on honest benchmarks still matters more than raw scale. And Fastpotify at 772 points is the community’s way of saying: we’ll build our own clients, thanks, if the official ones keep getting worse.