Hacker News Top 10
English Edition

1. Muse Spark Open (675 points by bigwheels)

   Muse Spark Open — Meta AI Research released Muse Spark Open, an open-weight model whose weights can be downloaded and run by others. It is natively multimodal with 816 billion total parameters and 42 billion active parameters, plus a 1-million-token context window. The stated focus is coding, agentic work, long-horizon tasks, and multi-step reasoning. Meta says it is the first model in its class to run on a desktop using two Apple M5 Ultra Mac Studios with Apple's MLX framework. It is also offered through several API platforms and supported by many inference engines.

2. RPKI now protects 70% of internet routes (74 points by emot)

   RPKI now protects 70% of internet routes — The article reports that RPKI now protects 70% of internet routes. RPKI is a system for cryptographically signing routing announcements so networks can verify who is allowed to announce an IP address block. The content preview is not available, so the article's measurement method, timeline, and remaining gaps are not established here.

3. Business accounts will consent to Meta scraping their websites Starting Oct. 30 (94 points by 0in)

   Business accounts will consent to Meta scraping their websites Starting Oct. 30 — The article reports that Meta's new commercial terms, effective October 30, will allow Meta to scrape websites belonging to business accounts, according to legal experts. The URL and title indicate the focus is news sites and business accounts. The content preview is not available, so the exact scope, opt-out options, and whether the scraped data is used for AI training are not confirmed by the supplied text.

4. Any window from your Mac, on your iPhone and iPad (50 points by chanana)

   Sill — Sill is an open-source project that shows any window from a Mac, or the whole desktop, on an iPhone or iPad. It captures the window, encodes it as HEVC video, a video compression format, and sends it to the Sill app on the device. Touch, trackpad, keyboard, and Apple Pencil input are sent back to the Mac. The connection can run on the same network, over USB, through Direct Wireless Connection, or through your own VPN such as Tailscale. There is no account and no server in between.

5. Neural networks are clean (555 points by themgt)

   Neural networks are clean — This Goodfire essay by Nick Cammarata argues that neural networks are more legible than many researchers feared. It claims that the usual evidence for polysemanticity (one neuron responding to many unrelated concepts) and superposition (models packing more concepts than they have neurons) was misread. What looked like arbitrary packing is described instead as a motif called "recursive generalization." The author says we can now understand most neurons in models up to trillions of parameters, but presents this as a working hypothesis rather than settled proof.

6. Apple routed a graphite pipe through the iPhone Duo's hinge to dissipate heat (71 points by giuliomagnifico)

   Apple routed a graphite pipe through the iPhone Duo's hinge to dissipate heat — The article reports that Apple routed a graphite pipe through the hinge of the iPhone Duo to dissipate heat. The content preview is not available, so the article's details about thermal performance, components, or design tradeoffs are not included here. The title indicates the focus is hardware cooling in a phone with a hinge.

7. What I self-host (330 points by BrunoBernardino)

   What I self-host — The author describes a long history of self-hosting, from a Linux web server under a university bed to MythTV and later XBMC/Kodi. After leaving India and not wanting to pay Western streaming prices, they replaced streaming services with self-hosted Jellyfin, a media server. They wanted more control over their son's viewing options and disliked recommendation and autoplay engagement features and constant redesigns. Over time, self-hosting expanded to other services, reducing cloud dependence and providing stability.

8. Wayland desktop for Linux with skeuomorphic design (545 points by FinalDestiny)

   DEsk — DEsk is a Wayland desktop environment for Linux with a skeuomorphic design, meaning it imitates physical objects. The GitHub repository describes it as v0.1.2 beta, with v0.1 as the first public release and an installer for Fedora. The name came from the Reddit community, and its commands include desk, desk-session, and desk-settings. It is a desktop environment built for the Wayland display protocol.

9. Non-Determinism Doesn't Matter (21 points by sklarsa)

   Non-Determinism Doesn't Matter — The author initially objected to using large language models for software development because the same prompt can produce different output on different runs. After using AI as their primary way to write code, they changed their view: the final result must be correct, but the path to it does not have to be deterministic. They argue that LLMs and their harnesses are actors in a system, while code is just an output. The post is an opinion and experience report, not a controlled study or benchmark.

10. AdFection: Targeted ads turned into a pipeline for zero-click mercenary spyware (409 points by jcomp)

    AdFection — Citizen Lab reports that devices at the news network Al Jazeera were infected with mercenary spyware from Candiru, now rebranded as Integrity Labs, through targeted online advertisements. The researchers call the attack AdFection and describe it as the first public technical confirmation of zero-click spyware, meaning spyware that does not require a user click, delivered via targeted advertising. The malicious ads were purchased by Discover-Tech, an Israeli company that presents itself as a legitimate advertiser but allegedly facilitates spyware partners. Discover-Tech also operated Controlium, a malicious ad targeting platform that ran at billion-bid scale.

AI/ML Insights & Trends

1. Open-weight models now come with real deployment paths

   • Trend/insight: Article 1, "Muse Spark Open," shows a large open-weight model released with downloadable weights, API providers, and support in many inference engines, which are programs that run trained models. The model has 816 billion total parameters and 42 billion active parameters; active parameters are the part used for each token, so the model can be very large to store but cheaper to run than its total size suggests. This is more than a research release because it gives teams several ways to run or serve the model.
   • Why it matters for AI development: Developers can choose between paying an API provider and hosting the model themselves. That choice affects cost, latency, data privacy, and how much control you have over customization.
   • Implications and practical tips: Check the license, supported inference engines, and hardware needs before committing. Self-hosting an 816B-parameter model is not a small task, but using an API provider can reduce operations work.

2. Long-context agents are being aimed at software work

   • Trend/insight: Article 1 says Muse Spark Open is built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning, with a 1M-token context window. A 1M-token context window lets the model process a very large amount of text or code at once, roughly a large codebase or many documents. This suggests vendors are optimizing models for multi-step workflows, not just single-turn chat.
   • Why it matters for AI development: Such models can support assistants that read large codebases, plan changes, call tools, and revise work over longer tasks. That can reduce manual context management for developers.
   • Implications and practical tips: Long context can increase memory, cost, and latency, so test real workloads. Give agents limited tool permissions, log actions, and require review before code is merged or deployed.

3. Local inference is improving, but serious models still need serious hardware

   • Trend/insight: Article 1 says Muse Spark Open can run on a desktop using two Apple M5 Ultra Mac Studios with Apple's MLX framework. That is a notable step for local use, but "desktop" here means multiple high-end machines, not a typical laptop. Article 7, "What I self-host," shows a related preference for running services on your own hardware for control and stability, though it does not discuss AI models.
   • Why it matters for AI development: Local inference can help with privacy, offline use, and avoiding API costs, but it shifts cost and complexity to hardware and operations. Not every application needs or can afford a large local model.
   • Implications and practical tips: For privacy-sensitive work, consider smaller local models or quantized versions first. If you need a very large model, compare API pricing and latency against the cost, power, and maintenance of local hardware.

4. Interpretability may be more practical than many developers assume

   • Trend/insight: Article 5, "Neural networks are clean," argues that the evidence for polysemanticity and superposition was misread and that neurons and circuits can be more legible than feared. It claims most neurons in models up to trillions of parameters can be understood, describing the pattern as "recursive generalization." The author presents this as a working hypothesis, not a proven result.
   • Why it matters for AI development: If models are easier to inspect, developers may be able to find the internal causes of errors, unwanted behavior, or unsafe outputs. That could improve debugging and safety work.
   • Implications and practical tips: Treat interpretability as a promising research direction, not a production guarantee. Use behavioral tests, logging, and monitoring now, and watch for tooling that makes neuron and circuit inspection usable outside research labs.

5. Non-determinism changes testing, not the need for correct results

   • Trend/insight: Article 9, "Non-Determinism Doesn't Matter," describes an author who stopped treating different LLM outputs as a fatal flaw. The argument is that the final result must be exact or correct, but the path to it does not need to be deterministic. The author frames LLMs and their harnesses as actors in a system, with code as one output.
   • Why it matters for AI development: If you generate code or actions with an LLM, you cannot rely on the same output every run. You need verification that checks behavior, not just text equality.
   • Implications and practical tips: Run generated code in a sandbox, use tests, type checks, linters, and human review, and assert on outcomes rather than exact generated text. Keep retries and fallbacks for cases where the model produces invalid output.

6. AI data collection is increasingly governed by ordinary commercial terms

   • Trend/insight: Article 3 reports that Meta's new commercial terms, effective October 30, will let Meta scrape websites belonging to business accounts, according to legal experts. The preview is unavailable, so it is not confirmed here whether the scraped data feeds AI training. The key point is that data collection rights can be set through platform agreements rather than separate opt-in datasets.
   • Why it matters for AI development: If your site, app, or business uses a platform, its terms may permit scraping of content you consider yours. That affects privacy, competitive exposure, and the data you can control or license.
   • Implications and practical tips: Review the terms of platforms you rely on, keep copies of agreements, and check what content is exposed by business accounts. If control matters, consider data licensing, self-hosting, or technical measures such as robots.txt where they are respected.

7. Security of data and ad pipelines is a practical AI concern

   • Trend/insight: Article 10, "AdFection," documents targeted ads being used to deliver zero-click spyware to Al Jazeera devices through a malicious ad targeting platform. The article does not discuss AI models directly. However, AI applications often ingest web data, user data, or ad-tech signals, so the same untrusted data pipelines can become an attack surface.
   • Why it matters for AI development: If an AI system treats ads, scraped pages, or third-party data as trusted input, it can inherit malware, privacy, and data-poisoning risks. This is especially relevant for agents that browse, retrieve, or act on web content.
   • Implications and practical tips: Isolate untrusted content, sanitize inputs, limit tool permissions, and audit third-party SDKs and data sources. Do not let an agent execute actions from an ad or web page without confirmation and sandboxing.

Analysis generated by deepseek-flash