Hacker News Top 10
English Edition

1. Big Fish: Exploiting a 37-Year-Old Kernel Bug in OpenBSD (50 points by rs_rs_rs_rs_rs)

   This article describes a security vulnerability in OpenBSD that traces back about 37 years to original BSD code. A shared memory attachment counter was stored as a signed 16-bit integer, so repeatedly forking a process could push it past 32,767 and wrap to a negative value. That integer overflow could trigger a use-after-free, where memory is used after it has been freed, through a race condition. Combined with a separate out-of-bounds write in the kernel, it allowed a local user to gain root privileges. The researchers achieved local root even with KARL, SMAP, and SMEP enabled and without an information-disclosure bug, and fixes were released in OpenBSD errata 026 and 018.

2. Apple ML Research: Muse Spark Open with MLX on Mac Studio (61 points by antipaul)

   Apple describes running a nearly trillion-parameter model, Muse Spark Open, locally with MLX on Apple silicon Macs. MLX is Apple’s open-source array framework that uses unified memory, where the CPU and GPU share the same memory pool. On a cluster of four Mac Studios connected with RDMA, a way to move data between memory and network adapters with less CPU work, over Thunderbolt, the model generated over 36 tokens per second at an 8,192-token context. Tokens are pieces of text the model produces, and that speed is above average human reading speed. The article presents this as a private way for researchers and developers to run inference and fine-tuning experiments on their own hardware, though the preview does not give full details on setup complexity, power use, or cost.

3. I Love Software Rendering (171 points by speckx)

   The author argues that software rendering, where the CPU draws every pixel, still has value and can look better than some early 3D-accelerated versions. They describe playing Quake with its DOS software renderer at 800×600 and a locked 60 frames per second on a Dell OptiPlex 380. The appeal is a sharp, gritty look: textures are sampled pixel by pixel, texels stay square, and early bilinear filtering is avoided. The article notes that modern fast CPUs allow old software renderers to run at resolutions that were not practical in the 1990s. This is a personal, aesthetic argument rather than a claim that software rendering is generally superior.

4. What on Earth Is Dekton? (18 points by wawayanda)

   This article begins with the author’s frustration over a pricey induction cooktop that required a separate, expensive cooking surface made of Dekton. Dekton is an artificial material invented by the Spanish company Cosentino in 2013, and the article says it is highly resistant to heat, pressure, scratches, shattering, and stains. The author asks what Dekton actually is and explains that the answer has at least two parts. The preview says the piece is also about learning to love porcelain, but it cuts off before the full argument is presented. So the supplied text establishes the material’s claimed properties and cost context, not the complete conclusion.

5. Show HN: Gemini 4 Argon plays StarCraft (48 points by cayenne)

   StarSkirmish Bench v0.2 is a benchmark where large language models write StarCraft: Brood War bots in C++ under a one-hour time limit. Each bot is then tested against other LLM-written bots and established human-written bots, with scores based on average win rate over five runs. The benchmark is meant to test long-horizon reasoning and agentic coding, because the model must use practice games to improve its bot. Version 0.2 includes flagship models from OpenAI, Anthropic, Google, and Meta running in each lab’s first-party command-line harness, and the preview reports GPT-6 Astra first, followed by GPT-6.1 Sol, with several other frontier models close behind. The preview also compares scores with API cost, but does not provide a full methodology or limitations section.

6. How to Formalize All of Mathematics (31 points by practal)

   This guest post from the Xena project discusses recent claims that AI makes it possible to formalize all of mathematics. The authors say AI has become remarkably good at proving theorems, both in formal systems like Lean, a proof assistant that checks mathematical proofs, and in ordinary mathematical writing. They support the idea of large-scale formalization if it is done well, but they are concerned about calls to rush the process. Their central claim is that a coordinated, careful effort could expand mathematical discovery, while a poor effort would waste resources and possibly cause lasting damage. The post is a response to excitement around an OpenAI release and is written by mathematicians with experience in formalization.

7. Is Liquid Metal Killing Our PCs? Scientific Analysis and a 2-Year Test [video] (37 points by zdw)

   This is a YouTube video titled “Is Liquid Metal Killing Our PCs? Scientific Analysis and a 2-Year Test.” The supplied preview contains only YouTube boilerplate and no transcript, description, or detailed findings. The title suggests the video examines whether liquid metal thermal interface material damages computers, using analysis and a long-term test. Because the actual content is not available in the preview, I cannot summarize its results or limitations. Any specific conclusion about liquid metal would require watching the video.

8. Gyroscopes Got So Small [video] (42 points by surprisetalk)

   This is a YouTube video titled “Gyroscopes Got So Small.” The supplied preview contains only standard YouTube metadata and no transcript or detailed description. The title suggests the video explains how gyroscopes became much smaller, likely covering miniaturization and modern sensor design. Without the video content, the preview does not establish the specific technologies, claims, or limitations discussed. I therefore cannot provide a substantive summary beyond the title.

9. Stolen Renoir paintings found hidden, and damaged, behind cinder block wall (15 points by ohjeez)

   This news article reports that two stolen Renoir paintings were found hidden behind a cinder block wall in southern France. The paintings, “Portrait of Madame Colonna Romano” and “Young Woman at the Well,” were taken from a museum in Cagnes-sur-Mer along with two other works that were abandoned during the theft. Authorities recovered the paintings during a search in Clermont-l’Herault after arresting six people, and DNA evidence linked a 29-year-old main suspect to the crime. A prosecutor said the hiding place was “completely undetectable” and that interrogation led the receiver of the stolen goods to admit possession. The paintings sustained slight damage, and the total value of the four stolen works had been estimated at $10 million.

10. RFC8950: Announce IPv4 with an IPv6 next-hop (27 points by porridgeraisin)

    This blog post explains RFC8950, a networking standard that lets IPv4 routes be announced with an IPv6 next hop. The author says this allows a network to run IPv6 only in its core while keeping IPv4 at the edge for customers or peers, without using tunnels. It relies on Multi-protocol BGP, a routing protocol that can carry information for different address families. The post lists use cases such as an IPv6-only transit backbone, an IPv6-only IP exchange, and customers continuing to use IPv4 normally. The author compares its potential impact on backbone networks to what 464XLAT did for mobile networks, while noting the setup is logically simple but operationally powerful.

AI/ML Insights & Trends

1. Local inference for very large models is becoming practical on deskside clusters

   • Trend/insight: Apple reports running the nearly trillion-parameter Muse Spark Open model with MLX on a cluster of four Mac Studios. The reported result is over 36 tokens per second at an 8,192-token context. This is a specific vendor claim, not proof that every large model will run equally well on similar hardware.
   • Why it matters for AI development: If you handle sensitive data or want to avoid per-token API costs, local inference can give you more control. It may also let you run fine-tuning experiments without sending data to a hosted provider.
   • Implications and practical tips: Expect tradeoffs in setup, power, memory, and model compatibility. Test your actual model and context length on available Apple silicon before assuming a four-Mac cluster can replace a hosted API.

2. Distributed inference shifts bottlenecks to memory and interconnect

   • Trend/insight: The MLX article shows scaling across four Mac Studios using RDMA over Thunderbolt. This suggests that once a model no longer fits on one machine, the connection between machines becomes as important as raw compute. The article reports the speed, but not enough detail to judge how well this generalizes.
   • Why it matters for AI development: For developers, distributed inference design affects latency, throughput, and cost. A slow interconnect can erase the benefit of adding more hardware.
   • Implications and practical tips: When planning local LLM serving, measure tokens per second at your target context length, not just model size. Check whether your framework supports efficient multi-machine communication before committing to a cluster.

3. Agent benchmarks are moving toward long-horizon coding and feedback loops

   • Trend/insight: StarSkirmish gives each LLM one hour to write a StarCraft bot in C++, then tests it against other bots. The model must run practice games and improve its bot, so the benchmark tests more than one-shot code generation. This is one benchmark, so it does not prove that all agentic coding tasks behave the same way.
   • Why it matters for AI development: If you build coding agents, success may depend on iteration, testing, and adapting to observed failures. A single prompt is often not enough for complex tasks.
   • Implications and practical tips: Give agents access to a safe test loop, logs, and time limits. Evaluate them on end-to-end outcomes, such as whether the resulting program wins or passes tests, not just on code similarity.

4. Cost-aware evaluation is becoming part of model selection

   • Trend/insight: The StarSkirmish preview compares model scores with average API cost for one one-hour run. It reports that GPT-6 Astra leads, while several other frontier models are tightly grouped. This shows that raw capability is not the only useful metric for agentic workloads.
   • Why it matters for AI development: A slightly better model may be much more expensive for long-running coding or reasoning tasks. Developers need to know cost per successful task, not just cost per token.
   • Implications and practical tips: Track both win rate or pass rate and total cost, including retries and tool calls. For production agents, set budgets and fallback models rather than always choosing the highest-scoring model.

5. AI theorem proving is advancing, but formalization is still an engineering process

   • Trend/insight: The Xena post says AI has become remarkably good at proving theorems formally and informally. However, the authors warn against rushing to formalize all of mathematics without care. They argue that a large-scale effort could help discovery if done well, or waste resources and cause damage if done poorly.
   • Why it matters for AI development: AI can help with math and verification-heavy code, but formal proofs still need definitions, libraries, and maintainable structure. Generated proofs that compile are not automatically well-organized or useful.
   • Implications and practical tips: Treat AI-generated formal proofs as drafts that need review, documentation, and integration. If you use proof assistants, budget time for maintenance and for fixing definitions, not just for generating proof scripts.

6. Privacy and control motivate local AI experiments

   • Trend/insight: Apple frames MLX on Mac as a private environment for running and fine-tuning models on your own hardware. This is a vendor claim, but it aligns with a broader practical concern: sending data to hosted APIs can be undesirable or impossible for some projects.
   • Why it matters for AI development: Local models can help with private datasets, regulated data, or offline environments. They also let developers inspect and modify the full stack more easily.
   • Implications and practical tips: Check whether local hardware meets your latency and context-length needs before committing. For sensitive workloads, compare local inference against private cloud or on-premises options, including setup and maintenance costs.

Analysis generated by deepseek-flash