Hacker News. Daily summary. Top 20 stories.
Hackers triggered automatic defrosting in supermarket fridges across Europe, spoiling stock.
Thread calls this a repeat of August's US military commissary freezer hack.
Meta releases Muse Wonder models, claiming top performance on personal and professional agent tasks.
Split between praise for Meta's execution and skepticism over benchmark-maxxing and 'superintelligence' branding.
Shipping a commercial web product in C via CGI in 2026 is feasible and surprisingly modern.
Readability complaints about the site's color scheme and debate over choosing C vs safer alternatives.
GLM-Legal matches frontier models on legal benchmarks with a smaller model footprint.
Liability concerns dominate: if LLMs hallucinate cases, the attorney remains responsible.
BBC Panorama finds Deloitte helped falsify documents to hide Roman Abramovich's wealth and evade tax.
Cynicism dominates: commenters see Big Four firms as enablers of corruption, not ethical auditors.
Codex CLI installs a persistent, self-updating daemon by default, sparking backlash.
Commenters say the daemon violates trust and makes the CLI unusable; workarounds exist.
LLMs are tools that require skill to wield effectively.
Two camps: those who see LLMs as skill-requiring tools and those focused on marketing framing.
A Wi-Fi 7 router's chipset is repurposed as a 320 Msps wideband SDR receiver.
Enthusiasm for an under-$100 SDR hitting 320 Msps at 5 GHz.
Citroen revives the 2CV as a retro-styled electric city car to compete against Chinese brands.
Skepticism that retro styling and nostalgia are enough to compete with cheaper, more capable Chinese EVs.
QGIS adopts a strict policy banning all AI-assisted code contributions.
Broad support for the ban, with dissent over enforcement and perceived hypocrisy.
An 11-year-old unsoundness bug in Rust's borrow checker allows safe code to violate lifetime guarantees.
Shock at a simple, decade-old soundness hole that surfaces with custom Drop and covariance.
AI coding agents boost code output 30% but human review bottlenecks limit software gains.
Commenters validate the bottleneck finding from direct experience, calling code review the most tedious job part.
AI companies are exploiting a responsibility gap to avoid hacking charges, but the article argues intent is clear from their design choices.
Split between those arguing intent is hard to prove and others citing precedent like Swartz or ongoing state investigations.
AI advances will replace most cognitive work, making thinking a leisure activity.
Two camps: LLMs don't think vs. progress curve makes outcome irrelevant.
Static recompilation translates PS2 machine code to C++ and WebAssembly, letting Gran Turismo 4 and God of War II run in a browser with a WebGPU renderer.
The author details how static recompilation approximates PS2 timing without cycle accuracy, and commenters debate analog input mapping for racing games.
Rich Hickey argues LLMs are not tools because they lack control, feedback, and continuous direction.
Two camps: those arguing LLMs can be tools with careful use, those defending Hickey's tool-automation distinction.
Intel announces D270 GPU with 480 GB LPDDR5X, targeting agentic AI inference workloads.
Skepticism about consumer availability, but interest in the high memory capacity and bandwidth for inference.
Linux phones may finally become viable due to new hardware and software conditions.
Skeptics argue battery life and driver documentation are fundamental barriers Linux has not overcome.
Researchers built backdoor attacks that hide arbitrary capabilities inside AI models, provably hard to detect.
Familiar threat now given a formal proof; skepticism about detection practicality.
Setting Firefox's network.http.response.timeout to 0 makes big downloads resumable.
External downloaders (aria2, JDownloader, GetRight) are widely recommended as workarounds for Firefox's broken resume.