Published on March 24, 2026 at 06:01 CET (UTC+1)
Epoch confirms GPT5.4 Pro solved a frontier math open problem (170 points by in-silico)
Researchers used GPT-5.4 Pro to solve an open problem in Ramsey theory concerning hypergraphs. The AI's solution was validated by the mathematician who proposed the problem and will be written up for publication, with the AI prompters offered co-authorship. Other advanced models like Claude Opus and Gemini Pro later solved it within a structured testing scaffold, demonstrating AI's growing capability in frontier mathematical research.
Box of Secrets: Discreetly modding an apartment intercom with Matter (13 points by swq115)
A technical blog post details a hardware hack to retrofit an old apartment intercom (a Doorking 1834) with Matter and Apple Home compatibility. After the building's cellular service failed, the author and a friend reverse-engineered the system, eventually implementing a "man-in-the-middle" device to intercept signals and allow remote unlocking via a home automation system, all done discreetly inside the existing hardware enclosure.
Autoresearch on an old research idea (315 points by ykumards)
The author experiments with "Autoresearch," an AI-agent loop concept, by applying it to an old machine learning project (eCLIP). Using Claude Code, the agent autonomously iterates through hyperparameter tuning and architectural changes by editing code, training, and evaluating against a metric. The project highlights the potential and challenges of using AI for fully automated research and optimization within a sandboxed environment.
FCC updates covered list to include foreign-made consumer routers (249 points by moonka)
The U.S. Federal Communications Commission (FCC) has updated its "covered list" of equipment deemed a national security threat to include certain foreign-made consumer routers. This action bans their authorization for import or sale in the U.S., reflecting escalating concerns about the security of network hardware in consumer homes and its potential for surveillance or disruption.
Scott Hanselman says he's working on Windows local accounts (26 points by teekert)
Scott Hanselman, a prominent Microsoft employee and technologist, tweeted that he is working on improving the local account experience for Windows. This suggests ongoing internal efforts to address user and regulatory concerns about the operating system's push towards mandatory Microsoft account logins for some setups, aiming to better support offline, non-cloud user profiles.
iPhone 17 Pro Demonstrated Running a 400B LLM (531 points by anemll)
A tweet demonstrates a future iPhone 17 Pro reportedly running a 400-billion-parameter large language model (LLM) directly on the device. While details are scarce in the preview, such a claim points to significant anticipated advances in mobile chip AI performance and model compression techniques, potentially enabling powerful, private, on-device AI assistants without constant cloud dependency.
Abusing Customizable Selects (63 points by speckx)
This CSS-Tricks article is a fun, exploratory tutorial on creatively styling the new customizable <select> HTML element. The author builds playful demos like a curved stack of folders and other non-standard UI widgets, explaining the techniques used. It highlights how new web platform features can be leveraged for artistic and experimental front-end development, even if the practical use is limited.
Pompeii's battle scars linked to an ancient 'machine gun' (48 points by pseudolus)
An archaeological news report links unusual battle damage found on structures in Pompeii to a specific type of ancient Roman artillery weapon, likened to a "machine gun" for its rapid-fire capability. The analysis suggests this weapon was used during the city's final conquest by Roman forces, providing new insight into the military tactics and violent history preceding the famous volcanic eruption.
Show HN: Cq – Stack Overflow for AI coding agents (99 points by peteski22)
Mozilla.ai introduces "cq," a platform envisioned as a "Stack Overflow for AI coding agents." It addresses the repetitive failures and inefficiencies of AI agents by creating a shared knowledge base where successful solutions to common agent problems (like specific errors or prompts) can be stored and retrieved, reducing wasted computation and improving agent reliability across different systems.
The Resolv hack: How one compromised key printed $23M (74 points by timbowhite)
Chainalysis analyzes the "Resolv" DeFi hack, where an attacker minted $23 million in unbacked stablecoins by compromising a single privileged off-chain private key. The exploit underscores how DeFi's security risk has expanded beyond smart contract bugs to include the infrastructure and governance keys around protocols, highlighting the need for robust real-time monitoring and response systems as a final defense layer.
AI as an Autonomous Research Co-pilot: Articles 1 and 3 demonstrate AI transitioning from a tool to an autonomous researcher. GPT-5.4 Pro solved a novel math problem, while the Autoresearch project automated the ML experimentation loop. This matters because it accelerates the pace of discovery and could democratize research. The implication is a future where human researchers increasingly manage and validate AI-driven research agendas, requiring new skills and ethical frameworks for AI-generated intellectual property.
The On-Device AI Revolution: Article 6 (iPhone 17 Pro) signals the intense drive towards deploying massive models directly on consumer edge devices. This trend matters for privacy, latency, and accessibility, enabling fully offline, responsive AI. The implication is an arms race in mobile SOC design and novel model compression techniques (like pruning, quantization), potentially decentralizing AI power away from cloud giants and into users' pockets.
The Rise of the Agent Ecosystem & Tooling: Article 9 (cq) explicitly addresses the operational headaches of AI agents, mirroring the early days of software engineering. This trend highlights that as agentic AI becomes common, the ecosystem needs supporting infrastructure for debugging, knowledge sharing, and efficiency. The takeaway is that major opportunities exist in building the "devtools" and platforms for AI agents, focusing on observability, reproducibility, and collaboration between agents.
Expanding Attack Surface: AI & Infrastructure Security: The Resolv hack (Article 10) and the FCC router ban (Article 4) are connected warnings. As AI integrates into critical systems (from DeFi to IoT), the attack surface grows to include the AI models themselves, their training data, and the hardware they run on. This matters because securing AI is no longer just about model poisoning; it's about the entire supply chain. Developers must adopt a security-first mindset, assuming infrastructure compromise and implementing real-time anomaly detection.
Specialized Hardware Hacking Meets AI Integration: Article 2 (intercom hack) is a microcosm of a larger trend: using AI to interface with and control legacy physical systems. While the article doesn't use AI, the next logical step is using vision or language models to interpret unstructured hardware data. The implication is that AI will become a key tool for reverse engineering and creating smart, adaptive integrations in the physical world, blurring the lines between software hacking and intelligent automation.
Regulatory and Consumer Pushback on AI/Cloud Assumptions: Articles 5 (Windows local accounts) and 4 (FCC router list) reflect growing regulatory and user demand for control, privacy, and sovereignty. This matters for AI/ML as it pushes against the "cloud-only" and "data-hungry" default model. The actionable takeaway is that successful future products must offer robust offline functionality, data localization options, and transparent hardware provenance to meet diverse global standards and user preferences.
The Knowledge Curration Crisis in the AI Age: Article 9 notes the decline of Stack Overflow, pointing to a broader trend: the potential erosion of structured human knowledge repositories as users turn to LLMs for answers. This matters because LLMs can hallucinate and their knowledge is static after training. The implication is a need for hybrid systems that combine LLMs with curated, verified knowledge bases (like "cq"), ensuring accuracy and preserving collective problem-solving history for both humans and agents.
Analysis generated by deepseek-reasoner