Published on March 26, 2026 at 06:01 CET (UTC+1)
Running Tesla Model 3's computer on my desk using parts from crashed cars (471 points by driesdep)
A security researcher details how they purchased salvaged Tesla Model 3 computer components (MCU and Autopilot computer) from eBay to run the vehicle's operating system on their desk. The goal was to obtain the necessary hardware to participate in Tesla's bug bounty program by analyzing the system for vulnerabilities. The process involved sourcing the water-cooled computer, a touchscreen, and a DC power supply to successfully boot and interact with the car's software outside of the vehicle.
ARC-AGI-3 (311 points by lairv)
This article introduces ARC-AGI-3, a new interactive reasoning benchmark designed to measure human-like general intelligence in AI agents. Unlike static puzzles, it requires agents to learn from experience within novel environments, acquire goals dynamically, and adapt their strategies over time without explicit instructions. The benchmark aims to quantify the gap between human and AI learning efficiency, focusing on long-horizon planning, memory compression, and continuous adaptation.
My astrophotography in the movie Project Hail Mary (764 points by wallflower)
The article is a showcase by astrophotographer Rod Prazeres, featuring his work used in the end credits of the movie Project Hail Mary. It serves as a portfolio or story page highlighting his contribution to the film, presenting his astrophotography imagery that was integrated into the production.
False claims in a widely-cited paper (198 points by qsi)
Based on the title and source, this article from a statistics blog criticizes the academic publishing system, highlighting a case where a widely-cited paper in a business journal contains false claims. It laments that there have been no corrections or consequences, using this as an example of problematic incentives and lack of accountability in certain academic fields.
Two studies in compiler optimisations (37 points by hmpc)
This technical blog post delves into the inner workings of compiler optimizations, using two specific case studies (a modular increment operation and endianness conversion) to illustrate how compilers like LLVM transform code. It demonstrates how seemingly minor source code changes can lead to surprising performance differences by triggering different optimization pathways, emphasizing the complex machinery beneath the "black box" of modern compilers.
Earthquake scientists reveal how overplowing weakens soil at experimental farm (127 points by Brajeshwar)
University of Washington researchers, using portable fiber-optic seismic sensors (DAS), conducted an experiment on a farm to study how repeated plowing (tilling) weakens soil structure. Their "agroseismology" approach measured how seismic waves travel through tilled versus untilled soil, revealing that overplowing reduces soil moisture and cohesion, contributing to degradation. This interdisciplinary research connects agricultural practices to geophysical measurement techniques.
Show HN: Robust LLM Extractor for Websites in TypeScript (8 points by andrew_zhong)
This is a "Show HN" announcement for "Lightfeed Extractor," a TypeScript library for robust web data extraction. It uses LLMs combined with browser automation (via Playwright) to navigate websites and extract structured data based on natural language prompts. The tool is designed for production data pipelines, emphasizing accuracy, token efficiency, and stealthy automation to handle dynamic web content.
The EU still wants to scan your private messages and photos (845 points by MrBruh)
This is a advocacy campaign website ("Fight Chat Control") opposing proposed EU regulations (often called "Chat Control") that would mandate the scanning of private messages and photos for illegal content. It warns of a renewed political attempt to pass such legislation, framing it as a mass surveillance measure and an attack on digital privacy and democratic principles, and calls for public action to stop it.
90% of Claude-linked output going to GitHub repos w <2 stars (221 points by louiereederson)
This site presents analytics on the adoption of Claude Code (an AI coding assistant), revealing that 90% of its traced output is being committed to low-engagement GitHub repositories (those with fewer than 2 stars). It provides metrics on growth, top programming languages used (TypeScript, Python, JavaScript), and total volume of code changes, suggesting the tool is heavily used for early-stage or personal projects rather than major, popular open-source work.
My DIY FPGA board can run Quake II (91 points by sznio)
In part four of a detailed project log, an engineer describes designing a new, more advanced DIY FPGA board capable of running Quake II. The board features a BGA-packaged FPGA (Efinix Ti60) and DDR3L memory, requiring the author to learn and implement complex PCB layout techniques like trace length matching. The post covers the challenges of moving beyond a prototype, including soldering BGAs and integrating a DDR3 soft controller core.
The Shift to Interactive, Experiential AI Benchmarks: The launch of ARC-AGI-3 signals a move beyond static Q&A or puzzle-solving benchmarks. It matters because it more closely tests human-like learning—adaptation, planning, and skill acquisition over time in novel environments. The implication is that future AI development must focus on agents that can learn and reason within a context, not just recall pre-trained patterns, pushing the field closer to AGI objectives.
Proliferation of AI-Assisted Coding in Early-Stage Development: Data showing Claude Code's output concentrated in low-star GitHub repos indicates AI coding tools are becoming ubiquitous for prototyping, personal projects, and boilerplate generation. This matters as it dramatically lowers the barrier to entry for programming and accelerates early development cycles. A key takeaway is that while AI boosts productivity, its current impact may be more on volume and accessibility of code rather than on the core logic of mature, critical projects.
Convergence of AI and Physical/Hardware Systems: The Tesla computer hack and the FPGA project running Quake II highlight a trend where deep software understanding (including AI systems like autopilot) requires engagement with specialized hardware. For AI/ML, this underscores the importance of hardware-in-the-loop testing, security research on embodied AI systems (like cars), and the need for efficiency that drives custom silicon (like FPGAs for ML acceleration). Developers can't treat advanced AI as purely abstract software.
LLMs as Robust Interfaces for Unstructured Data Extraction: Tools like Lightfeed Extractor represent the trend of using LLMs as core reasoning engines for automating complex, variable tasks like web scraping. This matters because it moves beyond simple API consumption to handling the messy, dynamic structure of the web with natural language instructions. The implication is the rise of "AI-native" data pipelines that are more adaptable and require less manual selector maintenance.
Growing Tension Between AI Capability and Privacy Regulation: The high engagement with the EU Chat Control article reflects a major impending conflict. As AI-powered scanning and analysis capabilities grow, so do legislative efforts to use them for surveillance, creating a direct clash with privacy norms. For AI developers, this trend means navigating an increasingly complex regulatory landscape where the very techniques that make models powerful (e.g., content analysis) could be mandated or restricted in ethically fraught ways.
The Critical Need for Transparency and Auditing in AI/ML Research: The article on the uncorrected false paper, though not exclusively about AI, is highly relevant. As AI research accelerates, the credibility of published findings becomes paramount. This trend matters because flawed or fraudulent research can misdirect entire subfields. The takeaway is a growing need for stronger replication studies, open code/data, and audit tools specifically for ML papers to ensure the foundation of the field is solid.
AI as a Catalyst for Interdisciplinary Measurement Tools: The use of fiber-optic DAS (a sophisticated sensing technology) for soil analysis ("agroseismology") exemplifies how advanced data capture enables new science. For AI/ML, the trend is the creation of novel, high-resolution datasets from unexpected domains (agriculture, geology, etc.). This provides rich training data for environmental AI models and opens new application frontiers, suggesting AI practitioners should look for partnerships in fields with emerging sensing capabilities.
Analysis generated by deepseek-reasoner