Published on May 11, 2026 at 18:00 CEST (UTC+2)
Ratty – A terminal emulator with inline 3D graphics (389 points by orhunp_)
Ratty – A terminal emulator with inline 3D graphics
This article introduces Ratty, a GPU-rendered terminal emulator that supports inline 3D graphics directly within the terminal. It is designed to leverage modern graphics hardware for smooth rendering and interactive 3D content alongside traditional terminal workflows. The project is open-source and aims to push the boundaries of what a terminal can display. The post also links to a blog with further technical details and a download source.
Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s (76 points by zdw)
Training an LLM in Swift, Part 1: Taking matrix mult from Gflop/s to Tflop/s
Matt Gallagher details his journey optimizing handwritten matrix multiplication kernels in Swift for training a large language model on Apple Silicon. He compares ten implementations ranging from plain C to Metal, targeting CPU, SIMD, AMX, and GPU units. The goal is to achieve teraflop-level performance and provide insight into low-level optimization on Apple hardware. This is the first in a series that will later cover Apple’s ML frameworks and a full LLM implementation.
Hardware Attestation as Monopoly Enabler (1921 points by ChuckMcM)
Hardware Attestation as Monopoly Enabler
A post from GrapheneOS argues that Apple and Google are using hardware attestation features to create lock-in and hinder competition. By requiring verified hardware tokens for security-sensitive operations, they make it harder for users to switch platforms or use third-party services. The critique suggests that while attestation improves security, it also centralizes control and enables anti-competitive behavior.
Gmail registration now requires scanning a QR code and sending a text message (229 points by negura)
Gmail registration now requires scanning a QR code and sending a text message
Google has changed its account registration process to require users to scan a QR code and send an SMS from their phone to verify a number, replacing the previous SMS receipt method. The change is presented as a security measure against bots and phishing, but it also makes anonymous account creation much harder. Privacy advocates argue it harms ordinary users who rely on privacy tools or disposable numbers, while not preventing large-scale account farming.
Amália and the Future of European Portuguese LLMs (23 points by johnbarron)
Amália and the Future of European Portuguese LLMs
The article examines AMÁLIA, a €5.5 million Portuguese government-funded open-source LLM for European Portuguese. It is a continuation of pre-training from EuroLLM, not trained from scratch. The author praises the effort but raises critical questions about transparency, architecture choices, and whether the investment truly benefits the Portuguese-speaking community. Comparisons are drawn to Italy’s Minerva model and other national LLM initiatives.
Local AI needs to be the norm (1550 points by cylo)
Local AI needs to be the norm
This essay argues that relying on cloud-hosted AI APIs for application features creates fragile, privacy-invasive software. The author advocates for running AI models locally on user devices, which already contain powerful neural engines. Dependencies on third-party APIs introduce data retention issues, network dependency, and vendor lock-in. The post calls for a return to building software that respects user autonomy and work offline or with local compute.
Driver accused of DUI tracks missing laptop to Illinois State trooper's house (308 points by bryan0)
Driver accused of DUI tracks missing laptop to Illinois State trooper's house
A restaurant executive arrested for DUI discovered his MacBook was missing after his arrest and used its tracking feature to locate it at the home of the state trooper who arrested him. The story unfolds with a 911 call, video evidence, and an internal investigation. The trooper had been named “Top Cop” for most DUI arrests, but has faced allegations of fabricating evidence. The case highlights issues of police misconduct and the role of device tracking.
Venom and Hot Peppers Offer a Key to Killing Resistant Bacteria (85 points by littlexsparkee)
Venom and Hot Peppers Offer a Key to Killing Resistant Bacteria
Researchers at UNAM developed three new antibiotics from scorpion venom (Diplocentrus melici) and habanero peppers. They isolated benzoquinone molecules that turn blue or red upon oxidation, which show efficacy against Mycobacterium tuberculosis and Staphylococcus aureus. The work addresses antibiotic resistance by finding novel natural compounds. The blue benzoquinone demonstrated particular antimicrobial activity.
Building a web server in aarch64 assembly to give my life (a lack of) meaning (32 points by theanonymousone)
Building a web server in aarch64 assembly to give my life (a lack of) meaning
The author describes creating ymawky, a static HTTP server written entirely in aarch64 assembly for macOS using raw Darwin syscalls (no libc). It supports GET, HEAD, PUT, OPTIONS, DELETE, byte ranges, directory listing, and custom error pages. The project was done for educational purposes to understand low-level web server internals. It is not intended for production but as a demonstration of extreme minimalism.
Create a 90s GeoCities style website in seconds (Python) (22 points by whatsupdog)
Create a 90s GeoCities style website in seconds (Python)
This Python package, create-geocities-app, generates a retro GeoCities-style website with a single command. It automatically includes classic 90s web elements such as marquees, animated GIFs, guestbooks, and hit counters. The tool is aimed at nostalgia and fun, allowing users to quickly create a vintage-looking personal site. It is published on PyPI and requires only a few dependencies.
On-Device AI is gaining momentum as a counter to cloud dependency
The strong support for “Local AI needs to be the norm” (1550 points) reflects a growing backlash against always-online AI APIs. Developers and users are increasingly concerned about privacy, reliability, and latency. This trend is driving investment in edge inference hardware (Apple Neural Engine, Qualcomm AI Engine) and frameworks like CoreML, ONNX Runtime, and MediaPipe. Expect more apps to offer local inference options, especially for sensitive tasks like text completion, image generation, and document analysis.
National LLM initiatives highlight geopolitical and linguistic fragmentation
Portugal’s AMÁLIA project (€5.5M) joins Italy’s Minerva, France’s Mistral, and others in building language-specific models. This signals a shift away from English-centric LLMs toward sovereign AI infrastructure. Such projects risk duplicating efforts but also enable better representation of low-resource languages and cultural contexts. The open-source nature of AMÁLIA is promising, but the article questions cost-benefit and technical independence. The trend will likely accelerate as governments seek AI sovereignty.
Low-level optimization for LLM training on consumer hardware is a hot topic
The Swift matrix multiplication series (76 points but rich in technical depth) illustrates a niche but important area: hand-tuning ML kernels for Apple Silicon. As LLM training moves beyond large clusters, individual developers and small teams are trying to train usable models on personal Macs or PCs. This dovetails with tools like MLX, tinygrad, and llama.cpp. The insight: performance gains from hardware-aware code (SIMD, AMX, Metal) can be dramatic, and the community is eager to share techniques.
GPU-accelerated terminals hint at convergence of HCI and AI
Ratty’s inline 3D graphics in a terminal emulator (389 points) may seem unrelated to AI, but it points to a future where terminals become richer interfaces for AI interactions. Imagine AI assistants that can display 3D visualizations, real-time data plots, or interactive simulations directly in the CLI. Combined with local AI, this could redefine developer tools and dashboards. The trend is about making AI outputs more immersive without requiring a separate GUI.
Privacy and security backlash drives alternative AI architectures
Articles on hardware attestation as monopoly enabler (1921 points) and Gmail’s new QR/SMS registration (229 points) show growing distrust of centralized authentication and data collection. For AI, this means users want models that operate without sending personal data to cloud providers. Federated learning, differential privacy, and on-device fine-tuning are becoming more than research topics—they are demanded by the market. The high scores on these posts indicate strong community sentiment.
Extreme minimalism in computing (assembly web servers, retro web generators) reflects a counter-culture in the age of bloated AI
The aarch64 web server (32 points) and GeoCities generator (22 points) are outliers, but they represent a playful or educational pushback against complex AI stacks. The former is a deep-dive into bare-metal syscalls; the latter is a nostalgic reaction to over-engineered modern web. For AI/ML, this suggests an appetite for simplified, transparent, and even “retro” approaches—such as building tiny models from scratch or using rule-based systems for specific tasks rather than huge models.
Bioinformatics and drug discovery remain promising ML application areas
The scorpion venom/antibiotics story (85 points) is not directly about AI, but it underscores the data-intensive nature of modern drug discovery. ML models (e.g., AlphaFold, molecular generative models) are increasingly used to identify novel compounds from natural sources. The trend is that AI accelerates the screening of venom, plant extracts, and microbial samples for therapeutic properties. This article is a reminder that AI/ML’s biggest societal impact may still come from life sciences, not just chatbots.
Analysis generated by deepseek-reasoner