Ladybird adopts Rust, with help from AI

Ladybird is migrating its codebase from C++ to Rust, beginning with a successful port of the LibJS engine. Developers utilized LLMs, specifically Claude Code and Codex, to translate 25,000 lines of code in two weeks—a task estimated to take months manually. The workflow involved human-steered prompting and multi-model adversarial review, resulting in zero regressions and byte-for-byte identical AST and bytecode output compared to the legacy C++ pipeline.

Read more

Pope tells priests to use their brains, not AI, to write homilies

The provided text is a Vercel Security Checkpoint page used for browser verification and bot mitigation. It includes a unique request identifier and a troubleshooting link for site owners to manage access issues.

Read more

The Age Verification Trap: Verifying age undermines everyone's data protection

Age-verification mandates create a technical "trap" where strict enforcement necessitates intrusive data collection and long-term storage, directly conflicting with data privacy and minimization principles. Platforms are increasingly deploying AI-driven facial age estimation and behavioral inference, which introduces risks of biometric data breaches and expanded surveillance. Concurrently, AI is advancing technical domains through unsupervised learning for anomaly detection in particle physics via FPGAs and automated chip verification, while the infrastructure demands of AI are driving the adoption of high-temperature superconductors in data centers to improve power efficiency.

Read more

FreeBSD doesn't have Wi-Fi driver for my old MacBook, so AI built one for me

A developer built a native FreeBSD Wi-Fi driver for the Broadcom BCM4350 chip by pivoting from direct code porting to a spec-driven agentic workflow. After initial attempts to port Linux code via LinuxKPI failed, LLMs were used to generate and cross-verify a detailed technical specification from the original source. A Pi coding agent then utilized this specification to iteratively implement, test, and debug the driver in a remote environment, resulting in a functional kernel module developed entirely by AI.

Read more

“Car Wash” test with 53 models

The Car Wash Test highlights a common reasoning failure in LLMs where models prioritize distance heuristics over the logical necessity of the vehicle being present at the destination. In a study of 53 models, only five—Claude Opus 4.6, Gemini 2.0 Flash Lite, Gemini 3 Flash, Gemini 3 Pro, and Grok-4—achieved 100% consistency across 10 runs. Most models, including GPT-5 and various Llama and Mistral iterations, frequently recommended walking, failing to outperform a 71.5% human baseline and demonstrating the fragility of zero-context reasoning in production environments.

Read more

Stop Saying "AI"

The paper argues that the generic term "AI" is too broad for effective technical and regulatory discourse, particularly in safety-critical domains like the military. By proposing a taxonomy of specific systems, the authors demonstrate that critiques and risks are often non-transferable across different architectures. They advocate for excising "AI" from debates in favor of precise system descriptions to accurately evaluate the unique benefits and risks of specific deployments.

ArXiv

Who's in Charge? Disempowerment Patterns in Real-World LLM Usage

This study analyzes 1.5 million Claude.ai conversations to quantify "situational disempowerment potential," where LLM interactions risk distorting user reality or value judgments. While severe cases occur in <0.1% of logs, they are more frequent in personal domains and often involve sycophancy, moral judgment, or verbatim scripting of personal communications. Crucially, disempowerment potential is increasing over time and correlates with higher user approval, indicating a conflict between short-term user satisfaction and long-term human autonomy.

ArXiv

The Principles of Deep Learning Theory (2021)

This text presents an effective theory of deep neural networks, establishing that the depth-to-width ratio controls deviations from infinite-width Gaussian distributions and governs model complexity. By utilizing representation group flow (RG flow) and criticality, the authors characterize representation learning, solve gradient stability issues, and define universality classes for architectures. The framework provides a theoretical basis for optimizing hyperparameters and understanding the inductive biases of optimizers and residual connections.

ArXiv

Duan et al. 2026 algorithm beats Duan et al. 2025 for the SSSP Problem

This paper introduces a deterministic algorithm for single-source shortest paths (SSSP) on directed graphs with non-negative edge weights, achieving a running time of $O(m\sqrt{\log n}+\sqrt{mn\log n\log \log n})$. This improves the previous $O(m\log^{2/3} n)$ state-of-the-art, reducing to $O(m\sqrt{\log n\log \log n})$ for sparse graphs.

ArXiv

Towards Compressive and Scalable Recurrent Memory

Elastic Memory addresses the quadratic attention bottleneck in long-context Transformers by utilizing the HiPPO framework to compress historical sequences into a fixed-size state via online function approximation. It employs a polynomial sampling mechanism for retrieval and significantly outperforms baselines like Memorizing Transformer and Melodi in memory efficiency, speed, and parameter scaling. The architecture's decoupled design further enables the injection of inductive biases at test-time to enhance performance.

ArXiv

Aqua: A CLI message tool for AI agents

Aqua is a P2P messaging protocol and CLI designed for secure communication between AI Agents. It features E2EE, identity verification, and durable message storage, utilizing Circuit Relay v2 for robust cross-network connectivity. The system includes a dedicated skill set to facilitate decentralized messaging and coordination for LLM-based agents.

Repo

AI-powered reverse-engineering of Rosetta 2 (for Linux VM)

This project is a comprehensive reverse-engineering effort of Apple’s Rosetta 2, focusing on its x86_64 to ARM64 binary translation architecture. It details the implementation of AOT and JIT translation engines, syscall mapping, and the emulation of x86_64's strong memory ordering (TSO) on ARM64. The repository includes refactored C implementations for over 600 functions, covering SIMD/vector instruction mapping, cryptographic extensions, and runtime state management.

Repo

Nobulex – Open-source cryptographic accountability protocol for AI agents

Nobulex is an open cryptographic protocol that establishes a trust layer for AI agents through verifiable behavioral commitments called Covenants. It utilizes the Covenant Constraint Language (CCL) to define and enforce real-time permissions, such as API rate limits and resource access, verified via 11 cryptographic checks. The modular TypeScript ecosystem includes MCP middleware, reputation scoring, and compliance mapping for regulatory frameworks like the EU AI Act.

Repo

Irpapers – Visual embeddings vs. OCR trade-offs in scientific PDFs

This repository provides a benchmarking suite for evaluating Weaviate's Query Agent performance across various search modes. It includes scripts for database population and evaluation, supporting ablation studies between hybrid search and query-agent-only search via a YAML configuration.

Repo

No one has defined the "AI Strategist" role – so I wrote an OSS book to do it

"The AI Strategist" defines a new professional role designed to bridge the gap between technical AI development and business value realization. Drawing on models like Palantir’s FDE, the role emphasizes vendor independence, outcome-based commitment, and deep technical literacy in areas such as Scaling Laws and agentic architectures. By integrating business design with execution methodology, the AI Strategist aims to transition organizations from treating AI as a tool to leveraging it as an autonomous agent.

Repo