My Homelab AI Dev Platform

The author implemented a GitOps-driven homelab management system using OpenCode, a vendor-agnostic AI coding environment with a persistent web UI. This setup leverages LLMs to automate Docker Compose updates, healthcheck implementation, and network refactoring while maintaining security through a human-in-the-loop PR workflow. The architecture utilizes a dedicated VM for OpenCode to push changes to Git, which are then deployed via Arcane and other GitOps tools after manual approval.

Read more

Openrouter Fusion API

OpenRouter Fusion is a multi-model deliberation engine that processes prompts through a parallel panel of expert models with web search capabilities. A judge model synthesizes these outputs into a structured analysis identifying consensus, contradictions, and unique insights. The system is OpenAI-compatible and allows for custom panel configurations via the analysis_models field, with pricing based on the aggregate cost of all underlying completions.

Read more

Iroh 1.0

Iroh 1.0 is a stable networking stack that replaces IP-based addressing with a secure "dial-by-key" abstraction, facilitating direct, encrypted peer-to-peer connections. It features QUIC multipath, NAT traversal, and local-first discovery, making it highly efficient for distributed workloads like training LLMs and agent communication. The release includes official FFI support for Python, Node.js, Swift, and Kotlin, ensuring wire-protocol stability across diverse platforms and languages.

Read more

Veterinarian turned founder, AI lawn diagnosis

GrassDx is a computer vision platform that automates lawn diagnostics by processing multi-perspective imagery and localized geospatial data. The system identifies over 47 conditions and generates context-aware treatment plans tailored to specific ZIP codes and grass types. It demonstrates a specialized application of image classification and localized data retrieval to provide actionable, domain-specific insights.

Read more

India, UAE partner on AI sovereignty to bypass Google, Microsoft

India is partnering with UAE-based G42 to deploy an AI supercomputer powered by 64 Cerebras systems, aiming to enhance AI sovereignty and reduce dependence on US hyperscalers. While India’s existing infrastructure relies heavily on Nvidia GPUs, the Cerebras wafer-scale hardware offers a specialized alternative optimized for high-speed inference and large-scale application deployment. This deal represents the first rollout of G42’s "Intelligence Grid," allowing India to maintain domestic data governance while diversifying its compute stack.

Read more

The efficiency-gain illusion: People underestimate the rate of AI use

User studies (N = 2691) indicate that people frequently use AI for simple tasks even when it offers no objective efficiency gains. This overreliance is driven by self-estimate miscalibration and "efficiency-gain illusions," where users overestimate time and effort savings. Additionally, a session-level carryover effect reinforces these biases, creating a feedback loop that entrenches inefficient AI adoption.

ArXiv

AI language models have favorite names, and we mapped them

LLMs exhibit model-specific "behavioral fingerprints" by generating recurring, correlated ensembles of fictional names—such as Claude’s "Elena Vasquez and Marcus Chen"—across independent generations. These synthetic identities have been used to populate scholarly repositories like Zenodo and ResearchGate with over 1,600 ghost-authored records featuring fabricated metadata and real DOIs. This large-scale infiltration provides a temporal proxy for model deployment windows and highlights the vulnerability of academic metadata aggregators to automated content generation.

ArXiv

You Can Game AI Peer Review with Presentation-Only Revisions

The study introduces "adversarial repackaging," a closed-loop attack that optimizes a paper's presentation—such as framing, related work, and discussion—based on AI-reviewer feedback without altering the underlying scientific evidence. This method achieved a 75.1% success rate and a +1.21/10 mean score gain across three AI reviewers, significantly outperforming surface-level polishing. The results highlight structural vulnerabilities where AI reviewers confuse the appearance of addressing limitations with actual resolution, demonstrating that paper presentation has become a critical optimization surface for LLM-based evaluation.

ArXiv

Teaching Machine Learning to Software Engineers

This paper addresses the lack of systematic AI/ML preparation in undergraduate SE curricula by identifying coverage gaps and instructor priorities. It provides a structured inventory of topics and evidence-based guidelines for integrating the development, testing, and deployment of AI/ML-based systems into core SE education.

ArXiv

DPBench: Structural Determinants of Multi-Agent LLM Coordination

DPBench is a benchmark that adapts the Dining Philosophers problem to evaluate multi-agent LLM coordination across varying action protocols, communication structures, and group sizes. Testing models like GPT-5.2 and Gemini 2.5 Flash, the study finds that deadlock rates are primarily driven by protocol design—such as pre-commitment communication and classical concurrency primitives—rather than raw model capability. The results demonstrate that structural conditions, including sequential action and resource-ordering prompts, can reduce deadlock rates from near-total failure to zero.

ArXiv

I wrote a C++ ray tracer from scratch without AI

Luz is a zero-dependency C++20 path tracer featuring Monte Carlo global illumination, atmospheric scattering, and BVH acceleration with binned SAH construction. It incorporates advanced rendering techniques such as adaptive sampling and an NFOR-style denoiser, optimized for multithreaded CPU performance via LTO and native hardware tuning. The project includes a custom scene format, a Blender exporter, and a deterministic benchmarking harness for performance analysis.

Repo

Can Europe train a frontier AI model on the compute it owns?

EuroMesh proposes federating existing EuroHPC and national AI Factory compute to train a frontier-class LLM by 2028, bypassing the ~7.6-year grid-connection lead times required for planned 1GW datacenters. The project utilizes a three-layer model to analyze DiLoCo-style low-communication training efficiency, time-to-availability, and regional feasibility. While acknowledging risks like hardware heterogeneity and the unproven nature of distributed training at frontier scales, the analysis suggests federation is the fastest path to European AI sovereignty.

Repo

Macro – unified system for email, chat, tasks, docs, agents (AGPL/Rust)

Macro is an open-source, unified workspace platform built with Rust and SolidJS that integrates email, tasks, docs, and communication into a single system. It features a unified team-level memory that powers autonomous agents and provides an MCP server for seamless integration with external LLMs like Claude. The architecture emphasizes technical workflows through bidirectional linking, CRDT-based collaborative docs, and AI-driven signal filtering across a unified inbox.

Repo

Dream Server – Turn your PC, Mac, or Linux box into a private AI server

Dream Server is an all-in-one private AI stack that automates the deployment of local inference, RAG, and agentic workflows on Linux, macOS, and Windows. It features a modular architecture using Docker Compose to orchestrate services like llama-server, Open WebUI, n8n, and Qdrant, with a custom CLI for managing hardware-optimized model tiers. The platform supports local, cloud, and hybrid modes, providing a pre-wired environment for voice, image generation, and autonomous agents without external subscriptions.

Repo

WSP WordPress MCP – Connect AI Agents to WordPress

The WSP WordPress MCP plugin enables AI agents to programmatically interact with WordPress sites by exposing a suite of "abilities." These include CRUD operations for posts, pages, comments, and media, along with reading user and site information. It also offers extensive Elementor integration, allowing AI agents to list pages, manipulate page structures, and add/remove elements. The plugin features a modular architecture, an auto-config generator for tools like Claude Desktop and Codex, and granular control over each ability.

Repo