Qwen3-TTS family is now open sourced: Voice design, clone, and generation

Qwen3-TTS is an open-source speech generation series featuring 1.7B and 0.6B models built on a discrete multi-codebook LM architecture. Utilizing a 12Hz tokenizer and a non-DiT Dual-Track hybrid streaming design, it achieves 97ms end-to-end latency and delivers the first audio packet after a single character input. The system supports zero-shot voice cloning, natural language-driven voice design, and fine-grained control over emotion and prosody across 10 languages.

Read more

Text-to-video model from scratch (2 brothers, 2 years, 2B params)

Linum-AI has released Linum-v2, a suite of Apache 2.0 licensed text-to-video models. The collection features a 2B parameter architecture capable of generating 2-5 second video clips at 360p and 720p resolutions.

Read more

I've been using AI to analyze every supplement on the market

Pillser is a supplement research and comparison platform that aggregates data from 15.9K supplements, 4.4K research papers, and 7.4K health outcomes. It features an "Ask AI" interface that leverages this dataset to provide automated, evidence-based responses to user queries. The platform integrates large-scale health data with LLM-driven search to facilitate informed decision-making.

Read more

Composing APIs and CLIs in the LLM era

LLM tool integration is evolving from fine-grained tool definitions toward shell-based CLI composition via exec_bash. This approach leverages Unix-style pipelines to reduce token costs and improve script reusability across human and machine interfaces. By using OpenAPI interpreters like Restish and OAuth clients like oauth2c, developers can automate SaaS interactions without writing boilerplate code. For undocumented services, HAR captures combined with LLMs allow for rapid reverse-engineering of private APIs into functional command-line tools.

Read more

Bible translated using LLMs from source Greek and Hebrew

Biblex Online Reader is a digital scripture platform featuring AI-integrated tools such as semantic search for conceptual verse discovery and an LLM-powered "Interpret" function for contextual Q&A. The interface supports standard reader features including translation switching, infinite scrolling, and customizable display modes.

Read more

Vibe Coding Kills Open Source

Vibe coding leverages AI agents to assemble OSS, increasing development productivity while decoupling users from direct maintainer engagement. This shift risks destabilizing the OSS ecosystem by reducing maintainer incentives, potentially leading to lower code quality and availability. Sustaining the ecosystem under widespread vibe coding requires fundamental changes to how OSS maintainers are compensated.

ArXiv

DaggerFFT: A Distributed FFT Framework Using Task Scheduling in Julia

DaggerFFT is a Julia-based distributed FFT framework that utilizes a dynamic task-graph scheduler and work-stealing to overcome synchronization bottlenecks in heterogeneous HPC environments. By representing FFT stages as tasks on distributed arrays, it achieves up to 2.6x speedup on CPU clusters and 1.35x on GPU clusters compared to state-of-the-art libraries. Its integration into Oceananigans.jl demonstrates the efficacy of high-level, task-based runtimes for large-scale scientific simulations.

ArXiv

Debunking the Myth of Join Ordering: Toward Robust SQL Analytics

Robust Predicate Transfer (RPT) introduces the LargestRoot and SafeSubjoin algorithms to provide provable join-order robustness for acyclic queries. Integrated into DuckDB, RPT limits execution time variance across random join orders to a 1.6x ratio while delivering a 1.5x improvement in end-to-end query performance. This approach addresses the instability of modern query optimizers, which often generate suboptimal join plans that are significantly slower than the theoretical optimum.

ArXiv

The unreasonable effectiveness of pattern matching

LLMs demonstrate a surprising ability to decode "Jabberwocky" text by mapping nonsense strings to meaningful content through structural pattern-matching. This capability suggests that advanced pattern-matching is a fundamental component of intelligence rather than mere mimicry or database retrieval.

ArXiv

Memory DisOrder: Memory Re-Orderings as a Timerless Side-Channel

Memory DisOrder is a timerless side-channel vulnerability that exploits memory re-ordering in relaxed memory models to infer cross-process activity on modern CPUs and GPUs. By monitoring out-of-order execution frequencies, researchers demonstrated covert channels with throughput up to 30K bps and reliable DNN architecture fingerprinting across X86, Arm, and Apple silicon. This research highlights how hardware-level optimizations for parallel processing can be leveraged for high-precision side-channel attacks.

ArXiv

BrowserOS – "Claude Cowork" in the browser

BrowserOS is an open-source Chromium fork designed to run AI agents natively while prioritizing local data privacy. It supports integration with OpenAI, Anthropic, and local models via Ollama or LMStudio, and can function as an MCP server for tools like claude-code. The platform enables local task automation, data scraping, and AI-driven ad blocking while maintaining full compatibility with Chrome extensions.

Repo

AI usage policy for Ghostty contributions

Ghostty is a high-performance, native terminal emulator featuring a Zig-based core and a multi-renderer architecture using Metal and OpenGL. It achieves low-jitter IO and high frame rates through a dedicated IO thread and platform-specific native UIs like SwiftUI and GTK. Beyond a standalone app, it provides libghostty, a C-compatible library for embedding standards-compliant terminal emulation into third-party projects and CLI tools.

Repo

LaReview, local open-source CodeRabbit alternative

LaReview is a local-first code review workbench written in Rust that leverages AI agents to transform PRs and diffs into structured, flow-based review plans. It integrates with various LLMs via the Agent Client Protocol to identify architectural risks and enforce team standards while maintaining data privacy. Key features include task-focused diffs, automated diagram generation, and a feedback loop that calibrates the AI to reduce noise and improve signal over time.

Repo

NeuralVoid – Block AI Telemetry from Copilot, Grammarly, Adobe

NeuralVoid is a system-level telemetry blocker for Windows 10/11 that null-routes AI-related endpoints via the hosts file. It targets data collection services from Microsoft Copilot, GitHub Copilot, Adobe Firefly, and Grammarly to prevent unauthorized AI telemetry. The tool is available as a standalone executable or a Python script requiring administrative privileges.

Repo

I made a Mac app for rate limiting and monitoring LLM requests

LLMWatcher is a local desktop app designed to monitor and control LLM requests, primarily for coding agents. It functions as a pass-through proxy, enabling users to block or get notified on high request/token usage, maintain a searchable history, and automatically redact sensitive information using pre-defined or custom patterns. The app supports Codex, Claude, and Cursor, and allows configuration of custom LLM endpoints for data monitoring and control.

Repo