Daily
10 selected stories for 2026-06-20, grouped by topic. The rest is folded into the archive.
RikkaHub Agent is an open-source Android app that transforms a local LLM chat client into a powerful on-device AI agent with over 80 native device tools, workflow automation, Telegram bot, SSH, voice transcription, and more, all running locally with privacy-first design.
samtSQL enables you to run SQL queries enhanced with AI operators on your existing PostgreSQL database, supporting multimodal data including text, images, and audio.
AI doesn't make bad engineers good, it makes them fast. It amplifies existing traits rather than improving skills. For lazy or sloppy coders, AI accelerates output of low-quality code, and because AI itself tends to copy existing patterns without questioning, it perpetuates and scales tech debt.
A cautious, nuanced case for AI optimism: why safety, interpretability, bias, and alignment matter as much as raw capability.
MIT researchers have developed a machine-learning approach that improves the accuracy of material simulations by constructing training datasets that capture the diversity of atomic environments in chemically disordered metal alloys, potentially accelerating materials discovery.
This article demonstrates how to run local AI inference on the 2017 AMD RX 580 GPU using the Vulkan backend of llama.cpp and stable-diffusion.cpp, without requiring CUDA or ROCm. It covers hardware setup, compilation steps, and performance results.
The EU Commission announced the winner of its “Frontier AI Grand Challenge” on Friday. The consortium “Europa”, led by the Italian company Domyn, prevailed in the competition. The alliance is to receive the necessary resources with the award to develop a state-of-the-art open-source AI model. The prestige project is intended to cover all 24 official EU languages and set a statement for the continent's technological ambitions.
Summer is an open-source local tool by Autumn for tracking AI coding tool usage and spend. It supports Claude Code, Codex, and OpenCode, requires no hosting, and provides a local dashboard to aggregate usage across a team per developer, model, and cost.
ByteDance, Tencent, MiniMax, Alibaba, and Xiaomi all cut AI token prices between 50% and 99% within the same competitive window. Bank of America Securities analysts attribute the pricing race to narrowing capability differences between China's major AI models. Alibaba's 50% discount on Qwen3.7-Max was tied to the 618 shopping event, blending AI competition with consumer promotions.
Large language models have moved from research labs into engineers' daily workflows. To help technical professionals stay ahead, IEEE offers a five-course online program that covers the engineering behind generative AI, from transformer architectures to deployment.
A research team including Huawei has successfully used Ascend 910C chips to complete post-training of the DeepSeek-V4-Pro model, marking a milestone in China's ability to perform complex AI training with domestic hardware. The project, involving over 1,000 chips and a 1.6 trillion parameter model, demonstrates a shift from inference-only capabilities to full training, bolstering China's AI self-sufficiency amid US sanctions.
PhD_fleet is a Python toolkit that lets a single researcher (advisor) spawn and converse with a fleet of Claude Code agents through Slack. Each agent has its own workspace directory, with Slack messages driving turns and the filesystem serving as long-term memory. A separate coach agent watches the advisor's mentoring and provides evidence-based feedback.
pm-claude-skills is an open-source library of 174 professional SKILL.md files for AI assistants, covering 18 professions. Each skill is eval-verified to produce professional-grade output usable as a first draft. Includes workflow recipes, skill memory, and cross-tool compatibility.
Qubot, our internal Copilot-powered analytics agent, allows any GitHub employee to ask questions about our data in plain language. Here's what we learned as we built it.
In the Thinkbench benchmark, GLM 5.2 led in correctness (92% full-pass) while MiniMax M3 was cheaper and faster. They performed similarly on code modification tasks, but GLM was steadier on greenfield builds. MiniMax tended to build more complete systems with ambiguous prompts.
Checkmarx releases a new SAST engine combining a deterministic rules-based scanner, an LLM trained on security data, and a third engine to classify findings as true or false positives before they reach developers. The company claims an F1 score of 0.499, far above the category average of 0.20, and found 327 true positives missed by a leading frontier model. The architecture emphasizes orchestration, automatically running three engines together to provide determinism, language coverage, and noise filtering.
Inferlay is a platform that helps developers choose the appropriate AI technology stack for their projects by comparing various tools and frameworks.