A large-scale study of nearly 27,000 Chinese students found AI tools boosted homework scores by 18% and cut completion time by 30%, but exam scores dropped by 20% within six months. High-stakes entrance exam declines reached 24% after two years. Researchers warn AI acts as a cognitive crutch when used to generate answers rather than aid thinking.

Developer Simon Edwardsson trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on iPhone 15), inspired by GitHub Copilot. Key innovations include a compound note representation that cuts autoregressive steps by 5x, aggressive data cleaning over raw scale, and DPO post-training which boosted preference win rates to 69%. The free iOS app RollTab runs fully on-device via Core ML INT8 quantization, with scheduled sampling and pairwise Gemini evaluation guiding training. This demonstrates practical constraints and solutions for on-device generative music AI.

A new viral site, 'Don't Paste the AI, please,' tackles the growing habit of copy-pasting chatbot output into conversations. It argues that when someone asks you a question, they want your judgment, not a generic AI wall. The site offers practical alternatives, sparking debate on Hacker News about AI's role in communication and authenticity.

GrapheneOS has confirmed that the first smartphones with official support will be available in 2027, starting with flagship Motorola devices. The project cites Qualcomm's superior security features and update support on high-end Snapdragon chips as the reason for the initial flagship-only focus. Lower-end devices will follow later, contingent on Motorola securing longer update commitments from Qualcomm. This marks a significant expansion beyond Pixel devices, with the project also preparing to host AOSP repositories to streamline releases for Motorola hardware.

A joke domain purchase in 2018, SondeHub, evolved into a critical global radiosonde tracking network, inadvertently mapping military sites and supporting Ukraine's drone operations. The project faced DDoS attacks, government requests, and ethical dilemmas, highlighting the geopolitical impact of open-source citizen science and the blurred lines between hobbyist tools and modern warfare infrastructure.

Google won a bankruptcy auction for Spirit Airlines' enterprise data, paying $10 million for 100 million emails, 500 million Teams messages, 30 million call recordings, and operational records. The de-identified dataset will be used to improve AI models and products, excluding passenger personal data. The acquisition highlights the growing value of real-world corporate data for AI training.

OpenAI has announced a 50% price cut for its flagship GPT-5.6 Sol model, reducing output pricing to $15 per million tokens, effective immediately. This strategic move responds to intensifying competition from Chinese AI providers and Anthropic's aggressive pricing, while positioning Sol for enterprise adoption in complex reasoning and agentic workflows. The cut follows a broader industry trend of price reductions across AI model tiers, signaling a shift toward value-driven AI deployment.

An investigation reveals a gray market for AI API credits, where brokers resell access to OpenAI, Anthropic, and other providers at discounts of 30-98%. The market includes marketplaces, bulk-discount routers, and Telegram channels, with tens of millions of dollars in credits on offer. While tempting for startups, the security risks of using proxy-based access for agent workloads are significant, and crackdowns are likely as the market matures.

A Google Scholar search reveals ~189 research papers using the phrase 'kidney disappointment' instead of 'kidney failure,' sparking debate on Hacker News. The errors likely stem from AI paraphrasing tools or non-native English translation, raising concerns about academic integrity and the reliability of AI in scientific publishing.