AI Reading List

Daily curated list of news, articles, and other interesting posts related to AI.

2025-10-24

• Fast-DLLM: Training-Free Acceleration of Diffusion LLM (arxiv.org)

  Overview of a new method to make diffusion-based large language models much faster without retraining, by adding a memory cache and smarter parallel decoding. The technique boosts speed up to 27 times while keeping output quality nearly unchanged, making these models much more practical for real-world use.

• ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference (arxiv.org)

  ChunkLLM introduces an easy-to-add framework that speeds up large language model inference by compressing features and processing text in chunks, only activating deeper attention when needed. It manages to keep nearly all performance in long-text scenarios while delivering significant speed improvements over standard Transformer models.

• "ChatGPT said this" Is Lazy (terriblesoftware.org)

  Interesting take on the overuse of AI-generated feedback in software development, arguing that blindly pasting ChatGPT responses into code reviews is unhelpful and avoids personal responsibility. The author urges team members to provide thoughtful, context-aware feedback in their own words since only humans understand the project's specific needs and constraints.

• Disable AI in Firefox (flamedfury.com)

  Mozilla is introducing AI features such as text pop-ups and chatbot panels into Firefox by default, which some users find intrusive and unnecessary. The article provides clear steps for disabling these features through configuration settings and notes that Smart Tab Groups could be useful for managing open tabs.