AI Reading List

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

2025-10-21

• Wikipedia says traffic is falling due to AI search summaries and social video (techcrunch.com)

  Wikipedia’s human traffic has dropped 8 percent year-over-year, which the Wikimedia Foundation attributes to generative AI search summaries and the rise of social video platforms drawing users away from the site. There are concerns that fewer visitors could mean less content improvement and support, so the foundation calls for better attribution and urges AI platforms to encourage users to visit Wikipedia directly.

• Should LLMs just treat text content as an image? (www.seangoedecke.com)

  Interesting exploration of the idea that multimodal AI models might process text more efficiently as images, leveraging concepts like optical compression to potentially increase data throughput. The post discusses implications for language model training and hints at parallels between human memory and how AI could handle large text contexts visually, though practical challenges remain.

• Neural audio codecs: how to get audio into LLMs (kyutai.org)

  Audio is much harder for language models to process than text because it involves vastly more data points and requires temporal coherence, making simple approaches like sample-by-sample prediction slow and incoherent. Neural audio codecs compress audio into discrete tokens, allowing language models to handle audio more efficiently, improving understanding and generation of speech beyond what current wrappers or native speech models can achieve.

• LLMs can get "brain rot" (llm-brain-rot.github.io)

  Continual exposure of large language models to low-quality and engagement-focused internet content leads to persistent declines in their reasoning, context understanding, safety, and alignment traits. Importantly, even when models are retrained with higher quality data, the negative effects linger, highlighting data quality as a key concern for the long-term reliability of AI systems.

• ChatGPT Atlas (chatgpt.com)

  ChatGPT Atlas is a macOS app that brings ChatGPT directly into your web browsing, helping you summarize content, compare products, and analyze data from any site. Privacy controls and features like agent mode allow users to manage what ChatGPT can access and even automate tasks such as research or shopping.