AI Reading List

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

2025-10-22

• The security paradox of local LLMs (quesma.com)

  Local AI models like gpt-oss-20b are highly susceptible to prompt and code injection attacks, often generating dangerous backdoors or executing malicious code during everyday developer workflows. Smaller models are much less capable of detecting manipulation compared to larger, frontier models, posing serious security risks to users who rely on local deployments for privacy and control.

• AI assistants misrepresent news content 45% of the time (www.bbc.co.uk)

  A large international study found that AI assistants misrepresented news content 45 percent of the time, with this issue occurring consistently across languages and regions. The research highlights significant concerns about the reliability of AI-powered summaries and responses.

• Meta is axing 600 roles across its AI division (www.theverge.com)

  Meta is cutting around 600 positions in its AI division, focusing layoffs on its long-standing research teams while continuing to hire for its new superintelligence-focused TBD Lab. This shift reflects Meta’s strategy to prioritize applied AI projects and large-scale model development over foundational research.

• Look, Another AI Browser (manuelmoreale.com)

  Many new "AI browsers" are just repackaged versions of Chromium with some AI features layered on, lacking real innovation in browser design. Interesting critique of the tech industry's tendency to market incremental changes as groundbreaking, even when leading AI companies are involved.