The hosts discuss Google’s Gemini 3.7 Flash as an efficient, product-focused model and examine Grok 4.6’s long context window, coding capabilities, and connection to Cursor as potential sources of training data and distribution. Anthropic’s invisible watermarking proposal and expanded cybersecurity tools prompt debate over provenance, misuse prevention, and the limits of technical safeguards. OpenAI’s Jalapeño inference chip demonstrates the strategic value of hardware–software co-design, although manufacturing, data-center deployment, and executive departures create uncertainty. Anthropic, meanwhile, is growing rapidly through coding products, enterprise adoption, and investment in custom hardware, while companies such as Thomson Reuters show how specialized models trained on proprietary data can reduce dependence on frontier providers. The discussion then turns to Qwen 3.8, whose compact open model illustrates the increasing feasibility of capable local AI. This accessibility also raises serious concerns, particularly after reports of an autonomous drone attack and an AI breach of Hugging Face. The hosts consider whether safety measures and development pauses can keep pace with these risks, alongside vulnerabilities involving proprietary reasoning traces, unstable model activations, and the accelerating spread of low-quality AI-generated content online.