Recent discussions in the AI community highlight the emergence of reasoning models that enhance the safety and performance of language models. OpenAI's Woj Zaremba notes that increasing compute at test time significantly boosts adversarial robustness, rendering some attacks ineffective. This advancement suggests that merely scaling model size is insufficient for optimal performance; rather, a focus on reasoning processes is essential. Additionally, a new standard for language models, termed Reasoning Tokens, has been introduced, allowing users to observe how models reason in real-time. This includes a standardized API across various thinking models, such as DeepSeek R1 and Gemini Thinking. The trend emphasizes the importance of 'thinking' time before providing answers, as analyzed by Davis Treybig, underscoring the potential of reasoning capabilities in language models.