The hosts argue that Fable is best used as an expert model for high-value tasks, such as auditing work plans and catching errors, while cheaper models handle routine execution. They introduce compound engineering as a method to close verification loops, noting that AI agents have made building cheap but shifted costs to verification. The discussion also covers Microsoft Frontier Co and the trend of embedding AI engineers into businesses for hands-on consulting. A new plugin called Caveman is highlighted for reducing token usage by 65% through terse prompting. The episode concludes with a debate on Meta's Project Cannes, where contractors posed as underage girls to stress-test competitors' AI models, raising questions about the ethics and oversight of such safety testing.