Rohit Krishnan on Training AI to Write Better

n this episode, Samuel Arbesman speaks with Rohit Krishnan, one of the most playful and insightful thinkers in the world of ideas. Rohit writes Strange Loop Canon, a newsletter devoted to exploring delightfully wide-ranging concepts—including artificial intelligence. He also recently collaborated with Jon Evans on Walter, a project that trained an AI to be good at social media.Together, Samuel and Rohit discuss Walter and the curious question of why large language models remain so poor at good writing, despite being built around text. Their conversation branches into topics such as reinforcement learning for writing, useful metaphors for understanding LLMs—like “fuzzy processors”—and whether emphasizing the alienness of AIs is a fruitful endeavor. They also touch on the idea of nurturing AI, how to use these systems in one’s work and life, and what all this means for the future of labor—perhaps a future that feels more like a video game. Above all, they explore the importance of continuing to play with these strange new tools.

There are no comments yet.

Please log in to write the first comment.