Avisha discusses the significant advancements in email generation from RNNs to modern models like GPT. By breaking down the email creation process into parts, she found that fine-tuning GPT on email data led to remarkably coherent outputs, even if they weren't as flawless as human-written emails. The conversation highlights the challenges of training data, particularly regarding the availability of examples for spear phishing, and emphasizes the transformative impact of larger models on text generation quality.
