The AI Revolution Marches On

Episode Notes

The episode opens with Andrew detailing OpenAI's Dev Day announcements, especially the real-time API for continuous text or audio conversations and demos aimed at customer support and phone ordering. The hosts then debate AI as a replacement or augmentation for customer service, with Brian and Justin emphasizing how frustrating human support can be and how useful a capable AI agent might be if it can actually solve problems.

Andrew walks through several additional OpenAI features, including prompt caching, easier fine-tuning, memory controls in ChatGPT, and the new Canvas document-editing mode. The conversation also covers a study comparing doctors and GPT-4 on diagnostic tasks, Meta's new video model and the Movie Gen demos, and then shifts into picks where Justin praises Agatha All Along, Brian promotes Achewillow, and Andrew recommends Only Murders in the Building.

Key topics

• Real-time conversational APIs for voice and text: Andrew explains OpenAI's real-time API for nonstop voice or text conversation and compares it with similar work from other companies like Deepgram.
• AI as a replacement or augmentation for customer support: The hosts describe AI as potentially better than scripted human support if it can resolve issues quickly on either the calling or receiving end.
• Superhuman persuasion and AI negotiation risks: Andrew raises the possibility of highly persuasive AI being used for manipulation, while also imagining AI lawyers and negotiators.
• Prompt caching and cheaper repeated context: Andrew explains that caching repeated prompt context can cut token costs roughly in half for apps that resend long instructions.
• Fine-tuning custom models from conversation data: Andrew says fine-tuning is now simple from the dashboard and that the main challenge is collecting good training data.
• ChatGPT memory controls and persistent personalization: The hosts discuss ChatGPT's memory feature, how it can be viewed in settings, and how it can be edited or deleted.
• ChatGPT Canvas vs. Anthropic Artifacts: They compare OpenAI's Canvas to Anthropic's Artifacts, focusing on document editing, code workflows, and how different users may benefit from each.
• AI-assisted writing as an iterative process: Andrew argues that AI writing works best when humans add ideas first and then use the model to polish and refine the output.
• Doctors vs. AI in medical reasoning: The episode discusses a diagnostic study involving doctors, GPT-4, and doctors using GPT-4, using it to argue that AI is already in the range of doctors on some tasks.
• The current limits of video generation: Andrew notes that video models still struggle with detailed, complex scenes and text-heavy visual elements.
• AI video generation and its practical uses: The hosts discuss Meta's Movie Gen demos, including scene replacement, personalized videos from a supplied photo, and VFX-style edits.
• Streaming television economics and bloated production budgets: Justin and Andrew argue that some streaming genre shows became too expensive and still looked cheap, while lower-budget shows like Agatha All Along can work well.
• Executive bloat and creative inefficiency at Disney: Andrew suggests that too many executives and middle managers may be inflating costs without improving what appears on screen.

Picks

• Justin Robert Young: Agatha All Along — Justin explicitly says he really enjoys the show and recommends it as a fun, lower-budget Marvel series with strong production design and a clear vibe.
• Brian Brushwood: Achewillow — Brian clearly promotes the pre-order for the book version of Achewillow and describes it enthusiastically as a friend-of-the-show project.
• Andrew Mayne: Only Murders in the Building — Andrew directly calls this his pick and says he has been enjoying it, though he notes it has gotten weirder than usual.