Live from DevDay — the OpenAI Podcast Ep. 7

Quick Overview

The OpenAI Podcast live from DevDay featured interviews with developers from School AI, Jam.dev, and Abridge, focusing on how recent OpenAI announcements, particularly the Agent Builder and SDKs, accelerate their work in education, web development tooling, and healthcare documentation by reducing orchestration overhead and enabling faster iteration.

Key Points: Caleb Hicks from School AI emphasized that model progression provided leaps in intelligence and cost improvements, enabling them to offer safe, managed AI tutors to students, moving educators from banning AI to recognizing that every student must know how to use it. School AI utilizes a three-part product: a basic AI assistant tuned for schools (Dot), tools for creating structured outputs like lesson plans, and the special feature of creating guard-railed, safe AI tutors with real-time teacher dashboards. Danny Grant from Jam.dev announced 'Please Fix,' a tool allowing non-engineers to edit a live site like a Google Doc via a browser extension and submit clean, design-system-compliant pull requests, aiming to eliminate bottlenecks in development. Both School AI and Jam.dev highlighted the value of using OpenAI's tools internally to build their own products faster, echoing the sentiment that internal tool creation often yields the best products, contrasting with traditional startup advice to 'do things that don't scale.' Zach Lipton of Abridge, an AI platform for doctor-patient conversations, shared that their tool saves doctors up to an hour or more daily, alleviating 'pajama time' documentation and leading to profound emotional relief, with one doctor citing the tool was 'saving my marriage.' Zach expressed excitement for the Agent Developer Kit creating a common platform for orchestration, allowing Abridge to focus more on content and specialized models, especially concerning hallucination management where they achieve about 97% recall in error detection. Lee Robinson from Cursor noted that while initial hopes were that code completion would 'solve code,' the reality requires more complex agents, and their internal culture prioritizes dogfooding, measuring feature success by internal adoption rates before external release.