CHAPTERS
Live from DevDay: Spotlight on real-world builders
Andrew Mayne opens the episode from OpenAI DevDay, framing the conversation as a series of short interviews with developers using OpenAI tooling in production. The goal is to understand what changed with the latest announcements and how teams are translating tools into products people trust and rely on.
SchoolAI’s mission: safe, managed AI tutors for classrooms
Caleb Hicks explains SchoolAI’s focus on putting AI into students’ hands in a safe, managed way—more like one-time, guardrailed tutors than open chat. He emphasizes how model improvements over the past year unlocked better intelligence and lower costs, which is critical for education budgets.
Educator adoption journey: from bans to AI literacy to connected tutoring
Caleb outlines the typical progression schools follow: initial bans, then teacher productivity use, then recognition that students must learn AI to stay competitive. He argues the next step is classroom-connected tutoring that understands what students are doing in class and helps the whole system—teachers, families, and leaders—support students.
Inside the SchoolAI product stack: Dot, teacher tools, and tutor dashboards
SchoolAI offers a tuned assistant (“Dot”), form-based tools for outputs like lesson plans, and the more advanced managed tutoring experiences. The standout capability is letting teachers create guardrailed tutors and then view real-time dashboards showing student progress and needs.
Exit tickets and “GPS for impact”: identifying which students need help now
Caleb describes a concrete workflow: AI-powered exit tickets that quiz, coach, and perform social-emotional check-ins, rolling results into a teacher view. The system helps teachers triage attention—especially when responsible for hundreds of students—by flagging which students need immediate support.
DevDay tooling excitement: Agents SDK/Agent Builder and built-in safeguards
Caleb reacts to the Agents SDK and especially the Agent Builder’s drag-and-drop approach, including permissions and file search. He notes SchoolAI built similar internal tooling and is eager to replace parts with OpenAI’s platform—especially to help non-technical educators benefit without dealing with models and infrastructure.
MCP and integrations: turning partner ecosystems into safe classroom experiences
Caleb highlights momentum around MCP servers as a standard way to connect AI to tools and partners. OpenAI’s “line in the sand” around MCP makes it easier for SchoolAI to bring existing ChatGPT integrations into a managed, guardrailed school environment.
Evals and reliability at scale: why 2–3% matters
Andrew and Caleb discuss evaluation tooling, noting how small error-rate changes become huge at scale (millions of students). They point to DevDay demos showing how quickly agents can be built—and the need to make eval creation and monitoring equally easy to avoid deferring quality work indefinitely.
Jam.dev introduces “Please Fix”: edit websites like a doc and auto-create PRs
Dani Grant explains Jam.dev’s new tool, Please Fix, which lets non-engineers modify a site directly in the browser (like Google Docs/Figma) and submit changes as a clean GitHub PR that respects the design system. The pitch: eliminate bottlenecks where small fixes languish behind ticket queues.
A new way to experience the web: apps inside ChatGPT and “read, write, think”
Dani reacts to DevDay announcements as a shift in what browsers and the web mean, driven by ChatGPT apps and more agentic interaction. She connects this to usability: tiny UI tweaks matter, and tools that make iteration easy can raise product quality from “fine” to world-changing.
The Cambrian explosion of software: when anyone can build
Dani shares user stories (e.g., firefighters, church community builders) to illustrate how AI-enabled creation lowers barriers for non-traditional developers. Jam’s internal process is intensely user-centric—optimizing for a “wow” moment through constant feedback loops.
Abridge: ambient AI for doctor-patient conversations and paperwork relief
Zach Lipton describes Abridge as an AI platform for doctor-patient conversations that reduces documentation burden and returns attention to patients. He discusses “pajama time” and reports of saving an hour or more daily, plus emotional relief that reduces burnout.
High-stakes AI: defining hallucinations, detection pipelines, and trust-building
Zach explains that hallucinations in medical documentation are context-dependent—information can be plausible yet unacceptable if unsupported by the visit. Abridge uses an error ontology, model-based judging, and specialized pipelines to detect unacceptable statements with high recall, while emphasizing that trust is earned continuously via delivery and security commitments.
Cursor’s evolution: from autocomplete to coding agents, and how teams evaluate models
Lee Robinson discusses how Cursor moved beyond simple autocomplete into autonomous agents that can refactor, self-correct, and incorporate external info. Cursor dogfoods heavily, uses internal adoption as a filter for feature viability, and invests significant time evaluating models in real IDE workflows—where benchmarks alone don’t predict user experience.
Multiple models and rapid learning loops: custom autocomplete + online RL
Lee outlines Cursor’s “all-of-the-above” model strategy: foundational models for broad tasks plus Cursor-trained models for autocomplete. Cursor applies online reinforcement learning and can update autocomplete behavior frequently based on accept/reject signals from developers.
The next shift: making IDEs approachable, teaching context, and the future of software work
Lee describes a growing user base beyond professional engineers and a new UI direction that’s more agent-centric and less intimidating than traditional IDEs. He predicts AI will take on more of the non-coding parts of software engineering (on-call triage, bug handling, packaging/shipping), and argues education must add modern AI-tool literacy alongside CS fundamentals.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome