What did OpenAI Announce at DevDay? Apps SDK, MCP UI & Impact to SaaS - EP99.20-APPS

Chris highlights the importance of recent developments for developers using ChatGPT, indicating that advancements in AGI are imminent. Simon shares his experience creating a ChatGPT app that integrates with Philips Hue lights, which has gained popularity among developers. He also discusses a Festivus app showcasing seasonal lights, which Chris finds intriguing.

During OpenAI's recent Dev Day, significant announcements included the introduction of apps in ChatGPT, functioning as Model Control Protocols (MCPs) with user interface components. The launch of AgentKit aims to enhance chat experiences in SaaS applications, while Sora 2's API release allows for affordable video production without watermarks. New models announced include GPT-5 Pro and a cost-effective mini model for voice experiences, alongside GPT Image 1 mini for lower-quality image generation.

Chris invites discussion on the ChatGPT apps SDK, emphasizing the importance of user input and output components in app development. OpenAI has standardized the rendering of output components, similar to React, facilitating complex UI creation. The new MCP standard developed by Facebook allows developers to return components as resource references, simplifying custom MCP building. The concept of "glass" suggests that if an MCP lacks a UI, a fast model can generate one based on outputs, enhancing user experience.

The new UI framework supports dynamic fields and controls, such as calendar entries, which can be saved and sent back to applications. This UI for MCPs is expected to become a standard due to its broad install base. However, early tests revealed challenges, particularly with the Canva app's confusing installation process and its often contextually irrelevant design suggestions.

The speaker critiques the current app SDK for limiting varied outputs across different modalities, arguing that focusing on a single MCP restricts research use cases requiring data from multiple sources. They express concern that these limitations hinder the ability to build context and perform tasks autonomously, likening the tools to early ChatGPT connectors that have since been forgotten.

The podcast discusses "human in the loop" interactivity, with skepticism about its effectiveness in enhancing task completion. The speaker envisions a future where AI operates autonomously, reducing the need for constant human input, and emphasizes empowering AI agents to set and achieve goals independently.

Current SaaS applications, such as Figma, are criticized for their limited functionality, raising questions about the necessity of traditional design tools if AI can generate complete app concepts. The conversation also touches on integrating ChatGPT-like experiences into SaaS applications, suggesting this could streamline features like note-taking.

The speaker critiques the current state of Multi-Channel Platform (MCP) protocols, arguing they prioritize brand protection over practical utility. Many MCPs are seen as ineffective, lacking proper definitions and failing to enhance productivity for expert workers. Concerns are raised about companies relinquishing user interactions to platforms like ChatGPT, potentially leading to a loss of control over user data.

The discussion concludes with a warning about a potential "great replacement" strategy, where emerging super apps overshadow existing platforms. The speaker notes that companies are motivated to attract daily users, often prioritizing consumer engagement over the needs of specialized industries.

The future of technology is anticipated to involve genetic CRUD and instantaneous generative UIs, emphasizing the importance of timely access to these tools. An experimental support agent is being utilized, integrating knowledge, prompts, and Managed Control Points (MCPs) to enhance support through SIM theory. The custom SIM MCP can perform tasks such as fixing accounts and updating email addresses, with a preference for AI handling these tasks over traditional support UIs.

Users are increasingly switching tools or requesting custom MCPs to boost productivity, although confusion exists around terminology, as MCPs are often misidentified as apps. Current AI models, such as GPT-5, are criticized for not being fully utilized, particularly in executing multiple parallel tool calls. Transparency in AI research is deemed essential, as current systems are viewed as black boxes lacking clarity in their methodologies.

The speaker expresses frustration with existing AI tools, particularly in presentation creation, noting that tools like Canva often yield unsatisfactory results. They argue that while AI can transform workflows, current tools may hinder productivity. The speaker emphasizes the need for tools that facilitate efficient email management and advocates for leveraging generic AI capabilities to create customized solutions that boost productivity.

While acknowledging AI's usefulness for certain tasks, the speaker notes it may not be suitable for all scenarios, such as medical research, and stresses the importance of selecting the right mix of tools. The discussion includes the relevance of personas in AI agents and introduces a new visual interface for building agents, akin to tools like Zapier and N8N. This visual canvas allows users to create agents through a drag-and-drop interface.

The speaker argues that flowcharts do not represent true agency, as they impose rigid procedures on agents. Concerns are raised about the target audience for the agent builder, with skepticism about whether the complexity of the visual editor will effectively empower business users.

The podcast discusses the release of the GPT-5 Pro model, highlighting concerns over its high cost, which is significantly higher than GPT-5. The discussion includes specifications for the GPT-5 Pro model, noting its high reasoning capability but slow speed. The speaker hopes for a decrease in the model's price over time to make it more accessible.

The podcast also evaluates the performance of Sora 2 and VO3, with the speaker expressing a clear preference for VO3 for serious video work. The launch strategy for Sora 2 appears to focus on generating hype through TikTok-style distribution. The episode concludes with a light-hearted farewell and a promise to return next week.