# I just want AI to rename my photos | The Vergecast Source: https://www.youtube.com/watch?v=wPSpqIh2xtc Recap page: https://rapidrecap.app/video/wPSpqIh2xtc Generated: 2025-11-30T13:32:47.982+00:00 --- ## Quick Overview Raycast's CEO, Thomas Paul Man, confirms that users can already rename photos on their desktop using AI prompts within Raycast, claiming they are "90% there" on achieving fully agentic workflows that integrate local context with various AI models. **Key Points:** - Raycast integrated GPT-3 shortly after its release in late 2022, initially focusing on allowing users to ask questions and search information using natural language prompts. - Thomas Paul Man believes the future involves an 'AI OS' where users prompt software into existence for short-term needs, potentially making traditional static extensions less relevant over time. - Raycast aims to orchestrate different AI models, abstracting complexity into an 'automatic mode' that selects the best model for tasks like image generation or deep research, while still offering configurability for advanced users. - For concrete, repetitive tasks like renaming photos, Paul Man argues that reliable, predictable software artifacts (like current Raycast extensions) are superior to the unpredictability of current AI models. - The company prioritizes user control and trust, implementing guardrails like requiring confirmation for disruptive actions (like deleting files) because of Raycast's deep system access, contrasting with more 'yolo' approaches in other AI integrations. - Paul Man cited a potential distraction-blocking feature that required constant screen analysis as something they decided against building due to privacy concerns, opting instead for users to manually define distractions. **Context:** The Vergecast episode features host David Pierce interviewing Thomas Paul Man, the founder and CEO of Raycast, a popular productivity launcher application for Mac, iOS, and Windows. This interview is part of a two-part series exploring how AI tool builders are integrating artificial intelligence into their software, moving beyond simply shoving ChatGPT into existing features toward creating more functional and useful experiences, contrasting successful integrations like text transcription with potentially nonsensical ones. ## Detailed Analysis David Pierce and Thomas Paul Man discuss the evolution of integrating AI into productivity tools, specifically focusing on Raycast's approach. Raycast was uniquely positioned to adopt natural language prompting due to its existing search-bar functionality, integrating OpenAI's models quickly in late 2022 to allow users to ask questions and search data. Paul Man described an ongoing effort to make AI agentic—using AI to interact with local files, calendars, and over 2,000 third-party extensions—though achieving reliable, complex agentic workflows remains challenging, often resulting in demos that fail in real-world use. Raycast is actively working on abstracting model selection, creating an automatic mode where the system intelligently routes prompts to the best available LLM for the specific task, like image generation or research, while retaining user control for manual selection. Regarding the core question of renaming photos, Paul Man asserts that this specific, concrete task requires predictable reliability, making a static, well-tested extension preferable over an unpredictable AI prompt. He emphasizes that for Raycast, which has deep system access, user trust and control are paramount; disruptive actions require explicit user confirmation, and they avoid collecting data unnecessarily, citing a planned screen-monitoring feature that was scrapped due to privacy implications, concluding that the best path forward involves building stable software artifacts for concrete tasks and using AI's open-ended nature for more nuanced, exploratory problems. ### AI Integration Philosophy - Paul Man notes the initial successful integration of GPT-3 was natural given Raycast's search box nature, allowing users to ask questions directly instead of relying on keywords for search engines like Stack Overflow - Early challenges involved overcoming hallucinations, leading to models that search the web and distill information in a specific tone. ### The Agentic Promise vs. Reality - The 'promised land' involves flipping computer usage from transpiling ideas into clicks/keystrokes to speaking or writing intent and letting the computer handle execution using local data and extensions - Paul Man admits that many complex agentic workflows do not work reliably yet, and developers must find the middle ground between functioning tools and science fiction dreams. ### Model Orchestration Strategy - Raycast supports multiple LLMs because users have personal preferences for different model tones and speeds - They are building abstraction layers to automatically route requests to the best model (e.g., one for image generation, another for research), aiming for an 'automatic mode' that hides complexity from the user. ### Raycast's Unique Contextual Position - Raycast has unparalleled context by accessing local files, extensions (over 2,000), and usage patterns, positioning them perfectly for 'contextual AI' that learns user behavior - This personalization layer allows the system to predict user intent, essentially building up knowledge similar to what the user has consumed. ### Reliability vs. Open-Endedness - For concrete tasks like file renaming, predictability is crucial; software artifacts (extensions) that work the same way every time are superior to unpredictable AI outputs - AI excels in open-ended tasks where the user does not know the single best solution. ### Privacy and System Access - Due to deep system access, Raycast must maintain higher quality bars and implement strong guardrails, ensuring users are prompted before disruptive actions - They prioritize user control, defaulting to requiring confirmation for actions like file deletion, though offering configuration to disable prompts for recoverable actions.