The Vergecast: "I Just Want AI to Rename My Photos"

Date: November 30, 2025
Host: David Pierce
Guest: Thomas Paul Mann (Founder & CEO, Raycast)

Episode Overview

This episode launches a two-part series exploring how makers of AI-powered software are thinking through the possibilities and pitfalls of integrating AI into their products. Host David Pierce talks with Thomas Paul Mann, CEO of Raycast, about the real and practical ways generative AI can change how we use our computers— focusing specifically on workflows, device control, and just how far we are from a world where AI automates all our digital busywork, like renaming your photos with a single prompt.

Key Discussion Points & Insights

1. The State of AI in Consumer Apps

(00:32)–(02:47)

• The tech industry is racing to embed AI into everything, sometimes unnecessarily.
• Genuinely useful applications exist, such as reliable transcription (OpenAI's Whisper), organizing tools (Todoist "Ramble"), and AI-centric search (MyMind).
• The central question: Where should AI augment our workflows, and how do we balance utility, complexity, and privacy?

2. Introducing Raycast and Its AI Ambitions

(05:58)–(07:42)

• Raycast originated as a Spotlight alternative for Mac, allowing quick app launching, text expansion, and window management.
• Recently, Raycast added the ability to run AI models and even let AI interact with your apps and files— for instance, asking Raycast to grab browser tabs or find recent files.
• The app is uniquely positioned, sitting at the intersection of user commands, app extensions, and now, AI-driven automation.

“Raycast itself is like a massive search box ... so you can just type something in … and it just used to be very static text ... but then it felt quite natural to extend this and just put in natural language like a prompt and then get going.”
— Thomas Paul Mann (06:44)

3. Early AI Integrations and Evolving Use Cases

(07:42)–(09:55)

• Early days: Integrated OpenAI's GPT-3 right after it launched—noting demand from users for more natural ways to interact with their devices.
• Initial predictable tasks: answer questions, summarize content, search the web.
• Obstacles: Model hallucinations, finding solutions for accuracy, and layering in web and local data access.

“As with every new technology, people kind of adapt to what’s possible and then they take the next step and push the boundaries.”
— Thomas Paul Mann (09:35)

4. To Build or Integrate? The Model Dilemma

(09:55)–(13:21)

• Raycast decided not to build its own LLM, but rather integrate multiple external models (OpenAI, etc.), as users have varying preferences and use cases.
• Some tasks require fast models, some need more depth—model choice is both technical and personal.
• Raycast fine-tunes prompts and workflows for reliability in their context.

5. The Push Towards Agentic Workflows

(13:21)–(18:51)

• Raycast envisioned early on an "agentic" system where users simply describe tasks and the computer acts.
• Reality: Reliable agentic workflows are difficult—AI remains unpredictable. Hybrid UI (text box + suggested actions) is necessary.
• Discoverability is a challenge: users must both trust and understand what’s possible with natural-language prompts, echoing earlier voice assistant frustrations.

“Prompting is still like a skill thing ... as those systems become much more proactive, I think this will be better.”
— Thomas Paul Mann (17:39)

6. Model Routing and Orchestration

(18:51)–(22:44)

• Current frustration: Users must pick the best AI model for each task.
• Ideal future: Raycast (and similar tools) should auto-route prompts to the most appropriate model for each use case—abstracting away complexity while letting advanced users customize.
• Orchestration across access points (apps, files, extensions, and models) is where Raycast aims to differentiate.

“We started basically now abstracting that away ... the best experience is you sort of have an automatic mode which just does what you want.”
— Thomas Paul Mann (20:14)

7. The Path to Reliable, Everyday AI

(24:47)–(29:53)

• Sustainable AI integration is about building out incremental, reliable workflows—avoid science-fiction "100% agentic" promises that break on first use.
• Focus on improving repetitive, mundane tasks like reformatting text, spelling corrections, and—crucially—file operations like naming and sorting photos.
• The “last ten percent” of reliability remains elusive; demos can mislead, but even 90% solutions provide value.

“We’ve all seen the shiny demos in launch videos and then they fall apart the moment you use it ... It’s science fiction, right?”
— David Pierce (26:07)

Memorable Segment:

Renaming Your Photos or Cleaning Up Files—Are We There Yet?
(29:53)–(33:26)

• David underscores the real-world need for AI-powered batch operations, like renaming photos based on date/content—a task Raycast is “90%” able to do now.
• The vision: The rise of disposable, personalized tools called into existence via prompts, rather than ever-fatter utility catalogs.

“You can do this today in Raycast, we have that ... And then the 90%—every now and then, it doesn’t work.”
— Thomas Paul Mann (31:26)

“What if you could have this app just by asking AI and it builds this little app for you, and then you have it for yourself ... Software becomes malleable and you can change it ad hoc and it becomes just what you want.”
— Thomas Paul Mann (32:10)

Notable Quote

On AI’s role at the OS Level:

“If you purely think from a user’s standpoint, AI should be on the operating system level. It just makes so much more sense to be there instead of in every app ... from a user’s point, the best thing is if you have a smart operating system that helps you to get your job done.”
— Thomas Paul Mann (61:22)

Safety, Privacy, and the Human in the Loop

(43:02)–(50:35)

• Raycast takes extra care with security and trust due to its deep system access; users are alerted before destructive actions, human override is always possible.
• Not all AI-powered features should be built. Example: discarding plans for an AI to always-actively monitor the screen for distractions due to privacy and user unease.
• It's a value exchange—users will give up data for utility, but developers must responsibly draw lines and retain user control.

“We put a lot of effort into making [Raycast] super stable. ... For that it’s even more important to have the guardrails right.”
— Thomas Paul Mann (44:43)

The AI-First Workflow and the End of Apps?

(53:11)–(62:10)

• Thomas: “My brain is completely rewired and it’s like I’m prompt first by now ... I basically just put things [into Raycast] ... and then iterate on that.”
• Raycast is used to orchestrate across browser tabs, apps, note-taking, and even code snippets—lowering the barrier for non-coders to automate or script computer behavior.
• The once-clear boundary between extensions, mini-apps, and AI-driven workflows is blurring: sometimes the artifact should be concrete and repeatable; sometimes open-endedness is a strength.
• Designers at Raycast now code; creativity is accelerated by AI.
• But: Discoverability and reliability still need work. Open-ended prompts are powerful, but for true productivity, repeatable tools are still king.

Closing Thoughts

Both Pierce and Mann agree the industry often races for an “end state” vision (fully autonomous computers, etc.), but there's value—and a real user need—in practical, incremental AI that augments, not replaces, human control. For now, the best AI-powered experiences will be context-aware, OS-level, user-driven, and above all, helpful for repetitive or complex-but-patterned digital tasks.

Key Timestamps

• 00:32–02:47: Why AI everywhere? What works and what’s silly.
• 05:58–07:42: Raycast’s roots and AI ambitions.
• 13:21–18:51: The challenge (and limits) of agentic AI.
• 18:51–22:44: Why we need AI to auto-route tasks to best models.
• 29:53–33:26: AI for renaming photos—how close are we?
• 43:02–50:35: Privacy, guardrails, and where Raycast draws the line.
• 53:11–62:10: Living “prompt first”; the blurring of apps and AI workflows.

Notable Quotes

• “Raycast itself is like a massive search box … it felt natural to extend this and just put in natural language like a prompt and then get going.” — Thomas Paul Mann (06:44)

• “We started basically now abstracting that away … the best experience is you sort of have an automatic mode which just does what you want.” — Thomas Paul Mann (20:14)

• “We’ve all seen the shiny demos in launch videos and then they fall apart the moment you use it ... It's science fiction, right?” — David Pierce (26:07)

• “If you purely think from a user's standpoint, AI should be on the operating system level. … From a user's point, the best thing is if you have a smart operating system that helps you to get your job done." — Thomas Paul Mann (61:22)

Summary by The Vergecast Summarizer — Listen to the full show for deeper insights on the future of AI-powered productivity and the real, messy ways it could change your digital life.