Key Concepts
AI agents, system prompts, user prompts, AI-powered software development, automation of repetitive tasks, tools, AI horseless carriage, user programmability, natural language programming, AI-native applications.
The Problem with Current AI Implementations
Pete Kumman argues that many current AI implementations, particularly in existing applications, are not fully leveraging the potential of AI. He contrasts the empowering experience of using AI coding tools like Cursor and Windsurf with the frustrating experience of using AI features in established apps like Gmail.
Example: The Gmail's "draftw writing agent" is criticized for producing generic, impersonal emails that require prompts nearly as long as the generated drafts.
Key Point: Kumman contends that developers are using "old software development techniques" and a "one-size-fits-all" approach, failing to recognize AI's capacity for user-specific customization and automation.
System Prompts: The Key to Personalization
The core of Kumman's argument revolves around the concept of system prompts.
Definition: A system prompt is a set of instructions given to an AI model that defines its role, behavior, and tone.
Problem: Current implementations often hide or restrict user access to system prompts, resulting in generic outputs.
Example: The Gmail draft writing agent likely uses a generic system prompt that instructs the AI to use a "formal businessy tone" and "correct punctuation," leading to emails that don't reflect the user's personal style.
Solution: Kumman advocates for allowing users to view and edit their system prompts, enabling them to personalize the AI's behavior and tailor its output to their specific needs.
Example: By changing the system prompt to "You're Pete, a 43-year-old husband, father, and YC partner," the Gmail agent generates a concise, informal email that sounds like something Pete would actually write.
Argument: Hiding the system prompt is akin to treating AI like traditional software, where the developer defines the system and the user interacts through a limited interface.
The AI Horseless Carriage
Kumman uses the analogy of the "AI horseless carriage" to illustrate the current state of AI development.
Definition: The AI horseless carriage refers to AI applications that simply replace existing functionalities without fundamentally rethinking the underlying design.
Example: Early automobiles that resembled horse-drawn carriages, failing to take advantage of the new technology's potential.
Argument: Developers are "slotting AI into" existing applications instead of designing AI-native tools that automate repetitive tasks and empower users.
The Promise of AI: User-Programmable Software
Kumman believes that the true potential of AI lies in creating software that users can program using natural language.
Key Point: AI should automate repetitive busy work, freeing users to focus on more important tasks.
Example: An email reading agent that automatically labels, archives, and drafts replies based on user-defined rules.
Process:
- The user defines a system prompt that instructs the agent on how to handle different types of emails.
- The agent uses tools to label, archive, and draft replies based on the system prompt.
- The user can refine the system prompt based on the agent's performance.
Argument: This approach empowers users to customize the AI to their specific needs and workflows.
Tools: Extending AI Capabilities
Definition: Tools are functionalities that AI agents can use to interact with the world and accomplish tasks.
Example: In the email reading agent, tools include the ability to label emails, archive emails, and write drafts.
Key Point: Tools enable AI agents to go beyond simple question-answering and perform actions on behalf of the user.
Example: A company called Den is building a platform that allows agents to chain together tools from different applications, such as Google Docs, email, and GitHub, to automate complex workflows.
Addressing Concerns about User Adoption
Kumman acknowledges that not everyone may be comfortable writing system prompts from scratch.
Argument: He believes that prompting will become more intuitive and accessible over time, similar to how computer usage became widespread.
Solution: AI-powered tools can assist users in writing and refining system prompts, providing a "system prompt writer sitting next to you."
Future Vision: In the future, AI will automatically generate and customize system prompts based on user behavior and feedback, with users only needing to intervene in rare cases.
Implications for Founders
Kumman encourages founders to rethink existing tools from the ground up with AI.
Key Question: Instead of asking "How do I insert AI into my tool?", founders should ask "How would I design this tool from scratch to offload as much repetitive work from the user as possible?"
Argument: AI-native tools will look different from traditional tools, focusing on automation, personalization, and user empowerment.
Synthesis/Conclusion
Pete Kumman's central argument is that current AI implementations are limited by outdated software development practices. By embracing user-programmable AI, allowing access to system prompts, and focusing on automation through tools, developers can unlock the true potential of AI and create applications that empower users to accomplish more with less effort. The future of AI lies in building "rocket ships for the mind," not just "horseless carriages."
AI summaries can miss context or contain errors. Check important details against the original video.
