YC
AI Apps Are Broken — Here's How To Fix Them
@ Y Combinator
05/23/25
Takeaways
PK
Pete Koomen
05/23/25
@ Y Combinator
We're using outdated software development techniques to build AI features, which prevents us from fully leveraging AI's potential.
PK
Pete Koomen
05/23/25
@ Y Combinator
The promise of AI is to enable users to program software using only natural language, transforming it into a powerful tool for creativity.
PK
Pete Koomen
05/23/25
@ Y Combinator
When I use AI tools like Cursor and Wind Surf, it feels incredibly empowering, allowing me to create anything I envision.
PK
Pete Koomen
05/23/25
@ Y Combinator
Conversely, using existing AI-integrated apps often feels like a chore, creating more work instead of simplifying tasks.
PK
Pete Koomen
05/23/25
@ Y Combinator
The integration of AI in Gmail, specifically the draft writing agent, demonstrates how powerful AI can be, yet it often frustrates users due to its hidden complexities.
PK
Pete Koomen
05/23/25
@ Y Combinator
The AI model Gemini is impressive, but its capabilities are often obscured by a frustrating user interface that doesn't maximize its potential.
PK
Pete Koomen
05/23/25
@ Y Combinator
The tone of AI-generated drafts often doesn't reflect the user's voice, leading to misunderstandings about the authenticity of the communication.
PK
Pete Koomen
05/23/25
@ Y Combinator
The system prompt that guides AI behavior is typically hidden from users, preventing them from customizing it to better reflect their personal style.
PK
Pete Koomen
05/23/25
@ Y Combinator
If users could edit the system prompt, they could tailor AI responses to better match their communication style, enhancing the overall experience.
PK
Pete Koomen
05/23/25
@ Y Combinator
AI app developers often treat system prompts like traditional code, maintaining a divide between users and developers that limits customization and personalization, leading to AI applications functioning as lowest common denominator software with generic outputs that lack individual expression.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The Gmail system prompt serves as a basic email writer, producing emails that no individual would authentically write, which limits its effectiveness.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The concept of the 'AI horseless carriage' refers to early automobile designs that merely replaced horses with engines without addressing fundamental design issues, illustrating how technology often fails to fully leverage new capabilities.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The initial search engines were simply digitized yellow pages, and early mobile apps were just websites wrapped in a native app, showing a slow adaptation to new technologies.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The Gmail team approached AI integration by asking how to slot AI into existing applications, rather than rethinking the application design to fully utilize AI's potential.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The true promise of AI lies in automating repetitive tasks, allowing users to offload busy work that doesn't require their full attention.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
By using simple techniques, users can create an email reading agent that organizes their inbox based on personalized instructions, demonstrating how AI can be tailored to individual workflows.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The programming for the email reading agent is accessible and intuitive, allowing non-programmers to effectively instruct AI on how to manage their tasks.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The shift in responsibility from developers to users regarding AI outputs allows for more powerful and flexible applications, as users can directly interact with and modify system prompts.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
As technology evolves, the ability to write effective system prompts will become more accessible, similar to how computer usage has become commonplace for everyone, but not everyone will want to or have the initiative to do so.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
After using Gmail for 20 years, I believe a good product could create a draft prompt based on my email history.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
When hiring an assistant, you don't write 30 pages of instructions; instead, you provide gradual training and feedback.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The interaction model for building AI tools should be iterative, where users and developers collaborate to refine system prompts.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
There's a missing tool in AI app development that would allow an AI system prompt writer to assist users in refining prompts.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
We need to stop treating AI prompts like black boxes; users should be able to see and understand what the AI is instructed to do.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
The promise of AI is to allow users to program software using natural language, making it more accessible and powerful.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
An email reading agent could handle transactional emails, freeing users to focus on more meaningful communication.
YC
Y Combinator Cast
05/23/25
@ Y Combinator
Founders should rethink existing tools with AI, designing them from scratch to offload repetitive tasks from users.