How To Design Better AI Apps (AI column)
Interesting. interview Pete Koomen, general partner of Y Combinator and founder of Optimizely, discusses ideas from his essay.AI Horseless Carriages", in which he created a demo using AI to work with mail from scratch. In fact, this interview deals with the following topics:
1. The problem of AI-wheelchairs without horses Many developers use old software development approaches to create AI functions without revealing the full potential of the technology. As Koomen notes, “We’re using the old software development mentality, the old techniques to create these features, and we’re not really using all of what AI can do.” 2. Hidden system failures limit opportunities Gmail AI creates formal, faceless emails because of a hidden system prompt that users cannot see or change. This results in an AI assistant writing as a “universal email writer for everyone” rather than a specific user. 3. Users should have access to system prompts The ability to edit system prompts allows you to personalize AI for a specific user. Koomen demonstrates in his demo how changing the prompt from universal to personal dramatically improves the result. 4. Overcoming the developer-user division Traditional software requires separation: the developer writes the code, the user uses the interface. AI allows users to “program” applications in natural language, which changes this paradigm. 5. AI-agents for reading letters more effective than chatbots Koomen shows an example of an email agent that can automatically sort, tag and respond to emails based on user instructions. This is a more powerful approach than simple chatbots. 6. Importance of tools (tools) for AI agents AI agents become truly useful when they have access to tools to perform real-world activities. Developers should focus on building a rich ecosystem of tools. You can remember here. MCP protocol, which is now popular for allowing agents access to external tools n 7. Coding agents show the right direction Coding agents like Cursor and Windsurf perform better than other AI applications because they give full access to the model’s capabilities without restrictions. (or almost without limitation.) 8. Chatbot paradigm limits AI Most products simply embed chat agents into existing interfaces, which doesn’t take advantage of AI’s true capabilities. (We don't know what system bumps are on the chatbot side.) 9. Anyone can be a prompt engineer. Writing prompts is intuitive – it’s just an AI explanation of how you make decisions. “It’s not a skill we’re born with, but it’s pretty intuitive.” 10. Rethinking products from scratch for AI Startup founders should not be asked, “How do I put AI into my tool?” but “How would I design this tool from scratch to automate the repetitive user experience as much as possible?”
All in all, it’s an interesting look at product creation, but by giving access to AI model settings to everyday users, we find ourselves in a dangerous world – imagine that to drive a car you would have to assemble and customize its engine and running gear. But to tune the model, based on artifacts and user behavior, is much closer to a convenient and safe product, if you continue the analogy with cars, it would mean adjusting the configuration of the car to your typical behavior as a driver:)
#AI #ML #Software #DevEx #Engineering #Development