It’s been a while since I’ve written. With a moment to spare today, I want to share my experiences from the past two years of AI development and entrepreneurship.
Now, in the second half of 2025, having led several AI projects including my own startup, Hika, I can make a bold claim: building AI products has become deceptively easy, but building an AI business is harder than ever.
Prompt is All You Need
This is an insight I shared in an article a year ago, and it holds truer than ever today. Prompt Engineering, RAG, MCP, Agents—all these fancy terms are just different ways to feed a large model a prompt with the right context.
Let’s break it down with some quick, informal definitions:
RAG (Retrieval-Augmented Generation): A way to get around a model’s context window limits and “attention loss” by searching large volumes of text for “semantically similar” information to stuff into the prompt.
MCP (Multi-modal Content Protocol): Essentially a set of standards for how different applications can provide information to a large model. It hasn’t solved any fundamental problems, but Claude’s example has encouraged more companies to open up their functions and data.
Agents: To be honest, I’ve always felt the definition of “Agent” is problematic. But for now, let’s just say it’s about using various tools to acquire more information—again, for the prompt.
My point isn’t that these techniques are trivial. It’s that at this stage, AI development is still all about the Prompt, Prompt, Prompt. And as the underlying models become more powerful, the clever little prompt tricks of yesterday are becoming increasingly useless.
In other words, any competent programmer can get started with AI development today and, with some real-world experience, will be no different from a programmer who started a year or two ago. Outside of training a foundational model from scratch, there are no insurmountable technical barriers.
This means AI applications have no technical moat.
Data is All You Want
So, if “Prompt is All You Need,” what is the irreplaceable part of the equation? Data.
Forget for a moment how much information you can extract from it or how “intelligent” the model is. Data is the only real moat. If you have it and your competitor doesn’t, it’s only a matter of time before you surpass them. Imagine an AI companion, an AI search engine, or an AI assistant built with all of your personal information. That is a fortress.
Business is All You Desire
With the above established, we can talk about business. Manus is incredibly popular right now, but let me ask you a question: is today’s Agent fundamentally different from the AI applications of a year ago? The capabilities of Manus are clear for all to see—a general-purpose agent that seems capable of anything, but in reality, does nothing particularly well. It can’t even write a decent research report.
Here’s another question: since the major LLMs integrated their own search functions, have you still been using AI search tools like Perplexity?
The reality is dawning on everyone. In the race to achieve a certain level of “intelligence,” general-purpose applications have a first-mover advantage. But the moment a major model provider reaches that threshold, it will enter the field and integrate that capability itself. GPT adding search and file imports is a perfect example. Claude launching Claude Code is another. And believe me, when models are smart enough to handle Agent task orchestration, the model providers will launch their own Agent products. Aside from Deepseek, there isn’t a single company that is content with just being a model provider.
So what about vertical domains? A vast world of opportunity, you might say. True, but only if AI can solve a real, painful problem.
And what about my company, https://hika.fyi ? We recognized the limitations of tools like Perplexity a year ago—most users will just stick with the native model. When a major tech company was considering acquiring Hika, I wrote an introduction explaining our strategy and how we were focusing on building better thinking tools. You can read more about that here: https://www.reddit.com/r/macapps/comments/1lpp4ip/for_pkm_nerds_finally_an_ai_tool_that_helps_you/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
These are just some scattered thoughts I’ve had recently. There are bound to be omissions and oversights. I hope to spark more discussion, so please feel free to connect and share your thoughts.