7 min read ai

Daily Productive Sharing 1246 - Build a Truly Useful AI Product

Daily Productive Sharing 1246 - Build a Truly Useful AI Product
Photo by Ivo Sousa Martins / Unsplash

One helpful tip per day:)

Chris Pedrega, founder of Granola, believes that if building a startup is like playing a hard video game, then building in generative AI is like playing that game at 2x speed:

  1. If you're not careful, you might spend weeks building a feature only to see it instantly automated by the next major AI model release.
  2. Everyone has access to powerful APIs and cutting-edge LLMs—so your “brilliant product idea” can be built by anyone.
  3. LLMs unlock unprecedented capabilities—like code generation and research assistance—but you need to surf the AI wave, not get crushed by it.
  4. LLMs are evolving faster than almost any technology in history.
  5. Don’t waste time solving “soon-to-vanish problems.” It sounds obvious, but it's hard to follow—because it goes against instinct.
  6. Predicting the future is now part of your job.
  7. The marginal cost of serving an additional user is the same—but running state-of-the-art AI is really expensive.
  8. Large companies with millions of users often can’t compete with you—there simply isn’t enough compute to scale bleeding-edge experiences.
  9. As a startup, you can deliver a Ferrari-tier experience using the most advanced, expensive models. You don’t need to optimize costs yet. Giants like Google can only offer Honda-tier service. The magic? Today’s Ferrari will be tomorrow’s Honda. Build the Ferrari now.
  10. Don’t treat AI models like obedient tools. Treat them like interns on their first day—success depends on giving them enough context so they can “think like you.”
  11. “Context window design” may become one of this era’s most important ideas—far beyond just AI.
  12. One of the biggest challenges today is that you’re competing with general-purpose assistants like ChatGPT and Claude that do most things pretty well.
  13. The only winning strategy: go ultra-narrow. Pick a very specific use case—and nail it.
  14. A world-class experience in a niche use case often has little to do with the AI itself.
  15. Being use-case-specific also helps you improve model performance more easily.
  16. When AI is right, it feels magical. When it’s wrong, it can be confusing—or downright eerie.
  17. Specializing makes it easier to identify the most common failure cases and handle them gracefully.
  18. Building in generative AI is like sprinting on a treadmill—while traditional tech is still strolling along.

If you enjoy today's sharing, why not subscribe

Need a superb CV, please try our CV Consultation

Granola 的创始人 Chris Pedrega 认为如果说创建一家初创公司就像是在玩一款高难度电子游戏，那么在生成式 AI 领域创业就像是以 2 倍速玩这款游戏：

  1. 如果你不够小心，可能会花好几周开发一个功能，却发现下一个 AI 模型发布时已经把它自动化了。
  2. 因为每个人都能接触到强大的 API 和前沿的大语言模型，你的“惊艳产品创意”也可以被任何人实现。
  3. 大语言模型（LLMs）开启了前所未有的产品能力，比如代码生成、研究辅助等——但你需要确保自己是“在 AI 浪潮上冲浪”，而不是被浪拍翻。
  4. LLM 正经历历史上最快的技术迭代之一。
  5. 不要浪费时间在那些“即将消失的问题”上。听起来很简单，但做起来很难，因为它违背直觉。
  6. 预测未来现在是你工作的一部分。
  7. 每新增一个用户的边际成本是一样的，而运行最前沿的 AI 模型真的很昂贵。
  8. 拥有上百万用户的大公司甚至无法与你竞争，因为全球根本没有足够的算力让他们在规模化时也能提供前沿体验。
  9. 作为初创公司，你可以为每位用户提供“法拉利级”的体验。用最贵、最前沿的模型。不必过于担心成本优化。像 Google 这样的巨头，最多只能为用户提供“本田级”的体验。美妙的是，就算你的用户规模呈指数增长，AI 推理的成本也在以同样的速度下降。今天的“法拉利”，将是明天的“本田”。趁你还可以时，做那辆法拉利。
  10. 不要把 AI 模型只当作“听指令的工具”，而应该把它们当作“刚入职第一天的实习生”。实习生能否成功的关键在于：你是否给了他们足够的上下文，让他们能“像你一样思考”。
  11. 他相信，“上下文窗口选择”将成为我们这个时代的核心理念之一，其影响远超 AI 领域。
  12. 如今构建 AI 产品的一大挑战是：你正在和 ChatGPT、Claude 等“通用 AI 助理”竞争，它们大多数事情都做得还不错。唯一的应对方式是走“超窄垂直方向”：选择一个非常具体的应用场景，并在该领域做到极致。
  13. 面向特定用例的“卓越体验”，往往和 AI 本身没太大关系。
  14. 专注于具体场景，也能让你更容易改进 AI 部分的表现。
  15. 当 AI 答得准时，它像魔法一样；但答错时，常常表现得令人困惑甚至毛骨悚然。
  16. 聚焦在特定用例上，更容易识别最常见的 AI 失误情形，并尝试避免或更优雅地失败。
  17. 在生成式 AI 领域做产品，就像在跑步机上冲刺，而传统科技仍在慢慢散步。

如果你喜欢的话，不妨直接订阅这份电子报 ⬇️

Dr Selfie

Dr Selfie