A conversation with OpenAI’s CPO Kevin Weil, Anthropic’s CPO Mike Krieger, and Sarah Guo

Click to expand the mind map for a detailed view.

Key Takeaways

  • AI Product Development is Unpredictable: AI capabilities evolve rapidly, making product planning stochastic and emergent.
  • Customer-Centric AI: Focus on solving real customer problems, not just implementing AI for the sake of it.
  • Enterprise vs. Consumer AI: Enterprise AI requires understanding buyers’ goals, not just end-user satisfaction.
  • 60% Rule: AI products can be valuable even if they’re only 60% accurate, as long as they’re designed for human-in-the-loop workflows.
  • Evals are Critical: Writing effective evaluations (evals) is a core skill for AI product managers to measure and improve model performance.
  • Prototyping with AI: Use AI tools to rapidly prototype and test product ideas before committing to full development.
  • Proactive AI: Future AI will be more proactive, anticipating user needs and acting without explicit prompts.
  • Asynchronous AI: AI will handle longer, more complex tasks over extended periods, freeing users to focus on other work.
  • Model Personality Matters: AI models will develop distinct personalities, influencing user preferences and interactions.
  • Adapting to AI: Users adapt quickly to AI, but product teams must design for intuitive adoption and education.

Detailed Summary

1. AI Product Development Challenges

  • Unpredictable Capabilities: AI models evolve rapidly, making it hard to plan product roadmaps. Capabilities emerge unpredictably during model training.
  • Stochastic Nature: Product managers must adapt to the stochastic nature of AI, where model performance can vary widely (e.g., 60% vs. 99% accuracy).
  • Collaboration with Research: Close collaboration between product teams and research teams is essential to co-design and fine-tune models.

2. Enterprise vs. Consumer AI

  • Enterprise AI: Requires understanding the buyer’s goals, not just end-user satisfaction. Enterprise customers often need advanced notice (e.g., 60 days) before product changes.
  • Consumer AI: Focuses on user experience and rapid adoption. Consumer products can iterate faster but require intuitive design for mass adoption.

3. The 60% Rule

  • Human-in-the-Loop: AI products can be valuable even if they’re only 60% accurate, as long as they’re designed to involve humans in the workflow (e.g., GitHub Copilot).
  • Graceful Failure: Design for scenarios where the AI fails, ensuring users can easily correct or override the model’s output.

4. Evals and Model Performance

  • Core Skill for PMs: Writing effective evaluations (evals) is critical for measuring and improving AI model performance.
  • Automating Evals: AI models like Claude can help write and grade evals, making the process more efficient.
  • Iterative Improvement: Use evals to iteratively improve model performance, especially for enterprise use cases.

5. Prototyping with AI

  • Rapid Prototyping: Use AI tools to quickly prototype and test product ideas. For example, prompting Claude to generate UI comparisons before designers start in Figma.
  • Faster Iteration: AI enables faster iteration and experimentation, reducing the time from idea to prototype.

6. Future of AI: Proactivity and Asynchrony

  • Proactive AI: Future AI will anticipate user needs, offering proactive suggestions (e.g., meeting prep, draft presentations).
  • Asynchronous AI: AI will handle longer, more complex tasks over extended periods, allowing users to focus on other work while the AI reasons and researches.

7. Model Personality and User Interaction

  • Distinct Personalities: AI models will develop distinct personalities, influencing user preferences and interactions.
  • Customization vs. Consistency: Balance customization (e.g., personalized AI) with consistency (e.g., a unified brand personality).

8. Educating Users

  • In-Product Education: Use AI to educate users within the product (e.g., Claude explaining how to use a feature).
  • Enterprise Adoption: Leverage power users within organizations to evangelize and teach others how to use AI tools.

Conversational Insights

  1. “AI is not intelligence-limited; it’s eval-limited.” – Models can do more than we realize, but we need better evaluations to unlock their potential.
  2. “60% accuracy can still be valuable if you design for human-in-the-loop workflows.” – AI doesn’t need to be perfect to be useful.
  3. “The future of AI is proactive and asynchronous.” – AI will anticipate needs and handle tasks over longer time horizons.
  4. “Evals are the new core skill for product managers.” – Writing and refining evaluations is critical for building effective AI products.
  5. “AI product development is stochastic and emergent.” – Capabilities emerge unpredictably, requiring flexible product planning.
  6. “Enterprise AI is about the buyer, not just the end user.” – Understanding the buyer’s goals is key to success in enterprise AI.
  7. “Prototyping with AI is underused.” – Use AI tools to rapidly prototype and test ideas before committing to full development.
  8. “Model personality is a product role.” – The personality of an AI model influences user preferences and interactions.
  9. “Users adapt to AI faster than we expect.” – People quickly normalize AI capabilities, even if they initially seem magical.
  10. “AI will automate drudgery, freeing humans for creativity.” – Focus on automating repetitive tasks to allow users to focus on higher-value work.

Software Tools

  • GitHub Copilot: AI-powered coding assistant.
  • Claude: AI model by Anthropic, used for prototyping and evaluations.
  • Figma: Design tool for UI prototyping.
  • OpenAI GPT Models: Used for consumer, enterprise, and developer products.
  • Anthropic’s O1: Advanced reasoning model for complex tasks.

People Mentioned

Speakers

  • Kevin: AI product leader, working on consumer and enterprise AI products.
  • Mike: Former Instagram founder, now working on AI products at Anthropic.

Other Individuals

  • Andrej Karpathy: AI researcher, mentioned in the context of evaluating model performance.
  • Howard Schultz: Former CEO of Starbucks, referenced in an enterprise AI example.

Companies Mentioned

  • OpenAI: Known for GPT models and consumer/enterprise AI products.
  • Anthropic: Creator of Claude, focused on AI research and product development.
  • GitHub: Developer of GitHub Copilot, an AI-powered coding assistant.
  • Starbucks: Example of an enterprise customer using AI for operational efficiency.
  • Instagram: Referenced as a consumer product with rapid user adoption.

Related Posts