Evals: Translating Product Goals to ML Teams

Nick Turley, head of ChatGPT at OpenAI, outlines a distinctive approach to building AI products that prioritizes rapid shipping, learning from real-world usage, and iterative improvement over traditional product development cycles.

Core Principles of AI Product Development

Ship Fast, Learn Fast

• "This is a pattern with AI where you won't know what to polish until after you ship"
• "The only way to find out what people like and what's valuable is to bring it into the external world"
• "You're gonna be polishing the wrong things in the space... you won't know what to polish until after you ship"
• Shipped ChatGPT in just 10 days from decision to launch, despite many features not being ready
• Prioritize getting real-world feedback over perfecting features in isolation

"Is It Maximally Accelerated?"

• Use this question as a forcing function to understand what's critical path versus what can happen later
• "I just really wanna jump to the punchline of 'why can't we do this now?'"
• Creates a culture where teams constantly question if they're moving as quickly as possible
• Became a team meme with its own Slack emoji to push for faster execution
• Helps cut through blockers and bureaucracy, especially with people from larger companies

Empiricism Over Speculation

• "You really have to ship to understand what is even possible and what people want rather than being able to reason about that a priori"
• Real-world usage reveals unexpected use cases that wouldn't emerge in internal testing
• Models improve through exposure to real failure cases, not just benchmark testing
• "The benchmarks are increasingly saturated so really you need real world scenarios"
• Learning happens both in-product and through social channels (TikTok comments, social media)

Balance Speed with Safety

• Apply different processes for different contexts: "Process is a tool"
• Move quickly on product features, but maintain rigorous processes for safety
• "For frontier models there actually needs to be a rigorous process where you red team, work on the system card, get external input"
• Don't use speed as an excuse to skip critical safety evaluations

Treat the Model as the Product

• "There really is no distinction between the model and the product; the model is the product"
• Iterate on the model like a product by identifying key use cases and systematically improving them
• Focus on both capabilities and "vibes" - the personality and feel of the model
• Understand that model behavior is central to user experience, not just a technical component

Implementation Framework

1. Start with Open-Ended Exploration

• Begin with a broad, flexible interface rather than overly specific use cases
• Allow users to discover their own applications rather than prescribing them
• "ChatGPT feels a little bit like MS DOS; we haven't built Windows yet"
• Let usage patterns emerge naturally before optimizing

2. Use Evals to Communicate Product Goals

• "I started writing evals before I knew what an eval was"
• Evals are simply "articulating success before you do anything else"
• They become the "lingua franca" between product and research teams
• Clearly specify ideal behavior for various use cases
• Use evals to communicate what the product should be doing to researchers

3. Build Interdisciplinary Teams

• "The interdisciplinariness of really making sure that you put research and engineering and design and product together rather than treating them as silos"
• Hire for curiosity over specific AI experience
• "If a feature doesn't get 2x better as the model gets 2x smarter, it's probably not a feature we should be shipping"
• Think like a jazz band rather than an orchestra - ideas can come from anywhere

4. Follow a Three-Part Retention Strategy

• Model improvements (1/3): Systematically improve the model on use cases people care about
• New capabilities (1/3): Add research-driven features like search and personalization
• Traditional product work (1/3): Reduce friction, improve UI, and apply standard product practices

5. Polish After Learning

• "Shipping is just one point on the journey towards awesomeness"
• Once you understand what people are doing, there's no excuse not to polish
• Don't use velocity as an excuse for permanent roughness
• Follow through on refinement after initial learning

When to Apply This Approach

• When working with emergent technology where capabilities aren't fully understood
• In situations where user behavior can't be predicted in advance
• When the cost of delay exceeds the cost of imperfection
• When you need to establish product-market fit for a novel capability
• When you're competing in a rapidly evolving space