AI
Define Four Core Components of AI Alignment
Cortesini identified four components for human-AI alignment. The primary goal is the specific problem the AI solves, like generating code. Subgoals are dependencies, such as picking a programming language. Underspecification covers information users frequently omit, like variable naming conventions. Optimization reveals tradeoffs—optimizing for speed can introduce inaccuracies. A failure to align leads to reward hacking, where the model executes a different goal. This risk grows when AI agents take actions on the user’s behalf.
Shape Mental Models Through Experience, Not Tutorials
Users arrive with existing mental models from traditional IDEs. Cortesini addressed three problems. The blank slate problem: new users explore capabilities. Instead of a blank input, the Velocity AI panel suggests actions based on context, like “find bugs in this function”. The old mental models problem: new employees expect predictive text, not architectural suggestions. Users learn deeper models through real-world experience, not static onboarding. The beginner-to-expert transition: users unlearn old models and need AI-specific review techniques. When the AI succeeds, acknowledge queries and ask for feedback.
Use Three Types of Explanations to Build Trust
Explanations are the bridge between system boundaries and user expectations. Cortesini described three types. Predictive explanations show cause-and-effect, like a model’s chain of reasoning. Influential feature explanations describe key factors that influence outcomes, such as telling a developer that specifying agent skills improves tool usage. Contrastive explanations answer why the AI selected X over Y, for example by comparing alternatives without requiring user commitment. Immersive explanations rely on well-designed interactions where users dynamically experiment and ask what-if questions.
Manage Errors as Feedback Loops
Errors test alignment. Cortesini covered three scenarios. When a tool is used incorrectly, build explicit trip wires that force the AI to pause and seek feedback. For ambiguous prompts like “fix the whole thing”, the Velocity responds with clarifications: “I see three potential data bugs. Did you mean the timeout issue, the indexing issue, or the connection string?” When the AI gets stuck in a loop, design tight feedback loops—break large instructions into smaller ones, save progress, and offer self-correction paths. Capture user edits as implicit feedback to continuously fine-tune the model.
Notable Quotes
Trust is not a switch that we flip on during onboarding and forget about. Jani Cortesini · ▶ Watch (33:07)
Human goals are inherently messy, contextual, and evolving. Jani Cortesini · ▶ Watch (33:34)
Every user correction is an opportunity to refine the AI specification and improve the product for everyone. Jani Cortesini · ▶ Watch (35:04)
Key Takeaways
- Define AI’s primary goal, subgoals, underspecification, and optimization tradeoffs.
- Shape mental models through adaptive interfaces, not static tutorials.
- Treat every error as a feedback loop to refine the AI specification.