Nov 27, 2025 • Perfomance & AI Day

AI is an Electric Bike for the Brain - Stoyan Stefanov

Stoyan Stefanov calls AI an electric bike for the brain. But he warns it can also lead you straight off a performance cliff.

#1about 4 minutes

Analyzing the performance of AI-generated code

A practical example shows how an AI-generated UI with Vercel's V0 resulted in slow interactions and significant layout recalculation issues.

#2about 2 minutes

Navigating the emotional stages of adopting AI

Developers often move through stages of grief, from initial panic about job replacement to eventual acceptance and productivity with AI tools.

#3about 3 minutes

Using AI for prototyping and research in performance work

AI excels at creating isolated test pages and prototypes but is less suitable for shipping the meticulously researched code required for production performance optimizations.

#4about 6 minutes

Viewing AI as an electric bike for the brain

The metaphor of an electric bike illustrates how AI helps developers explore more complex features and ideas faster, reducing the dread of tedious tasks.

#5about 7 minutes

Critically evaluating AI-powered performance analysis tools

Using Google's MCP tool reveals that AI can identify some performance issues but may also hallucinate problems, highlighting the need for developers to ask critical follow-up questions.

#6about 2 minutes

Why developers may need their own custom AI models

Frustration with the growing restrictions and guardrails on commercial AI platforms motivates the exploration of building custom, domain-specific models.

#7about 7 minutes

Building a custom AI model to analyze web performance

A project called 'AI-slow' demonstrates how to build a custom performance analysis model using tabular data from the HTTP Archive and gradient boosting algorithms.

#8about 10 minutes

Training a model and interpreting its predictions

The process involves training a model on performance data and using techniques like SHAP to explain which features, such as page size, most impact its predictions.

#9about 2 minutes

Understanding the limitations of custom AI models

The custom AI model can provide useful suggestions but is not yet a replacement for an expert because its accuracy is fundamentally limited by the quality of its training data.