Lecture by Andrew Ng at Snowflake Build

Introduction

• Andrew Ng discusses AI's biggest opportunities.
• AI compared to electricity as a general-purpose technology.
• Focus on creating new applications with AI technology.

AI Stack and Opportunities

• AI Stack Levels:
  1. Semiconductors
  2. Cloud infrastructure tools (e.g., Snowflake)
  3. Foundation models and trainers
  4. Application layer (focus for generating value and revenue)
• Application layer viewed as offering the best opportunities.

Fast Machine Learning Development

• Growth in generative AI accelerating ML model development.
• Traditional supervised learning models took 6-12 months to build.
• Generative AI allows for developing prototypes in days.
• Enables fast experimentation, iteration, and invention.

Bottlenecks in Evaluation

• Testing (evals) becoming a bottleneck.
• Shift towards parallel rather than sequential data and prototype development.
• Innovations needed in building evals.

Prototyping and Development Speed

• Machine learning prototyping is fast but other steps like integration still take time.
• Pressure on organizations to speed up entire development processes.

Responsible Innovation

• Emphasis on "Move fast and be responsible."
• Rapid prototyping and testing to avoid causing harm.

Agentic AI Workflows

• Most exciting technical trend: Agentic AI.
• Agentic AI involves iterative processes for better output.
• Examples of Agentic Workflows:
  • Processing legal documents, healthcare diagnostics, compliance.
  • Visual AI applications (image and video processing).

Design Patterns in Agentic Workflows

• Reflection: Critique and improve outputs iteratively.
• Tool Use: LLMs making API calls and performing tasks.
• Planning: Sequential execution of complex tasks.
• Multi-Agent Collaboration: Different LLM roles for improved task performance.

Vision Agent Demo

• Demonstration of visual AI using agentic workflows.
• Examples include counting players on a field and video analysis.
• Vision Agent enables interaction with visual data and generates metadata.

AI Trends

1. Agentic Workflows: Enhanced by faster token generation.
2. Tuning LLMs for Tool Use: Better integration into workflows.
3. Rise of Data Engineering: Especially for unstructured data.
4. Image Processing Revolution: Expected to unlock new application possibilities.

Conclusion

• Exciting time for building new AI applications.
• Generative and agentic AI expanding possibilities.
• Encouragement to explore Visual AI demos and developments.

• Website for Demos: va.landing.ai
• Invitation to explore demos and apply them in personal projects.