Deconstructing AI

With GPT-4o, OpenAI has developed a multimodal AI able to perform several tasks: reading, dialogue, and vision. Such models favor a monolithic structure for AI-powered applications, where a single AI handles all tasks rather than relying on multiple specialized models. However, this approach clashes with traditional software architecture, which advocates for modular pipelines.

This talk explores the meanings of the word "deconstruct" to understand the challenges of AI. We'll examine the pros and cons of monolithic versus modular architectures in the context of AI. By shifting from specialized models to a single monolithic AI, we transfer complexity from code to the AI itself. Therefore, we will also deconstruct—or rather, take apart—an AI model to decipher its complex architecture. 

The Why:
As more developers work on AI-related features, this talk aims not only to share best practices in architecture but also to demonstrate techniques for better understanding models and reducing their biases. Interpreting AI models has become more critical than ever, as AI-powered applications can cause damage to companies or even exacerbate social inequalities. In this talk, we will explore the latest techniques to construct more robust and fair AI systems.

Key Takeaways:
Understand the trade-offs between monolithic and modular AI architectures, with real-world use cases.
Learn techniques for reverse-engineering and interpreting complex AI models. Hence, we will deconstruct together the myth of "black box" AI. 
Discover methods to identify and reduce biases in AI systems.

This session is for... you if you’re a developer, tech lead, data scientist, or a Chat GPT power user who wants to challenge the way AI is being built. While the current hype around AI has fostered both excitement and fear, this talk offers an opportunity to take a step back and question the narratives surrounding artificial general intelligence.