Maarten Grootendorst, creator of BERTopic and developer relations engineer at Google DeepMind, argues that foundational AI techniques remain essential even as large language models dominate the field. In a conversation with O'Reilly's Ben Lorica, Grootendorst emphasizes that embeddings and topic modeling still solve real problems that practitioners encounter when deploying AI systems at scale.
His core insight challenges the narrative that LLMs have rendered older methods obsolete. Embeddings continue to power semantic search, clustering, and retrieval systems that support LLM applications. Topic modeling helps practitioners understand what their data actually contains before feeding it to generative models. These aren't relics. They're infrastructure.
Grootendorst's work reflects a broader pattern in AI development: hype cycles obscure practical engineering. The shift from traditional NLP to transformer-based models captured attention and venture capital, but real-world applications require hybrid approaches. You still need embeddings to efficiently search document collections. You still need interpretability tools like topic models to explain what your system learned.
His perspective on agentic systems—the focus of the discussion—suggests that agents aren't fundamentally novel. Instead, they represent a particular way of orchestrating existing components: language models, retrieval systems, and decision logic. Understanding these building blocks matters more than fixating on the agent abstraction itself.
This grounded approach matters for practitioners who need to ship systems rather than chase headlines. Grootendorst's emphasis on intuition over prompts reflects his teaching style. Practitioners who understand why embeddings work at a mathematical level can debug failures, choose appropriate models, and make informed tradeoffs. Those who only know how to write prompts hit walls quickly.
The O'Reilly conversation likely explores how embeddings enable semantic understanding, why topic models reveal dataset characteristics that matter for fine-