21/09/2026
What did LLMs actually change in AI systems? 🤔
AI assistants existed long before ChatGPT.
They could already process user queries, search knowledge bases, compare possible matches, rank answers, apply thresholds, and interact with external systems. But all of this was usually built as an explicitly designed pipeline of specialized components.
One common approach was to represent text with TF-IDF, compare queries using cosine similarity, rank the closest known questions, and return a predefined answer if the confidence score was high enough.
The main limitation was simple: the system could only work with what had already been defined.
LLMs changed that.
The key shift was from retrieving and classifying to understanding and generating. Instead of choosing the closest known answer, an LLM can understand intent, use context and knowledge, and generate a response to a previously unseen query.
At the same time, one thing didn’t disappear: architecture. Retrieval, context, external tools, APIs, dialogue management and monitoring are still part of building AI systems.
This was one of the topics Artem, our Solution Architect, explored in our AI Talks webinar: The Pre-GPT Era.
Swipe through for a quick breakdown 👉