Conversational AI is older than large language models. Phone menus that understood “check my balance,” early voice assistants and rule-based chatbots all count. Many were task-oriented dialogue systems that tracked a “frame” of details to fill, such as destination, date and time for a flight booking, and asked follow-up questions until every slot was filled (Jurafsky and Martin). Voice versions add speech recognition on the way in and speech synthesis on the way out. Large language models changed the field by making open-ended conversation work without hand-built rules for each intent; the underlying discipline is natural language processing.
The hard part is the conversation, not the single reply. A system has to remember what was said, handle corrections and interruptions, and know when to ask rather than guess. Research finds that leading models still lose track more often when a request unfolds over several turns (Laban et al.); see multi-turn conversation. In business, conversation is increasingly an interface to action. In episode 16 at 11:49, DevRev’s Manoj Agarwal describes a system that can “converse with the customer, with the conversational AI interface” and then update the order system, which needs reliable two-way data sync. Galileo co-founder and COO Yash Sheth co-hosts episode 16 (0:24), recorded while Galileo sponsored the show.
Builders still debate how to make these systems dependable. In episode 54 at 35:54, Scaled Cognition’s Dan Klein says his team’s hardest problem was figuring out “how to do for conversational AI what RL has done for things like game playing or math or code,” by generating the right training data; that is his company’s approach, not a field consensus. Episode 54 thanks Galileo as its presenting sponsor in the opening read. Related: AI assistant, conversational analytics.
Hear it from the guest
“Converse with the customer, with the conversational AI interface”
“Figure out how to do for conversational AI what RL has done for things like game playing or math or code”
Quotes lightly edited to remove filler words.
Sources
- Jurafsky and Martin: Frame-Based Dialogue Systems — Textbook chapter on task-oriented dialogue systems built around frames, slots, dialogue state and dialogue policy; checked October 24, 2026.
- Laban et al.: LLMs Get Lost In Multi-Turn Conversation — Finds leading LLMs perform significantly worse when a task unfolds over several conversational turns than when given at once; checked October 24, 2026.
- Chain of Thought episode 16 at 11:49 — DevRev's Manoj Agarwal describes a conversational AI interface that talks with a customer and updates back-end systems.
- Chain of Thought episode 54 at 35:54 — Scaled Cognition's Dan Klein describes generating training data to do for conversational AI what reinforcement learning did for games, math and code.
