I recently conducted an informal experiment while driving between meetings. My goal was to discuss the concept of “vibe coding”—using generative AI to build software based on high-level intent rather than manual coding—from the perspective of a non-developer.
Because I needed time to think and pause while driving, I gave both ChatGPT and Gemini the same instruction: “Do not respond until I say ‘please begin.'”
The Results
The experience was night and day:
-
ChatGPT: Despite acknowledging the instruction and even repeating the “cue” back to me, it interrupted me every time I paused to think. It frequently apologized for interrupting, only to do it again seconds later.
-
Gemini: This model performed much better. It waited significantly longer during my silences. While the delay felt slightly extended at times, it was far preferable to being constantly cut off.
The Technology Behind the “Vibe”
The reason for this discrepancy lies in Voice Activity Detection (VAD).
Most voice AI models use VAD to determine when a user has finished speaking. ChatGPT’s current implementation relies heavily on silence duration to trigger a response. Even when explicitly told to wait for a verbal cue, the underlying VAD protocol often overrides the conversational instructions once a certain threshold of silence is met.
In contrast, Gemini’s architecture in Live mode appears to have a more flexible VAD buffer or a stronger integration between the audio processor and the large language model (LLM), allowing it to “hold its breath” longer when instructed.
Final Thoughts
For “vibe coding” to truly work, the “vibe” of the conversation needs to feel natural. If the AI cannot handle a thoughtful pause, it disrupts the flow of creative engineering. In this round, Gemini proved to be the more patient collaborator.
