The case against conversational interfaces

The essay argues that natural language is a data transfer mechanism, but not a fast or efficient one for most computer interaction. Mouse clicks, keyboard shortcuts, gestures, and other compressed inputs usually beat chat when the goal is speed and convenience.

It does not reject LLMs. It treats them as useful for thinking, drafting, and new workflows, and argues that the better future is an always-on OS-level layer that complements existing tools instead of replacing them.

Reading notes

#

• Conversational interfaces keep being announced as the next computing paradigm, but past waves like Siri, Alexa, chatbots, and AirPods-as-a-platform did not replace existing usage patterns.
• Natural language is framed as a data transfer mechanism, with speed and lossiness as the key trade-off.
• People receive information faster than they send it, and thoughts form much faster than spoken or written language can transmit them.
• Gestures and facial expressions work as compression techniques because they carry less information but do so faster.
• Command lines were text-based and exact, while GUIs solved discovery and made common actions easier and faster to find and use.
• Keyboard shortcuts and modern productivity tools compress routine actions into fast inputs that feel effortless after muscle memory develops.
• Touch interfaces helped mobile workflows, but typing on phones remains slower than desktop typing, and mobile has not produced an equivalent to desktop shortcuts.
• Voice input can be faster than typing, but speaking a prompt is still much slower than tapping an app or button for many tasks.
• The main problem with voice interfaces is not only output quality, but the inconvenience of the input method.
• LLMs are presented as useful and already used in the writing process, including for a first outline built through Advanced Voice Mode.
• The strongest use case shown is conversational UI as augmentation, not replacement, such as Alexa acting as an in-game voice assistant in StarCraft II.
• The proposed future is an AI meta-layer across tools, available at the OS level, so users can trigger actions without interrupting current work.
• The closing argument is against zero-sum thinking: AI should enhance existing paradigms instead of substituting them.