This is an original English companion essay to 為什麼 Chatbot 不會是 AI-native computing 的終點？ (Traditional Chinese). It explores related ideas; it is not a full or literal translation.
A conversation is a wonderful way to tell a computer what you want.
You can start with an imperfect thought: “Compare these plans,” “Help me understand this mistake,” or “Show me what matters.” You do not need to know the name of a menu or the structure of a database before asking.
That is a remarkable improvement in how people interact with software.
But it leads to an easy mistake: assuming that because conversation is a useful entrance to computing, it should become the whole computer.
Imagine managing a product through one long chat history. Somewhere inside are your roadmap, decisions, unanswered questions, experiments, files, customer evidence, and the reasons you changed your mind. The assistant can summarize them. Yet whenever you want to compare two versions or verify a decision, you have to ask the system to reconstruct the picture.
The more fluent the conversation becomes, the easier it is to overlook what has disappeared from view.
Conversation is not the same as state
A conversation unfolds in sequence. Most serious work does not.
A product team needs to inspect several hypotheses at once, compare evidence side by side, and see which decisions are final and which remain open. A learner may need to rearrange ideas visually. A designer may understand something by moving objects directly, not by describing every move in words.
When all of that is compressed into a transcript, we risk losing the ability to see, compare, manipulate, and preserve the structure of our work.
The answer is not to abandon chat. It is to give chat a more useful role.
Three kinds of interaction belong together
I think a humane AI-native environment needs at least three modes.
Direct manipulation: A person can point, move, edit, compare, and arrange things themselves. The interface makes the important state visible.
AI transformation: A model can reorganize existing material into a table, timeline, map, or summary. The representation changes without silently changing the underlying facts.
Agentic execution: A system can carry out delegated tasks across tools, with explicit permissions and verifiable results.
These modes complement one another. None is a universal replacement for the others.
For example, I might ask AI to propose a product plan, compare its options visually, edit the priorities myself, and authorize an agent to create only the approved tasks. I should be able to distinguish what was suggested, what I changed, and what the system actually did.
That distinction becomes more important as tools become more powerful.
Better intelligence needs clearer authority
Drafting an email is not sending it. Suggesting a database change is not applying it. Describing a purchase is not charging a card.
The transition from advice to action changes who bears the risk.
An AI-native system should make that boundary legible. What can the model read? What can it propose? What can it change? What requires confirmation? What can be reversed?
Power without understandable control is not necessarily progress.
And an interface that rearranges itself beautifully on every visit is not necessarily a better interface. People also need familiarity, stable references, history, and the ability to take their work elsewhere.
A question worth testing
I do not know whether the future will be organized around apps, files, semantic objects, or something else. Existing applications provide important benefits: coherent permissions, predictable interfaces, and shared conventions.
But I think we can test a smaller design hypothesis:
Can AI make software more flexible without making the person’s world harder to understand or control?
A good experiment would not ask only whether an assistant completes a task. It would also ask whether the user can inspect the result, correct it, understand what changed, and continue without losing context.
That, to me, is the more interesting promise of AI-native computing.
Not a computer that talks more.
A computer that helps us think and act with greater clarity.
For the full original discussion in Traditional Chinese, read Why Chatbots Are Not the End of AI-native Computing. The article above is a separately written English companion.
