From prompted intelligence to engineered intelligence
Jev from @TypeSafeAI finally makes the shift easier to see: intelligence is becoming part of how software works. Not as a chatbot or bolted on as some sidecar. But intelligence precision engineered into software systems from the foundation up.
Much of generative AI has worked through what I call prompted intelligence: we give a model instructions in plain language and get text back. Then we—or our software—have to interpret that text and decide what happens next. Or what I like to call "toss spaghetti against the wall intelligence", even tho its been enormously useful.
Everyone has had a lot to debate about AI, but the mere existence of LLMs requires a new philosophy of software engineering. That philosophy enables a new evolution of the craft itself, which enables the next phase of AI: engineered intelligence.
A program can encounter some state, ask a bounded question about what that state means, receive a structured judgment, and use that judgment as part of its execution.
These are semantic questions: questions about meaning. This is semantic engineering.
This means interpretation of meaning can become part of the machinery of software.
TypeSafe founder Diogo Almeida has called this smart software. Its perfect.
Smart software is software that uses intelligence as part of how it operates.
That is different from software that was written with AI. It is different from an application with a chatbot attached to it. And it is different from using a model as an assistant that waits for a person to ask a question.
The distinction is where intelligence sits in the architecture. In the familiar prompt-and-response model, a human gathers context, asks for a judgment, interprets the response, decides whether to trust it, and determines what happens next. Smart software begins when some of that interpretation moves inside the system itself.
The application encounters a situation, requests a bounded judgment, evaluates the result within its rules, takes an authorized action, observes what happened, and continues.
That produces a different loop:
State → judgment → bounded decision → authorized action → observation → new state
Once intelligence enters that loop, the engineering problem changes. It is no longer enough to ask whether a prompt produces a good answer. We have to ask what context the judgment receives. What outcomes it is allowed to produce. How uncertainty is represented. When the system may act. When a person must review the decision. What evidence is retained. What happens when the judgment is wrong.
That is the transition from prompted intelligence to engineered intelligence. Prompted intelligence asks a model to help us think. Engineered intelligence makes bounded judgment a designed component of a software system.
And smart software is what becomes possible when we do that well.