The First Tool That Can Speak Before It Knows Why

The Spark

Ever since generative AI appeared, we have struggled to explain why it feels fundamentally different from the software that came before it.

Our first attempts focused on probability, prediction, and hallucinations. Those explanations were technically accurate, but they never seemed to capture everything that felt different. When people nodded politely, we knew we had not yet found the right mental model.

Over months of conversations, experiments, and more than a few wrong turns, a different hypothesis began to emerge.

Generative AI may be the first widely available general-purpose tool that can produce fluent, context-rich output from a small nudge while we are still working out why the work matters.

We do not offer this as the final answer. We offer it as the best mental model we have found so far. Like all mental models, it is imperfect. Mental models are useful because they help us reason about reality, not because they are reality.

By “knows why,” we mean a human problem, not a claim about machine understanding. The system can produce plausible language before we have formed the purpose, audience, evidence, and judgment that should guide the work.

This mental model has helped us make sense of why working with generative AI often feels so different from working with almost every other tool we have used.

---

Most Tools Waited for More Intent

Think about a hammer.

The hammer contributes force. We bring everything else. We bring the nails. We bring the wood. We bring the measurements.

We bring the sequence. Most tools wait patiently until we provide enough intent for them to do useful work.

Earlier tools could create output from surprisingly little. A mail-merge template can turn a few fields into a standard letter. A procedural game can turn a seed into a map. Search can return a page of material from a few words.

Two early language systems come especially close. In 1966, ELIZA could respond to a typed sentence with a conversational turn. Its famous DOCTOR script matched key words to predetermined phrases and questions. In the late 1960s, SHRDLU could interpret ordinary-language commands and reason about them. Its entire world, though, was a small table of blocks, cones, and pyramids. The Computer History Museum’s short history of ELIZA and SHRDLU makes the contrast easy to see.

Those systems were remarkable. Each still operated inside a clear boundary: a template, a ruleset, a script, a corpus, or a tiny world. The human could be unsure of the final result, but the kind of work had already been selected.

Generative AI changes the combination. A sparse natural-language nudge can become fluent, context-sensitive output across many kinds of work. The conversation can continue while our intent develops. That is both the opportunity and the risk: expression no longer has to wait for purpose to finish forming.

I remember building Excel workbooks where entering a handful of numbers automatically produced dozens of charts, forecasts, and summary tables. The output felt almost magical the first time I saw it. But it wasn’t open-ended. The workbook already embodied the intent. I supplied the numbers; the spreadsheet followed the rules someone had already designed.

---

When Expression Outruns Intent

Consider the following conversation.

Finding Intent

• Mike Human

  Climate change.

• Zai AI

  Produces a fluent explanation about climate change...

• Mike Human

  My intent is to help my neighbor understand why our town has experienced more flooding.

• Zai AI

  The explanation shifts toward local examples and practical observations...

• Mike Human

  He's skeptical of politics but trusts engineering examples.

• Zai AI

  The conversation changes again, now emphasizing infrastructure, drainage, and rainfall rather than political framing.

Notice what happened. The first response wasn’t necessarily wrong. It simply solved a different problem.

Intent arrived. Then audience. Then context. Then purpose. The language was fluent from the beginning. Our understanding took longer.

---

A Workflow Race Condition

As software engineers, we see an analogy to a race condition. In a program, a race condition involves concurrent operations whose timing can change the outcome. We are extending that idea deliberately. This is not a claim about a concurrency bug inside a model. It is a description of a workflow whose parts no longer stay in step.

Here, expression can race ahead while intent, audience, evidence, and judgment are still catching up. Historically, the effort required to write often slowed us down. Thinking generally stayed ahead of expression. Today, expression can arrive almost immediately.

Sometimes that is wonderful. Sometimes the writing helps us discover what we think. Sometimes it hides the fact that we have not yet decided what we are trying to accomplish.

Hallucination remains important, but it does not explain the whole experience. In many everyday uses, the deeper challenge is synchronization.

How do we organize a workflow so that intent, audience, evidence, and judgment stay ahead of the language being produced?

---

Writing Isn’t the Only Thing Changing

Although we first noticed this while writing, the pattern extends much further: requirements and design; research and presentations; policies and planning; and much more.

Nearly every knowledge task can now begin with fluent language on demand.

In some kinds of writing, the bottleneck has moved. The question is increasingly whether the words are worth producing in the first place.

Generating words has become dramatically cheaper. Understanding has not.

---

Old Loops, New Loops

Education has continually adapted to new technologies. Printing, inexpensive paper, typewriters, calculators, computers, the internet, and search engines all influenced how we learn and communicate.

Generative AI is another transition.

The familiar loops of drafting, revising, questioning, and reflecting haven’t disappeared. Many have simply become faster.

Drafting, editing, and trying new ideas are faster. Other loops now deserve greater attention.

• Intent formation.
• Understanding.
• Verification.
• Workflow selection.

Each calls for a practice. We need time to form intent, ways to test understanding against evidence, and the judgment to choose a workflow that fits the task.

As the work changes, our loops continue evolving.

---

Perhaps the First Draft Isn’t Prose Anymore

For generations, the first draft was a document. Humans still need opportunities to understand what we are working with. Perhaps a first draft will become a living artifact such as:

• Conversations.
• A collection of notes.
• A concept map.
• A canvas.
• A graph of related ideas.
• Highlighted sources.
• Questions.
• Counterarguments.
• Voice notes.
• Sketches.

For years, we were surprised by how much professional engineering involved what looked like “arts and crafts.” Whiteboards, sticky notes, sketches, moving boxes around on a page. Eventually, we realized those activities were not separate from the work. They created space for the work. They slowed us down enough to notice patterns, challenge assumptions, and make better decisions.

More than once, a complicated ETL problem that had resisted hours of discussion became obvious after someone drew the data flow across a whiteboard. Nothing new had been discovered. We simply saw the relationships instead of reading about them. The drawing didn’t replace the thinking. It created the conditions for it.

While generative AI removes much of the friction involved in producing language, it also invites us to create new thinking spaces where understanding can emerge before decisions are made.

---

We Are Still Learning

We keep encountering advice about using AI better.

• Use larger prompts.
• Give the model more context.
• Think bigger.
• Ask it to solve larger problems at once.

Much of that advice is useful. It remains incomplete.

We are participating in a remarkable period of collective learning. Across disciplines, researchers, practitioners, hobbyists, teachers, students, and curious experimenters are independently discovering similar patterns.

Humanity has entered one of its periodic bursts of innovation, where millions of people are exploring a new tool at the same time, comparing notes, sharing successes, and learning from mistakes.

That’s exciting. It’s also an invitation.

---

An Invitation

We do not think we have discovered the workflow for writing with generative AI. We suspect there will not be just one.

Communication serves a purpose. Different tasks require different loops. Different audiences require different approaches.

We think we have found a better question: If language has become nearly frictionless, how do we redesign our thinking loops so that understanding grows along with our ability to generate words?