As someone who has spent more than 25 years as an engineer, architect, and CTO, I've always felt that one of the most important parts of the job is being able to explain complex things in plain language.
Not because people aren't capable of understanding them. Quite the opposite.
Most people are far more capable of understanding technical concepts than we often give them credit for. The challenge is usually not intelligence; it's translation. Every profession develops its own language, acronyms, and assumptions. Technology is particularly good at this. We can turn a straightforward idea into a three-letter acronym faster than almost any industry on earth.
The goal isn't to simplify things until they're no longer accurate. The goal is to explain them in a way that allows someone else to build a useful mental model.
With that in mind, I'd like to offer a mental model that I believe makes many aspects of modern AI easier to understand.
Start With Prediction, Not Intelligence
Many discussions about AI become unnecessarily complicated because they start with the latest tools, frameworks, or terminology.
Today it's agents.
Last year it was prompt engineering.
This year it's MCP servers, reasoning models, tool calling, orchestration frameworks, and whatever acronym the industry invents next Tuesday.
Those topics are interesting, but they often distract from a more fundamental question:
What is an AI system actually doing?
A useful place to start is this:
Large language models are predictive systems.
They learn patterns from enormous amounts of information and use those patterns to predict what should come next.
Sometimes the next thing is a word.
Sometimes it's a sentence.
Sometimes it's a tool call, a web search, a recommendation, or a clarifying question.
The underlying mechanism is the same.
Viewed through this lens, many things that initially seem mysterious become easier to understand.
Tool selection becomes a prediction problem.
Reasoning becomes a prediction problem.
Even hallucinations become a prediction problem.
Whether an AI is writing an email, summarizing a document, searching for information, or choosing between multiple tools, it is fundamentally doing the same thing: predicting the most appropriate next step based on the information available to it.
Why AI Can Use Tools
One question I hear frequently is:
"How does AI know when to search the web?"
Or:
"How does it know which tool to use?"
The answer is surprisingly straightforward.
AI systems are given descriptions of available tools and capabilities.
When you ask a question, the model evaluates the context and predicts what action is most appropriate.
Suppose the system has access to a weather tool.
You ask:
"What's the weather in Oakland tomorrow?"
The model has learned strong associations between concepts like weather, forecast, tomorrow, and location.
As a result, using the weather tool becomes the most likely next action.
The AI isn't consulting a decision tree or following a hard-coded rule.
It's using patterns it has learned to predict what should happen next.
Why Tool Selection Sometimes Fails
This mental model also explains why tool selection can occasionally be inconsistent.
Imagine an AI has access to four different systems:
- Google Drive
- SharePoint
- Notion
- An internal document repository
All four contain proposals.
You ask:
"Find the Chevron proposal."
From a human perspective, there may be an obvious follow-up question:
"Which repository should I search?"
But unless the system has been specifically instructed otherwise, the AI may simply choose the option that appears most likely based on the information available.
The challenge isn't that the AI is broken.
The challenge is that the environment itself is ambiguous.
This is one reason experienced AI practitioners spend so much time designing tools, workflows, and context carefully.
Better inputs often produce better decisions than larger models.
Why Hallucinations Happen
The same mental model helps explain one of the most discussed limitations of AI: hallucinations.
Many people think of hallucinations as a separate problem from intelligence.
In reality, they often emerge from the same mechanism.
Imagine asking an AI:
"Who was the Vice President of Acme Corporation in 2017?"
If the answer isn't available, the system still attempts to predict the most plausible continuation.
Sometimes that prediction is correct.
Sometimes it isn't.
The model isn't trying to deceive anyone.
It's not inventing information maliciously.
It's doing what it was designed to do: generate the most likely next response based on the information and patterns available to it.
The problem is that plausibility and truth are not the same thing.
Understanding that distinction explains a remarkable amount of AI behavior.
The Strength and Weakness Come From the Same Place
Perhaps the most important insight is that AI's greatest strengths and many of its weaknesses originate from the same source.
The reason these systems can:
- Explain complex concepts
- Draw useful analogies
- Summarize large amounts of information
- Generate creative ideas
- Connect ideas across domains
is because they operate on patterns and probabilities.
And the reason they can occasionally:
- Misinterpret a question
- Select the wrong tool
- Present incorrect information confidently
- Hallucinate facts
is because they operate on patterns and probabilities.
The intelligence and the mistakes are not separate systems.
They are often different outcomes produced by the same underlying mechanism.
Why This Matters
Most organizations are currently focused on what AI can do.
Can it write content?
Can it answer customer questions?
Can it automate workflows?
Can it analyze data?
Those are important questions.
But increasingly, I think the more valuable question is:
How does it arrive at those answers?
Because once you understand the mechanism, many things become easier to evaluate.
You begin to understand why context matters.
Why tool design matters.
Why human oversight still matters.
Why some AI implementations produce extraordinary results while others struggle.
And perhaps most importantly, you become less susceptible to both the hype and the fear.
The future may not belong to the organizations with the most AI.
It may belong to the organizations that develop the clearest understanding of how these systems actually work, and where human judgment still matters most.
Technology has always rewarded those who understand the underlying principles rather than just the latest terminology.
I suspect AI will be no different.
Ron Davis
Founder
Three decades building enterprise platforms. Started Joust to close the gap between strategy decks and the work they're supposed to change.