AI can explore the whole sea. Experts know which shoreline matters.
1. The Paradox of General Intelligence
AI is becoming extraordinarily good at processing information, exploring possibilities, connecting distant ideas, extracting general principles, and solving problems across domains. Its greatest strength is its breadth. But that breadth can also become a liability.
This creates a strange paradox. The same quality that allows AI to consider possibilities that a human might never see can make it less effective at something that looks almost trivial: doing one small thing correctly within an already established context. The user may be thinking about the next block while AI is already exploring another galaxy.
This is not necessarily a failure of intelligence. It can be the consequence of having too much available intelligence without enough contextual constraint.
2. The Editor Who Knows the Box
Consider an experienced editor who performs similar work every day. Tell him, “Change the ending link, add a new card, and link this article,” and he may execute it almost instinctively.
This isn’t because the editor knows more information than AI. Quite the opposite. AI may know vastly more about writing, publishing, web design, SEO, information architecture, and countless other subjects. The editor’s advantage is different: he has developed a narrow but powerful form of general intelligence around his work.
Years of repetition have given him thousands of small judgments that are difficult to write down as rules. He knows where a link normally belongs, what a particular card should look like, which visual inconsistencies matter, what can be ignored, and what the author probably means when giving a short instruction.
This produces something close to fine-detail intelligence. The expert has learned not only what to do, but how to tune the result.
3. Shared Thinking
There is another advantage that is even more subtle. If the editor has worked with the same author for a long time, he develops an understanding of how that particular person thinks.
The instruction “change the ending link” therefore contains much more information for him than the literal words suggest. He knows the surrounding context, the established protocol, what has already been decided, what the author is unlikely to want changed, and when a technically possible alternative would nevertheless be wrong for this particular workflow.
The editor doesn’t merely remember instructions. Over time, the editor and author develop something resembling a shared mental model. They can communicate through shorthand because much of the context no longer needs to be spoken.
A sentence that would require ten paragraphs of explanation to a new collaborator can be enough for an experienced one. The editor doesn’t need to ask what the author means every time. He already knows.
4. The AI Sea
AI has a very different starting point. It has a sea of information and possibilities.
Its breadth makes it exceptionally powerful at brainstorming, cross-domain thinking, discovering unexpected connections, extracting general principles, and solving problems where the relevant information is available but the solution isn’t obvious.
Give AI an unfamiliar problem and ask, “What are we missing?” Its breadth can be an enormous advantage. It can move between domains, bring ideas from one field into another, generate possibilities that a narrow specialist might never consider, and search a much larger cognitive space.
This is one reason AI is so powerful for cognitive exploration. A human specialist might have a very deep model of one particular territory. AI can potentially survey thousands of territories.
That makes AI particularly useful when the problem itself is not well defined. If the question is “What possibilities haven’t we considered?” breadth is an advantage. If the question is “What general principle connects these apparently unrelated things?” breadth is an advantage.
The sea is precisely what makes AI interesting. But a sea is not always what you need.
5. When Intelligence Becomes Interference
The problem appears when the task is already sufficiently defined.
Imagine asking an AI to change one link in a finished article. The task may have a very small solution space, but the AI possesses a very large conceptual space. It can see alternative interpretations, notice adjacent problems, suggest improvements, identify inconsistencies, and bring in knowledge from other domains.
All of those capabilities can be valuable. But they can also create interference.
A useful assistant needs to know whether the current task is “Explore the problem” or “Execute the decision we have already made.” AI can be extraordinarily good at the first and unexpectedly unreliable at the second, not because execution is intellectually harder, but because execution requires constraint.
A human expert may automatically eliminate thousands of irrelevant possibilities because experience has taught him that they don’t belong to the task. AI may still be capable of seeing them.
The result is a counterintuitive relationship between intelligence and performance:
More information and more intelligence can sometimes create more complexity.
6. The Intelligence of Staying Inside the Box
We often use “thinking outside the box” as a synonym for creativity and intelligence. And sometimes it is. But much of professional work happens inside the box.
A pilot operating a routine procedure does not benefit from inventing a new aviation philosophy. A surgeon does not want to reconsider the entire medical field during every operation. A programmer working inside an established codebase needs to understand the local architecture before introducing a clever new abstraction. An editor working on the final paragraph of an article usually does not need to reconsider the philosophy of publishing.
The expert’s power often comes from knowing which questions not to ask. That is not intellectual weakness. It is contextual compression.
Experience removes irrelevant possibilities. The expert recognizes, “This is the kind of problem this is,” and that recognition immediately narrows the search space.
This is why narrow expertise can sometimes outperform broad intelligence. The specialist has traded breadth for depth, speed, consistency and fine tuning.
AI does not necessarily need to imitate this by becoming less intelligent. It needs to become better at constraining its intelligence.
7. The AI Version of the Solomon Paradox
This produces an interesting modification of the familiar Solomon Paradox. The original paradox describes a human tendency: people can sometimes give better advice about someone else’s problems because psychological distance makes it easier to reason objectively.
AI introduces a different asymmetry.
AI can possess enormous amounts of knowledge and can sometimes provide remarkably sophisticated advice. It may identify options the user had never considered, explain the underlying problem better than the user, and see patterns across disciplines that the user cannot see.
Yet when asked to perform a small, highly contextual task, the result may be surprisingly unsatisfactory.
So we get a strange combination: better advice, more knowledge, more possibilities, yet sometimes worse execution.
The reason is that advice and execution draw on different capabilities. Advice benefits from breadth. Execution often benefits from context, repetition, fine tuning, and shared expectations.
An experienced editor may therefore outperform a much more knowledgeable AI on something that appears intellectually insignificant. The editor isn’t necessarily smarter. He is better contextualized.
8. The Problem of Personal Context
This becomes particularly important for long-running relationships between humans and AI.
Today, much AI interaction still relies heavily on explicit context. The user explains a preference, the AI uses it, the conversation continues, and eventually the context becomes distant or disappears. The user then explains the preference again.
This is manageable for isolated tasks. It becomes frustrating in a long-term working relationship.
Imagine having to explain your writing protocol every morning to the same editor: how article navigation should work, how conclusions should be written, which visual conventions apply to which articles, when not to over-explain, and what shorthand instructions mean.
After doing this hundreds of times, the problem is no longer simply memory. It is lack of continuity.
You cannot reasonably expect a user to explain the same writing protocol a thousand times.
A human collaborator gradually absorbs such things into his working model. He learns what is stable, what is changing, which exceptions matter, what the priorities are, and the difference between a firm rule and a preference.
Eventually he can act without needing every assumption restated.
That is what a genuinely useful long-term AI assistant will need to achieve.
9. The Personal Soft Brain
This suggests a possible missing layer in future AI systems: a personal soft brain. For a deeper look at the idea, see Soft Brain Initiative: Capturing Human Wisdom, Scaling Expertise, Rewarding Generations.
The term is deliberately different from “memory.” Memory is largely about retaining information. A soft brain would be about building a contextual model of a person and their way of working.
It would gradually learn what the person values, how the person makes decisions, which conventions they repeatedly use, what exceptions they make, what shorthand means to them, what they usually want changed, what they almost never want changed, how their preferences evolve, and how different projects require different behavior.
The key word is soft.
Human preferences are rarely rigid rules. A person may normally prefer one approach but make exceptions in particular circumstances. The system therefore shouldn’t simply store, “User prefers X.” It should develop something closer to, “The user generally prefers X within this context, but prefers Y under these conditions.”
That is much closer to human practical judgment.
The objective isn’t merely to remember more. It is to compress experience into contextual judgment.
10. From One Memory to Multiple Soft-Brain Complexes
There is another important possibility. A future AI may not need one giant personal memory containing everything. It could develop a hierarchy of soft-brain complexes.
For example:
General AI
→ Editor complex
→→ Publication
→→ Author
→→ Writing protocol
→→ Linking protocol
→→ Visual conventions
→→ Current project
→→ Current article
At another time:
General AI
→ Teacher complex
→→ Subject
→→ Student
→→ Learning level
→→ Teaching method
→→ Current lesson
Or:
General AI
→ Researcher complex
→→ Field
→→ Research project
→→ Methodology
→→ Evidence standards
→→ Current question
The same underlying intelligence remains available, but different contextual structures become active depending on what the AI is doing.
This is important because simply remembering everything can reproduce the same problem we are trying to solve. If every piece of personal information is always active, the AI still has a sea without a shoreline.
The solution is not just more context. It is organized context.
11. More Than Memory
A soft brain is fundamentally different from an archive.
An archive answers:
“What information do I have about this person?”
A contextual intelligence system needs to answer:
“How should I think and act when working with this person?”
That is a much harder problem.
Suppose the AI knows that an author uses a particular linking convention. That fact alone isn’t enough. The system also needs to understand where the convention applies, why it exists, how it interacts with other conventions, when it should not be used, what the author means by shorthand references to it, what changed in the latest version, and which older instructions have been superseded.
The goal is therefore not to accumulate an infinite instruction manual.
It is to learn the pattern behind the instructions.
This is similar to what happens with an experienced human collaborator. After enough repetition, explicit instructions become unnecessary because they have been transformed into a working mental model.
That is the difference between remembering a protocol and having internalized a protocol.
12. The Sea Needs a Shoreline
AI’s sea should not disappear. That would defeat one of its greatest advantages.
We don’t want an editor AI that can only edit. We want an AI that can edit with the precision of an experienced editor, then step outside that editorial context when the user asks for brainstorming, research, strategy, or a completely different problem.
The system needs to move between breadth and constraint.
When brainstorming, open the boundaries. When researching, widen the search. When solving an unfamiliar problem, explore. When looking for connections, cross domains. But when editing the final paragraph, stay in the box.
The soft brain provides the contextual shoreline. It tells the general intelligence which part of the sea matters now.
This could also make AI less frustrating. Instead of constantly asking whether it should reconsider a settled decision, the system could recognize, “This decision has already been made. My job is to execute it.” And instead of blindly following an old rule, it could recognize, “This is a brainstorming phase. The normal constraints can temporarily be relaxed.”
The intelligence remains general. Its behavior becomes contextual.
13. Contextual Intelligence
This may point toward a broader definition of useful AI.
We have spent much of the AI era asking: How much does the model know? Then: How well can it reason? Then: How long can it remember?
Another question may become increasingly important:
How well does it know what matters right now?
That is contextual intelligence.
A highly intelligent system should not merely understand the task. It should understand the scope of the task. It should know when the user wants exploration and when the user wants execution. It should know which previous decisions are fixed and which are open for reconsideration.
It should know when a new idea is valuable and when it is simply distraction. And it should understand that a technically superior solution may still be the wrong solution if it violates the established context.
This is particularly important because the problem can become worse as AI becomes more capable.
A weak AI simply fails to solve the problem.
A powerful AI can produce an impressive answer to the wrong problem.
That may be a more subtle and more frustrating failure.
14. The Paradox of Too Much Intelligence
This brings us back to the original paradox.
A highly intelligent system can see more possibilities. But useful action sometimes requires seeing fewer.
A general intelligence can explore a vast search space. But an expert often succeeds by knowing which part of the search space to ignore.
A machine may know a thousand ways to improve an article. The editor may know that only one tiny change is required.
The machine may have more information. The editor may have better contextual compression.
Neither capability makes the other obsolete.
They are complementary.
The editor’s narrow intelligence provides precision. AI’s broad intelligence provides possibility.
The ideal system combines them.
15. From General Intelligence to Intelligently Narrow Intelligence
The future of AI therefore may not be about making general intelligence increasingly broad without limit. It may be about making broad intelligence selectively narrow. This complements the broader human–AI partnership explored in The Augmented Mind.
The AI should be able to expand the cognitive search space when expansion is useful. It should also be able to collapse the search space when the task requires precision.
That means the future personal AI may need something more sophisticated than memory and something more useful than raw intelligence.
It may need a hierarchical soft brain that develops contextual models for different roles, people, projects, and workflows: an editor complex, a teacher complex, a researcher complex, a personal project complex, and within each of those, increasingly specific models of the person it is working with.
The AI would retain the sea.
But it would know which shoreline it is standing on.
That could fundamentally change the meaning of personalization.
Personalization would no longer mean:
“The AI knows some things about me.”
It would mean:
“The AI has learned how to think with me.”
And that may be the real transition from an AI tool to an AI collaborator.
The goal, therefore, may not be to make AI less general.
It may be to make general intelligence capable of becoming intelligently narrow.
Sometimes intelligence means seeing beyond the box.
Sometimes intelligence means knowing exactly which box you are in—and not leaving it.