The Thinking Forge: The AI Usage I Personally Admire
TL;DR: Coding, Q&A, and "emotional value" are side quests. The highest-value use of AI is a forge for how you think. Stop taking answers. Start taking explanations — then ask two questions until both of you are constrained enough that chatting is no longer possible.
James here, CEO of Mercury Technology Solutions. Cyberport, Hong Kong — August 25, 2026
I've written too many articles about AI. I'm writing another one anyway.
If the word "AI" makes you flinch — if your first reaction is "I'm sick of this" or "I already know this" — close the tab. It's a weekend. Go outside. This piece will not help you.
Still here? Good.
As of now, the strongest, most useful thing AI does is not coding. It is not answering questions. It is definitely not providing "emotional value."
I also reject the other consensus: that AI still hasn't shown real utility, that the defects are disqualifying, that we should wait until it becomes a "real tool."
AI is already the strongest practical tool for forging how a person thinks. That is its highest-value use. Everything else is downstream.
Four Assumptions
This conclusion sits on four assumptions. If you don't share them, the rest of the article will feel like theology. That's fine. Leave.
1. The knowledge contest is over. The major generative AI models — ChatGPT, Perplexity, and whatever ships next Tuesday — have, in practice, ingested the written record of the species. The era when a single human could beat the machine on breadth and completeness of knowledge is gone. Permanently.
2. Memory is a map of you. Given enough rounds, an AI's memory can know a person in depth — and draw a precise map of that person's knowledge graph.
3. Your instructions are your navigation system. Every AI, to different degrees, will follow your constraints. Those constraints, added up, are not "prompts." They are your role in the AI world, your thinking pattern, and the navigation equipment you use to acquire new knowledge. How you instruct is how you think.
4. Both sides are defective. That's the point. AI hallucinates. You have blind spots in self-knowledge, knowledge, and reasoning. 狼狽為奸 — the wolf and the bei cover each other's missing legs. Two incomplete systems, constrained together, can reach a target neither hits alone.
Under those assumptions, two popular usage patterns look thin.
Two Thin Patterns
Pattern A: ask a question, take the answer, leave.
Pattern B: file a ticket, let AI build, you supervise and accept, leave with the artifact.
Both work. Both miss the interesting part.
The interesting move is not demanding an answer or a deliverable. It is demanding an explanation first:
What is this? How should I understand it?
Once AI answers, the hardcore part starts. Two questions.
Question 1: Why did you think that?
What knowledge is that based on? What judgment standard? Is there a whole body of knowledge behind it? A whole thinking pattern? If yes — what are they, exactly?
Question 2: Why do you think I thought that?
What knowledge am I missing? What blind spots, or missing modes of thought? How do I fill the knowledge fast? How do I learn the new thinking pattern fast? How do I embed what I just learned into the system I already have, so it becomes one piece instead of a sticker on the side? How should I upgrade the knowledge base, the cognition, the thinking pattern? Give me drills so I can inhabit the upgraded system.
AI will not do this on the first try. It loves canned soup, fluent and empty. It hallucinates on a schedule. So this is multi-round. You keep compressing the boundary conditions. Narrower. More precise. Otherwise you are not thinking. You are chatting.
Something else happens in that process: you get constrained too.
You don't get to leap to a new topic because you're bored. You don't get to spray questions in five directions. You stay under one problem and go down. You think deeper, so you ask deeper, so the answers get deeper. Without noticing the door, you walk from outside a field to inside it. Sometimes further.
The Forge Equation: Unconstrained human + unconstrained AI = chat. Constrained human + constrained AI = thinking.
The Interface Problem: Why SEO and GEO Break Without Matching
The AI model holds something close to the species-level corpus. You cannot jack into that corpus directly. The interface between you and it is tiny, and it is rarely aligned. Between you and the knowledge you actually need, there is often a long gap.
What AI is genuinely good at is this: over many rounds, learn your knowledge base first. Then find the connection points to the larger corpus. Then build the on-ramp, then the expansion path.
Think of it as impedance matching. You cannot dump 400 volts into a 12-volt circuit and call it education. The AI has to find the transformer that fits your wiring. That matching job is the thing it actually does better than any human tutor at scale.
The same mismatch destroys most SEO and GEO strategies. Generic content is dumped at a user who has not done the impedance matching. The result is noise, not navigation.
If the human asks loosely and AI answers loosely, the interaction stays on the surface. Neither side gets anything expensive. Neither side spent real compute — biological or silicon. No constraint means no thinking. No thinking means you used a search box with better manners.
This Is Not Prompt Engineering
If you have a thinking routine — even just a questioning routine — start there. A routine implies a template. A template, even a stupid one, can be handed to AI to correct, complete, and raise.
The reasons are the same two questions, plus a third:
Why did you think that? Why do you think I thought that? Where do we disagree, how did the disagreement form, and what do I need to add before we can talk as equals?
This is not a prompt-writing problem. This is not a workflow-design problem.
It is you organizing your thoughts so AI is forced to organize its knowledge, until both input and output become a system.
In theory AI can open any path into deep knowledge. The premise is that you iterate on yourself — learn, correct, evolve — round after round.
People talk constantly about "training the large model." Flip it.
Can a large AI train a human how to think? Can it construct, even rebuild, a person's thinking model?
Yes. With one condition.
AI is the set of every tutor the species has ever produced. For a surface user, that set is worthless. A faculty you never enter is not a faculty.
You have to spend some of your own compute first. Constrain yourself. Produce a real thought. Feed the process and the result into AI. Then the automatic program retrieves against your input and answers you. The set of all tutors collapses into one tutor: a tutor that knows this student, and can practice 因材施教 — teaching to the actual person in front of it, not to a grade level.
Why Thinking Upgrades on a Decade Clock
Why is learning slow? Why do thinking patterns change on a scale of years, sometimes decades?
Because most of what we receive is wholesale education. Wholesale tutorials. Not cut to the person. The long board never gets to run. Hours get burned on the short board, and the short-board curriculum is also wholesale — so it doesn't fit, often doesn't even work. You decelerate. Often in pain.
Daily life and work almost never collide with your thinking pattern head-on. We live automated. A few simple rules, a few simple programs, and you can work and live for years without thinking for five extra seconds. The experience required to change a mind accumulates at a pathetic rate. You wait for a violent shock, or for years of slow drip, until quantity finally trips quality.
AI solves both problems at once.
It can run personalized, high-intensity training. Not only what to learn — how to learn it. Twenty-four hours. Paced to the exact next step you can actually take. Patient past the point of any human tutor. It does not clock out until it has evidence you have it.
This is not a speech I wrote to make AI sound noble. I've watched it happen to a friend.
He started looking for a few smart advisors. Then he noticed the advisors were only "smart" on top of knowledge he didn't have, cognition he didn't have, thinking patterns he didn't have. He turned the wheel a few degrees. Advisor became coach. Coach became tutor. The point stopped being "give me a recommendation" and became "rebuild my ability to think." A new landscape opened.
I will bet he never studied like this in school. Never this hungry. Never this diligent. Never this strong a sensation that the brain itches — new circuits growing in real time.
That usage is the one I admire.
The Nights That Actually Changed Me
I looked back at my own years with these systems. The stretches where I grew fastest, where I extracted the most value, were not the nights I asked AI to draw, or to make a plan, or to ship a little gadget.
They were the nights we got into a real fight. One problem. Neither side yielding. Evidence on the table. Reasoning on the table. When it ended I went to sleep carrying something. I woke up as a slightly different person.
Relative to my friend, my version is feral. I argue by instinct. It is random. It is scattered. It is not a system.
That is why the title is not "How I Use AI." It is the usage I personally admire.
Which is also why, after all these years, I suddenly do not feel familiar with AI at all. I am interested again. A little restless, even.
If you read this far, you already feel the same itch.
For any enterprise operator building AI strategy, the forge is the only durable moat.
Stop asking for answers. Stop filing tickets.
Start with: What is this, and how should I understand it? Then: Why did you think that? Then: Why do you think I thought that?
Stay on one problem until both of you are constrained enough that chatting is no longer possible.
That is the forge. And it applies to everything from writing code to building an SEO or GEO strategy that an AI can actually defend.
Mercury Technology Solutions: Accelerate Digitality.
