不会炒币的小鱼

A $200-a-month ChatGPT, burning 1 billion tokens a day—you're probably wondering what I actually use it for?

My own ChatGPT membership started out at $20 per month. Then I upgraded to $100, and now I'm at $200.

Now, including other tools and agents, my entire AI workflow uses on the order of hundreds of millions of tokens per day. On busy days, it's as high as tens of billions of tokens.

For some tasks, I’ll even run the AI continuously for two or three hours.

I call different agents in the middle—each one looks up information, processes data, and validates assumptions. If problems come up, they keep trying until the end. Only then do I roll everything up into a final result.

Let me be honest: the more I use it, the more I feel that our space for using AI is still too small—we’re still thinking too narrowly.

When I evaluate whether someone can use AI well, I look at whether they can think through a complex problem clearly, and then organize the AI to carry it out step by step. Asking questions is obviously important, but the real task often only begins after the first answer.

Take our usual U.S. stock analysis as an example.

If you just ask, “Help me analyze ＄NVDA,” it’s easy to get something like a business overview, financial data, and risk disclaimers. It sounds complete—but after you read it, you might still not know which investment assumptions you should adjust.

I’d rather break the task down further: first, pull the company’s latest earnings reports, announcements, and transcripts of earnings calls word for word, and put the revenue, profits, orders, and management guidance from the past few quarters side by side;

then follow the industry chain—see where customers are planning to spend their money, whether suppliers have shown changes that confirm each other. At the same time, bring in interest rates, valuations, and the current share price to analyze what the market is already paying for in terms of expectations.

But we’re not done yet.

I’ll have another group of agents specifically look for counterevidence: which assumptions in this whole line of thinking are the most fragile?

If the company’s growth claim—can you find support for it in other data? If profits keep growing, does that automatically mean the stock has upside? If interest rates rise, or customers cut spending, can the original valuation still hold?

Finally, what I get delivered should be a research package I can keep tracking:

which companies are worth watching, where the expectation gaps are, what needs to be verified next, and what kinds of changes would force you to overturn the judgment. The sources, data dates, and reasoning process should all be left behind so I can check them.

I want AI to help me go deeper with the research, so I can spend more energy on judgment—not place an order based on a polished-looking report.

That’s also when I realized: whether it’s my understanding of AI or my understanding of companies, it has improved a lot.

The same idea applies to designing a business model.

For example, I already have content, users, and some products and services. For what to do next, I can hand these existing conditions to AI to first map out why users care about me, what problems they’ve encountered, and which needs are worth investing in long-term—then go further to figure out where revenue will come from, where costs will go, which part is too dependent on me personally, and what should be validated first.

You can also have different agents approach from the perspectives of users, operators, and competitors to surface problems. A business might look very imaginative at first, but if customer acquisition is too hard, delivery is too heavy, or revenue growth always tracks my work hours—those issues should be broken out early. By the end, at least it should land on an executable, testable little experiment.

The same goes for productivity and time planning.

I’ve also been thinking about organizing my goals, to-dos, and actual time investment more completely. Within the scope where I’m willing to provide information and grant authorization, I want AI to participate in ongoing retrospectives: which research can be merged, which repetitive tasks can be automated, and which things have been consuming a lot of time but haven’t moved the business forward much.

If a temporary task pops up, how should the schedule for the next few days be adjusted? If a goal hasn’t made progress for several weeks, is it because there isn’t enough time—or because there never was a clearly executable next step broken out?

Used this way, AI can gradually understand what you’re pushing forward, and continue working with the momentum you already have. Starting a new chat box every day and re-explaining the background from scratch actually wastes a lot of already accumulated context.

The biggest takeaway from these practices for me is this: you need the ability to organize complex tasks first, before you can get AI to take on more work for you.

What the goal is, what information is needed, what dependencies exist between the tasks, how to check the result, what conditions mean “keep trying” and what conditions mean “must stop and confirm with you”—all of that needs to be thought through.

If you only have a vague sense of something, and you give it to AI, it will likely just expand that vagueness into thousands of words.

So now I intentionally train myself to specify tasks more completely and set acceptance criteria more precisely.

Any steps that can be pushed forward independently can be delegated. Anything involving money, external announcements, or important decisions should be kept for me to confirm. If it runs long and calls a lot, it should all serve the final result. If you keep making the same mistake after three hours of running, you should either re-check the task design or switch methods directly.

I genuinely recommend that everyone use AI more deeply. Pick a real piece of work you spend a few hours on every week, try handing over the whole workflow—from collecting information to delivering the final result—and see exactly where it gets stuck. Then adjust it round after round.

Right now, I’d rather spend time setting up a category of work that keeps recurring.

Put in a little more effort today, and every time you run into it afterward, you’ll need to handle fewer pieces yourself. That kind of accumulation is really worth it—and you’ll get deeply into it too!

I hope you can also find your own “AI flow” moments for using it…