Zhang Pinpin · Hands-On AI Coach
← Back to home

Don't have AI build you a horse-drawn carriage

Long-Form Video · EP0079 August 4, 2026 04:19
What this episode covers

What's this situation like? It's like we're holding the magic paintbrush of Ma Liang, which can create anything—and we ask it to build a horse-drawn carriage. A very fast horse-drawn carriage.

And then we start deciding for it: maybe this carriage needs twelve wheels, maybe it needs twenty horses to pull it.

The first minute of this episode is a rant; the real material starts at 1:10 (there’s an on-screen hint in the video saying exactly that). The rant is the first section below—skip straight past it if you only want the method.

First, let me finish the rant

“Burning 60 billion tokens a month”—that’s not me saying that. That’s an AI-community operator I keep running into online, talking about himself.

What does that even mean? 2 billion tokens a day.

Over Spring Festival this year I was burning about 2 billion tokens a day myself. What did I make in three days? A voice input method written in Rust, a web version of Claude Code at 50,000 lines, and an app for my smart glasses, Even Glasses.

Since then I’ve had one intensive development sprint after another, all at roughly the same burn rate. And the output was:

  • a commercialized comic-drama platform;
  • a commercialized e-commerce platform;
  • a commercialized digital-human platform;
  • a nutrition-management platform driven by genetic-test reports;
  • a platform for self-media and community operations;
  • business research and business analysis;
  • and tools for the Lean Canvas.

Some were one-off custom products clients bought outright. Others are continuous delivery—revenue from selling a service.

So what did these “compute kings and community lords” produce? Courses at 9.9, 99 and 199 yuan, and paid membership groups. Egging people on to learn AI they’ll never use, AI that doesn’t start from any real business. What’s the point of learning that?

Some supposed AI influencers had barely touched Codex and Claude Code for a month before they started saying on camera, “I don’t use AI any more, I use AI to manage AI.” Sounded very impressive. I dutifully watched the whole thing start to finish and contributed a completion-rate data point.

Fine. Enough of that—let’s get to the useful part.

Most people get the first step of AI research backwards

I want to use AI research as the example today, because there’s a misconception worth clearing up.

When you do research with AI, don’t you find yourself hunting for prompts everywhere, then piling all your own requirements and ideas onto the model, then fanning out a swarm of agents, burning a huge pile of tokens? And what comes back is:

  • missing certain perspectives;
  • or full of errors and hallucinations;
  • or a heap of technically-correct platitudes—completely useless for an actual business decision.

The magic brush, and a carriage with twelve wheels

What’s this situation like?

It’s like we’re holding the magic paintbrush of Ma Liang, a brush that can create anything—and what do we ask it to make? We ask it for a horse-drawn carriage. A very fast horse-drawn carriage. And then we start deciding for it: maybe this carriage needs twelve wheels, maybe it needs twenty horses to pull it.

Everything stays trapped inside the operator’s own understanding of the problem.

Go back to the bottom of it, back to the starting point, and what’s actually needed may not be a carriage at all. The world already has cars. It already has airplanes.

So the method is three moves

Whenever we put AI on a task, the groundwork is this:

  1. Help it find and connect to the best information sources for the job.
  2. Have it find what the current best practice in the world is—we can even ask for several different approaches and compare them across dimensions.
  3. Tell it the need behind the need.

I don’t want a research report. What I want is the business decision I can make once I have the report.

With the data, the best practice and the ultimate purpose in hand, a good tool and a good model can range freely instead of being capped by your own ability.

What that study actually looked like

What’s on screen is a study I ran yesterday: the AI speaking-practice tutor business in China’s mass consumer market, across different audience segments:

  • the fundamentals of the whole category, the market size, the audience map;
  • a panoramic analysis of direct competitors and adjacent ones.

My purpose was to judge whether it’s worth entering, so it lists 10 key assumptions, plus opportunity slots and choices, plus suggested ways to validate them quickly in the short term. It’s a thorough report, and it’s laid out so you can read it:

the demand is real, the category is crowded, what the big platforms do, where non-AI speaking improvement stands today, the real data of competitors in the app store, which parts of the demand are the base, a three-tier sizing of the market. I don’t walk through the rest in the video—look at how granular it gets, down to policy factors and the impact of the new rules on AI companionship.

Any boss who has hired a consulting firm for research knows what a report like this sells for.

And at the end, the report gives you the detailed sources behind every data point—so it can be verified.

30 minutes, a few million tokens

From asking to output, a report like that took about 30 minutes. It didn’t burn many tokens either—a few million, maybe a bit over ten million at most.

And research is already one of the fastest ways to burn compute: it fans out a lot of sub-agents searching and cross-checking along many dimensions, then aggregates and renders the result.

So I do wonder whether those AI influencers are producing a dozen, two dozen, dozens of reports like this every day.

Last thing

When you’re learning AI, spend the money where it matters. Use good models. Use good tools.

Don’t begrudge a hundred-something a month for an AI tool and then pay thousands, tens of thousands for a course. I’ve even seen an offline course at 59,800—what are you learning there? Someone transferring 2 billion tokens of internal energy straight into your body?

I’m speechless.

I don't want a research report. What I want is the business decision I can make once I have it.