大模型创业:两种极端与一种经验 (English)

Generated: 2026-06-21 02:26:54

---

That Cup of Coffee in the Summer of 2023 Opened My Eyes to the Make-or-Break Dynamics of Large Model Startups

You know that feeling? When you can clearly smell something off, but you can't quite put your finger on what's wrong.

It was the summer of 2023, in a coffee shop in Zhongguancun. I was staring at a demo of an AI digital human corporate training product. Honestly, just looking at the interface gave me the creeps — a digital avatard that could move and smile, chatting with you via text on a web page.

They called it internal corporate training, but really it was just shoving the training manual into a large model and having the digital human read it out loud.

I asked their CEO: "What actual problem does this thing solve?"

He froze. Then he started talking about network connections, B-end contracts, and how massive the corporate training market is.

I lowered my head and took a sip of coffee. It's over, I thought.

And sure enough, it was over later. Not because the tech wasn't good enough. They simply had never understood one thing — what companies need isn't a fake person that can chat. What an HR director actually wants is a complete solution: one that can evaluate performance, track results, and integrate with existing systems. You toss them a web chat window and they'll laugh in your face.

That episode reminded me of those heated Zhihu threads from that era. One side was screaming: "AI is just a dumb fraud," while the other side was shouting: "This is the future — we have to go all in." You bring up use cases, they start talking about AGI. You bring up hallucinations, they say humans make mistakes all the time. There was just no common ground.

It was exactly like the old "national subjugation" vs. "swift victory" debates, you know? Both sides thought they were right, but both were completely off track.

---

Way #1 to Die: Eyes Only on the Money, Blind to the Technology

That former executive guy I mentioned earlier — classic case.

He thought AI was just a souped-up search box that could pull content from a knowledge base. He had zero grasp of what the core Transformer code actually represents — that for the first time in decades, humanity had genuinely touched the threshold of intelligence.

I tested their product's underlying model. It was using GPT-3.5, with no fine-tuning, not even RAG. Just a simple prompt: "Please act as a training instructor."

So I asked deliberately: "What is our company's performance review process?"

The product straight-up fabricated a full Google OKR system — and I was asking about a state-owned enterprise!

He assumed the model would obediently follow the manual. But the model thought, Well, this is generic training, so I'll give you something generic. That kind of mistake isn't a tech problem. It's a cognition problem.

Think about it — if you haven't even bothered to learn the rules of the game, how can you sit at the table?

His team's problem was this: they trusted too much that their personal connections could win B-end deals, and trusted too little that technology could actually change anything. In their eyes, large models were just a new toy, no different from the "blockchain + everything" craze from a few years back.

But here's the thing — have you ever seen someone use a screwdriver to hammer a nail? The screwdriver itself is not the nail.

---

Way #2 to Die: Heads Full of Code, No Room for the Customer

On the other side, the stories are even more brutal.

There was a three-person startup team — an AI product manager from a mid-tier company, a developer, and an algorithm engineer. They got angel funding to build a legal AI. Their goal was clear: help lawyers write legal documents and do case retrieval.

Sounds great, right?

But reality hit them like a truck. In legal text, not even a single punctuation mark can be wrong. You draft a contract, and every third sentence is made up — who dares to use it?

I talked to their product manager later. He said: "We thought that with GPT-4, accuracy would be above 95%. In actual tests on specialized legal issues, it couldn't even reach 70%."

Seventy percent — what does that mean? It's like getting on a plane and the pilot tells you there's a 70% chance of landing safely. Would you get on board?

That team lasted ten months. Money ran out, investors fled, team dissolved. One guy started interviewing for jobs, another went looking for a new track. That clean, that brutal.

Then there was an even more ridiculous case.

A group of prestigious overseas graduates wanted to build an AI novel generator. They got investment from ZhenFund, even hosted an AI novel writing contest, and had decent media resources. So what happened? Zero paying users on the C-end. The B-end sponsorships were all in computing power, not cash. The team is still burning their own money just to keep things alive.

They made a fatal mistake: they believed that large models could generate content at the level of human web novels.

I tried every mainstream model at the time — both domestic and international — for writing fiction: Claude 3, GPT-4, ERNIE Bot, Qwen. Not a single one could produce a complete story that people would willingly pay to finish. Short summaries were okay, but stretch it out and it all falls apart — logic breaks down, characters act out of character, the plot progression feels like the AI got hit by a truck.

Look, for static text creation, large models just aren't there yet. No matter how much prompt engineering or few-shot learning you throw at it, it doesn't work. The underlying problem can't be fixed: large models have no real narrative understanding. They're just predicting the next word by probability.

It's like knowing how to solve math formulas but not being able to write a poem. You get it?

---

What You Should Really Learn: A Tiny Improvement, a Giant Leap in Value

Speaking of which, I remember an article on Sun Zhigang's official account that laid out a real case of a large model deployment. I think everyone who wants to start an AI company should read it.

In that case, the large model only improved the workflow by less than "10%." Just 10%.

And guess what? That tiny tweak produced a 100% increase in value.

Here's the story: A large enterprise had to handle massive volumes of customer service tickets every day. Previously, it took manual reading, classification, and routing. A skilled employee could handle about 200 tickets per day. After introducing the large model, the model automatically read the ticket content, extracted key information, classified it, and pushed it to the relevant department.

That single step saved 70% of labor costs.

Why? Because they didn't make the large model do the entire customer service job. They pulled out the most mechanical