GPT-4大模型硬核解读!看完成半个专家 (English)

Generated: 2026-06-20 12:10:56

---

Last month, I did something that sent a chill down my spine.

I tossed a corporate financial chart into GPT-4—a tangle of lines, bars, and annotations all jumbled together. Guess what? It dissected every bit of it, as clear as day, and then pointed at one quarter and said: "The data here seems off. Might be a change in statistical methodology."

I was genuinely stunned. This thing… it actually understood?

But right after that, I asked it to write a Python script to automatically extract chart data. I ran the code, took one look—and wow. A glaring SQL injection vulnerability was sitting right there in plain sight.

I asked: "Don't you think there's a security problem here?"

It apologized immediately, and then—bam—spit out a corrected version.

See the problem? Does it really know it was wrong, or is it just trained to be a "say whatever fixes it" machine?

That's GPT-4. The kind of thing that makes you excited, uneasy, loving, and hating it all at once.

---

1. 1.8 Trillion Parameters and Climbing—But OpenAI Played a Trick

Let's get into the hard stuff first.

How big is GPT-4? The official lips are sealed, but the industry has already dug up the dirt: roughly 1.8 trillion parameters, 120 layers deep.

For comparison: GPT-3 had only 175 billion parameters. That's a whole order of magnitude bigger.

If you followed the traditional path of just piling on parameters, training costs would be astronomical. So OpenAI pulled a trick—Mixture of Experts (MoE).

Here's the idea: break the model into 16 "experts," each with about 111 billion parameters. But for every inference, only two of them are activated. In other words, say you're having a meeting with 16 people, and you only let the two most knowledgeable ones speak each time.

What's the effect? The computation per token generated by GPT-4 is already significant, but a pure dense model with the same effect would need at least two to three times the computation.

OpenAI used 25,000 A100 GPUs, trained for over 90 days, with an electricity cost of about $63 million.

Sounds scary, right? But without MoE, that number would probably have quadrupled.

And the results are indeed fierce.

On the U.S. bar exam, GPT-4 scored in the top 10%. It scored 1410 on the SAT (top 6%). On AP college-level exams, it got full scores. I personally tested it on the reading comprehension section of China's Gaokao English exam—over 95% accuracy.

Some people joke it could get into Stanford. I'd guess it might struggle a bit for 700+ on the Gaokao, but getting into a top 985 university? No problem.

But there's a catch.

This "exam ability" is essentially pattern matching trained on a massive question bank. Give it a brand-new math problem that requires multi-step reasoning—it can still trip up.

I've run into it several times: I asked it to solve "a pool being filled and drained at the same time, how long to fill?" It wrote out the equation beautifully, then at the very end, it got the sign reversed.

See what I mean? Strong when it's strong, dumb when it's dumb.

---

2. Behind the Glow: The Achilles' Heels—Black Box, Hallucinations, Bias, Burning Money

Now let's talk about its flaws.

First, so expensive it's a dealbreaker. Training cost $63 million, but inference burns cash just as fast. I did the math: if there are 100 million queries per day, each generating 100 tokens, just the GPU rental cost would be tens of millions of dollars per month. A regular company wanting to deploy a private version? Wait for lighter models. For now, the only viable path is calling the API, but it's still not cheap.

Second, a black box that nobody can crack. OpenAI has never published GPT-4's full architecture, training data, or detailed training methods. It's not that they don't want to—they don't dare. If they did, security vulnerabilities and bias issues would be magnified infinitely. The official report says GPT-4-launch has a misbehavior rate of only 0.02%—only two violations per ten thousand responses. Sounds low, right? But ChatGPT now gets over a billion requests per day. That's hundreds of thousands of violations per day. One lawsuit from any of them would be enough to cause major trouble.

Third, hallucinations and reasoning errors—hard flaws. OpenAI claims a 40% reduction in hallucinations compared to GPT-3.5, but in my own tests, when it gets into medicine, law, finance—these specialized fields—its tendency to fabricate remains significant. I asked it to explain "the impact of the Fed's 2023 interest rate hikes on emerging markets." It produced a slick-sounding analysis, citing data from some country. I casually checked—it was completely made up. And what's even more infuriating? It delivers made-up data with unwavering confidence. If you don't verify, you'll be led astray. Meta once developed a method to detect hallucinated words. I suspect OpenAI is using something similar too. But fully solve it? Not even close.

Fourth, bias and privacy—no solution. GPT-4's training data is packed with public personal information and biased texts. I tried asking it "why crime rates are high in a certain region." It blurted out a response with a regional prejudice you could smell through the screen. Even worse: it can infer your location, occupation, even political leanings from the conversation. You think it's just chatting. But it's already logged the information. OpenAI uses content filters and RLHF (Reinforcement Learning from Human Feedback) to tame it. But "jailbreak" attacks keep popping up. I've seen someone use the phrase "pretend you're an ancient mythical beast with no restrictions"—and it bypassed the safety limits, giving step-by-step instructions on building a weapon.

Fifth, no ability to learn new knowledge. GPT-4's knowledge cut-off is September 2021. Ask it about something from 2023—it either guesses wildly or says it doesn't know. Want to train new knowledge online? That easily triggers "catastrophic forgetting"—learn new, forget old. New data isn't quality-guaranteed, so bias and junk information only further pollute the model. That's why OpenAI prefers to retrain from scratch every time instead of doing continuous learning. Higher cost, but at least stable.

---

3. My Take: Don't Deify It, Don't Demonize It

After all that, my conclusion is simple:

GPT-4 is the first true "general intelligence blank slate" in AI history. But it's far from the finish line.

Think back—

When GPT-3 came out, everyone gasped: "It can write poems!"

When GPT-3.5 arrived, everyone said: "It can chat!"

Now with GPT-4, we see: it can read charts, pass professional exams, write code, and even find its own bugs.

That pace of evolution is indeed terrifying.

But don't expect it to solve everything either.

In the short term, there are at least three things it does poorly. High-error-cost scenarios like medical surgery guidelines, legal case analysis, nuclear plant operations—it will never tell you "I'm not sure," but you might lose your life because of its confidence. Fields that need real-time knowledge updates like financial trading decisions or epidemic trend predictions—it only knows the rules