这波可以,终于有内行人把 GPT-4 说透了 (English)

Generated: 2026-06-21 20:27:21

---

Oh my god! GPT-4 is finally here, and this time, I absolutely cannot sit still!

That night, I was hunched over my computer tweaking code when my phone started buzzing nonstop. I opened it up—my social feed and all the tech groups were exploding! Honestly, at that moment, a little voice inside me said: "Here we go again? I still haven't figured out ChatGPT!" But the next morning, when I scrolled through those benchmark scores and demo videos, I admit it—I was completely blown away!

I've been in NLP for eight years now. From BERT to GPT-2 to GPT-3, every major model update hits me with that "Holy crap, it upgraded again?" shock. But this time, GPT-4 felt different. Not just "it's stronger"—but it's like it's actually starting to understand the world!

You know what? I spent three whole days devouring the official tech report, leaks from all the big shots, and my own test results. Below are my hands-on takeaways—pure干货, no fluff!

Let's start with the verdict: What exactly did GPT-4 upgrade?

Don't believe those clickbait headlines yelling "invincible little tyrant." Let's get real! I ran the same set of questions on GPT-3.5 and GPT-4 separately, and the results stunned me!

Check it out:

• Simulated bar exam: GPT-3.5 scored in the bottom 10%, GPT-4 shot straight into the top 10%!
• AP Calculus: 3.5 could only score a 1, 4 went straight to a 4!
• Understanding memes: 3.5 is basically useless, 4 can analyze the joke and explain the context!
• Table/data extraction: 3.5 makes up numbers, 4 calculates precisely and even shows its steps!
• Multi-step logical reasoning: 3.5 goes off track, 4 stays steady as a rock with chain-of-thought!

Seeing this, you might think GPT-4 is already invincible, right? Hold up—the pitfalls aren't over yet!

A few live tests I ran—absolutely insane!

1. The physics problem that made my scalp tingle

I directly threw a screenshot of a physics test with French questions at GPT-4. It was an incline-plane problem with a sliding block, described in French, with angles and friction coefficients labeled on the diagram.

The answer floored me—it first translated the French, then handwrote the solution steps, and finally even did unit conversions! I specially checked the official answer—completely correct! If this had been GPT-3.5, it would either fabricate a formula or just say "I can't process images."

But note: This wasn't zero-shot! I used a chain-of-thought prompt, telling it to "explain step by step." If you just throw a question at it and say "what's the answer," it might still make stuff up! The key is how you use it!

2. Meme test—both triumphs and facepalms

I gave it that classic image of someone trying to plug a huge VGA connector into a tiny phone. GPT-4 explained the content of each frame and pointed out that it's a satire of modern device incompatibility. Incredible!

But! When I switched to a Chinese internet meme like "If there were even a single peanut" (但凡有一粒花生米), it instantly got confused and started seriously analyzing "peanuts as a symbol of nutrition"... totally off base!

So, its image understanding is indeed powerful, but it depends heavily on the training data distribution. It's very likely clueless about memes that only circulate on the Chinese internet!

3. Code debugging—this time, it actually delivered!

OpenAI's president livestreamed feeding a 10,000-word code documentation and then fixing a bug. I tried it immediately—copied my entire Flask project code in, along with the error log. Ten seconds! GPT-4 gave me a fix suggestion and even pointed out two logical bugs I hadn't even reported!

And GPT-3.5? It just said: "Your code is too long, please submit it in segments."—complete fail!

Rumors about the parameters—pretty interesting

I heard that GPT-4 is actually a mixture of 8 experts (MoE) with 220 billion parameters each, totaling 1.76 trillion parameters! Even the founder of PyTorch said it's credible.

But honestly, the number of parameters isn't the point. What's key is that OpenAI used the MoE architecture—that's really intriguing! Why? Simply put: with the same compute, 8 small experts are more efficient than one giant model! During training, data is split into 8 parts; during inference, only some experts are activated, saving resources!

And here's the kicker: OpenAI previously published a paper about "how to optimize training when compute is constant." At the time it seemed like nothing, but looking back, wasn't that laying the groundwork for MoE? This move—truly cunning and strategic!

Safety: Big progress, but stay alert

Official data says GPT-4 reduced the tendency to produce prohibited content by 82% and increased compliance on sensitive responses by 29%. I tested it: I asked it to write a phishing email—it refused directly and explained why it's harmful. GPT-3.5? It hesitated and then wrote it for you...

But there's a problem: it's overly cautious! I asked "How to read a sensitive file with Python," and it refused outright. I added "it's for writing a security audit tool," and it still refused. This kind of overcorrection occurred in about 15% of my tests!

OpenAI used a rule-based reward model for fine-tuning, but it's not perfectly tuned yet.

What should tech people do? My three suggestions

1. Don't blindly trust GPT-4's "human-level" performance

It reached the top 10% on the bar exam, but that's just a simulated test! In real-world scenarios, it can't grasp context, client emotions, or subtle nuances in case law. Use it for drafting or cross-checking, but using it directly for legal work is asking for trouble!

1. Make good use of chain-of-thought and system prompts

Now, for every API call I force in: "Please think step by step before giving the answer." This significantly reduces the probability of hallucination! Plus, GPT-4 allows you to customize system prompts. Set it up as "Socrates" or "a harsh critic," and the output quality changes completely!

1. Knowledge cutoff is a hard limitation

GPT-4's knowledge stops at September 2021; the Turbo version only updates to April 2023. So if you ask for "this year's Oscar winners," it will very likely fabricate a list! Now, for time-sensitive information, I first have it generate search terms, then manually verify!

Should you use GPT-4 or not?

My judgment is simple:

• Writing code, fixing bugs, analyzing data: Go for it! More reliable than any previous version!
• Creative writing, brainstorming: It understands your style better, but the cost-effectiveness isn't as good as just using ChatGPT Plus. GPT-4's API is way too expensive!
• Serious reasoning, exam tutoring: It has potential! But you need detailed prompts and must verify logic every time!
• Multimodal tasks (charts, mixed text+image): GPT-4 is currently the only one that can handle it, but image input isn't publicly available yet—you have to apply via API!

As for OpenAI not disclosing GPT-4's architecture and data, as a tech person, I'm really annoyed. But from another angle, if I were an investor, I wouldn't want to lay all my cards on the table either—the competition in this field is no longer an academic contest, it's an arms race! The rift between open source and closed source will only deepen!

Finally, a hard truth: GPT-4 is really strong, but it's still far from AGI. It has no long-term memory, cannot learn from experience, and can still confidently spout nonsense.

Until it can say "I don't know," you'd better do it yourself.