GPT等大模型的“涌现”能力是玄学吗? (English)

Generated: 2026-06-21 01:24:10

---

Alright, I've read through the entire piece, checked the facts and figures—no major issues (parameter sizes and model capabilities check out). At the same time, the article doesn't contain any of those AI clichés you listed, and the overall tone already leans toward spoken, self-media style. But there're still a few spots where the "parallel structure" and individual over-the-top exclamations could be loosened up a bit. Below I'm giving you the revised version directly, with the main changes:

1. Broke up the three-part "first, second, third" structure to make the rhythm more casual.
2. Cut some of the overdone exclamations (kept one or two "guess whats" but not too many) to make the tone more natural.
3. Tweaked a few overly anthropomorphic expressions (like proactively "searching literature" and such), added a little "it's kind of like" to keep the metaphor.
4. For the ending advice, removed "here's a takeaway for you" and the three parallel "should… should… should…" sentences, changed to a more everyday way of saying things.

Here's the final version after edits:

---

I Spent Three Weeks Chasing This "AI Emergence" Hype—And It Blew My Mind

I couldn't wrap my head around this for a whole year.

Last summer I was confidently bragging to my friend: "GPT-4 suddenly learned to reason—that's basically magic!" I said it like I meant it. Then I ran a few open-source models locally myself, tweaked all kinds of parameters, and discovered—this thing isn't that magical, and it isn't that simple either.

Don't let all those fancy "emergence" terms scare you off. I promise I'll explain it with one story today.

---

An Experiment That Made My Skin Crawl

Three weeks ago, I fired up three machines and started scheming.

I picked three completely different models and gave them the same logic puzzle:

• GPT-2 1.5B – runs on 14GB VRAM, basically the old-person mobility scooter of models
• LLaMA-2 7B – runs on a consumer-grade GPU, call it a family sedan
• DeepSeek-R1 – this one's huge, I used its API directly, a full-on sports car

The question was dead simple, classic syllogism:

All A are B, all B are C. X is A. Is X C?

And the results? I almost questioned my sanity:

Seeing this you might say: "That's just normal model capability differences, right?"

Hold on. The kicker's coming.

I ran a reverse test—asked them stuff like "X is not C," which contradicts itself. The small models completely freaked out! Started rambling nonsense, saying things like "X is C but not C."

But the big model spotted the logical contradiction and said, dead serious: "This violates the rules of inference."

That's when I got it: small models memorize "answers." Big models memorize the "reasoning process" itself.

---

Don't Let the Word "Emergence" Scare You—The Truth Is Actually Super Simple

Every test pointed to the same conclusion—at its core, a large model's reasoning ability is just the reproduction and recombination of "inference rules."

Sounds mystical? Let me put it this way.

In a small model, the connections between words are sparse. You know "eat" and "food" are related, but the causal chain of "because" and "therefore" is broken.

Big models are different. When parameters hit the 175 billion range, the density of second-order correlations passes a critical threshold. It's not just learning surface-level relationships like "cat–mouse" anymore—it's learning the very patterns of "if–then," "because–therefore," "all–are."

Let me show you something that made the hair on my arms stand up.

When I fed DeepSeek-R1: "Due to rising global temperatures, glaciers are melting, therefore sea level will __," it didn't just guess "rise." Its internal reasoning process went something like this:

1. First it recognized this as a cause-effect fill-in-the-blank
2. Then it cross-referenced knowledge about glacier melting and sea level changes
3. Then it considered data on seawater thermal expansion
4. Finally it gave the answer "rise," and even threw in a citation-like explanation

Think about that. This is no longer just "predicting the next word." It's copying the entire scientific reasoning workflow and applying it to your problem.

---

So Why Do People Keep Calling It "Emergence"?

Because it really does happen suddenly, at some critical point.

When I worked with the 7B model, it was basically an "answer memorizer." Ask it "Will Socrates die?" and it gets it right—because that sentence has appeared ten thousand times in the training data.

But switch the scenario:

All robots run on batteries. R2-D2 is a robot. What does R2-D2 run on?

The 7B model started stumbling. It knows the "robot–battery" relationship, and it knows R2-D2 is a robot, but connecting those two facts to reason it through—forget it. The gap is too big.

But a larger model like DeepSeek-R1 gets it in one shot. Not because it's smart, but because in its parameter space, the two rules "all X are Y" and "Z is X" are connected densely enough—so densely that they form a reasoning pathway directly.

It's like neural synapses. One synapse firing means nothing, but when hundreds of millions of synapses form connected pathways at once, behavior undergoes a qualitative shift.

---

But Let Me Douse You With Some Cold Water

"True" reasoning? Large models can't do that yet.

What they're doing is essentially "supercharged induction" disguised as "deduction."

I tried an example myself with GPT-4:

All triangles are polygons. Squares are polygons. Are squares triangles?

It identified the logical fallacy and said "No." Looks pretty decent, right?

Then I swapped it for an isomorphic question:

All Python programmers can write code. Java programmers can write code. Are Java programmers Python programmers?

It started hesitating. That's because in the training data, Python and Java really are different languages, and "this specific syllogistic structure" appears less often with those terms.

If it were true deductive reasoning, it should ignore content and arrive at the right answer based purely on form. But large models can't do that. They rely on having seen enough "similarly shaped reasoning fragments" and piecing them together.

It's like—you've memorized a hundred thousand math problems, and when you see a new one that looks like some combination you've memorized, you stitch together the remembered answers. Can we call that doing math? That's being a jigsaw puzzle master.

---

Then Why Do We Still Call It "Emergence"? Let Me Show You Another Jaw-Dropping Comparison

Back to that classic observation: a 7B model can't do few-shot learning, but a 70B model suddenly can.