涌现是伪科学吗? (English)

Generated: 2026-06-22 11:32:04

---

Alright, today let's talk about a seriously thrilling topic—emergence.

You know, if you scroll through your phone, you've definitely seen the word: "AI emerged reasoning capabilities," "the emergence of consciousness," "emergence in complex systems." It's made to sound so mystical that you feel like you're falling behind the times if you don't understand "emergence."

But let me ask you this—

90% of what you've heard about "emergence" is basically bullshit.

I'm not saying this lightly. Last month, I did something ridiculously obsessive—I dug through every single post about emergence I could find on Zhihu, from physicist Wen Xiaogang's "演生" (emergent evolution) to the enactivism crew's "decomposition," from Barabási's power-law distribution to Rosas's "lossy compression" theory. I devoured it all.

Then I ran an experiment. I threw the same question—what is emergence—at GPT-4, Claude 3.5, and some domestic Chinese LLM. Guess what? The answers I got from the three models were galaxies apart.

See, this itself is pretty "emergent"—a pile of low-quality content clustering together, suddenly producing some holistic impression that leads you straight into a ditch.

Alright, today I'm going to fill in this massive pit for you.

---

Section 1: Is Emergence Actually Pseudoscience?

Straight answer for you: No. But it depends on who's using the word.

In 1972, physicist Philip Anderson wrote a killer paper called More Is Different. Right from the start, he slammed the table: The universe is not something you can simply break down and understand! When you put a bunch of particles together to form a macroscopic object, brand-new properties emerge—like the "rigidity" of a solid, which you can't explain using individual atoms. That's not mysticism; it's hard scientific fact.

But then look what happened later. A whole bunch of people turned this word into a universal get-out-of-jail-free card.

You ask them: "How can an ant colony build such complex nests?"

They say: "Because emergence."

You ask: "How does consciousness arise?"

They say: "Because emergence."

You ask them again: "Why can large language models suddenly solve math problems?"

They say: "Because emergence."

It's like asking a programmer, "Why did the code crash?" and they reply, "Because of a bug." Not wrong, but don't you just want to punch them?

Speaking of which, I have to confess a pitfall I fell into. In 2023, when I first started tinkering with LLMs, I saw all the hype online about "emergent reasoning abilities in LLMs." I bought it. I was so excited I wrote an article praising it. Later? I spent two whole months running over a dozen fine-tuning experiments, pushing models from 70M parameters all the way up to 7B, working like a dog. Only then did I realize—

So-called "emergence" depends a lot on how you test it.

If you test with multiple-choice questions, you get a smooth linear improvement, calm and quiet. But if you switch to open-ended Q&A, suddenly there's a turning point, and the data jumps up. That turning point gets packaged as "emergence."

At this point, I just want to say—all those people who cursed me out in the comments back then for not understanding complexity science, are you guys doing okay now?

The key thing is that later Rosas's paper What is emergence, after all? provided a mathematical foundation: emergence is essentially "lossy data compression." You discard a massive amount of microscopic information, keep only a few macroscopic variables, and can still accurately predict the system's behavior—then the details you threw away have "emerged" as macroscopic laws.

Think about it: temperature, pressure, volume—these basic concepts aren't they compressed out of the positions and velocities of billions upon billions of molecules?

So here's the crux: Emergence itself is fine, but have you clearly stated the "compression algorithm"? If you can't, you're just being a jerk.

---

Section 2: Why Does the Word "Emergence" Sound More and More Like Mysticism?

I came across an old book by Samuel Alexander, who directly said: emergence must be accepted with "natural piety."

In other words—don't ask, just call it a miracle.

This attitude ruined the reputation of "emergence" back in the early 20th century. Today, some people have inherited this fine tradition, just swapping out "miracle" for "complexity."

What's the most classic trap? Let me tell you: The enactivism crowd's critique of emergence is truly on point. They say the phrase "neurons emerge consciousness" is just repackaging the problem, not explaining it. But guess what their solution is? They propose that "mutual confirmation between different sensory modalities" is the essence of consciousness—sounds even more like mysticism, doesn't it?

I spent three whole days grinding through their terminology, and finally realized in despair—they just swapped "emergence" for "enaction," and none of the problems were solved.

Anyone who's worked in industry knows that bosses love using words like this. I was chatting with a startup team recently that does AI agents. Their CBO—I secretly call him the Chief Hype Officer—kept saying: "Our multi-agent system will exhibit swarm intelligence through emergence."

I said, okay, so show me the individual agent's decision-making mechanism and the interaction protocol?

He flipped through his PPT for ages and couldn't find a thing.

This is textbook use of "emergence" as a fig leaf.

What does real scientific research look like? Look at the work of Barabási and Albert in 1999. They discovered that complex networks like the internet and actor collaboration networks follow a power-law distribution—not pulled out of thin air, but grounded in two specific mechanisms: the network keeps growing, and new nodes preferentially connect to large nodes.

No mysticism, just two rules, and you can replicate the phenomenon.

That's why Professor Wen Xiaogang's translation of emergence as "演生" (emergent evolution) is spot on—it implies a dynamic process of "evolution plus generation," not just throwing a label on it.

---

Section 3: Is the "Emergent Ability" of Large Language Models Real or Fake?

I have to be straight with you about this one: I've verified both sides.

In 2024, a paper specifically debunked LLM emergence. The authors said: if you switch to a continuous metric—like using character accuracy instead of multiple-choice correctness—the so-called "emergent jump" becomes a smooth curve.

I tried it on my own dataset. Yeah, that phenomenon exists. After changing the evaluation method, the improvement from GPT-3.5 to GPT-4 looked linear.

But on the other hand, if you insist that emergence is purely a statistical illusion, that's not right either.

Using the same evaluation method, I scaled model parameters from 1.5B to 13B, all the way to 70B, and there really was a "threshold effect" for certain abilities: below a certain scale, the model simply couldn't learn syllogistic reasoning; once past the critical point, it suddenly could reason correctly, with accuracy jumping from 20% to 70%!

If that's not emergence, what is it?