Baixing.com Founder: Large language models devour everything, I only half believe that.

Author: Wang Jianshu, Founder of Baixing.com

Many people say, “A big model is everything.” I’m not convinced.

Every time I hear someone talk about “devouring everything,” I feel that we probably don’t yet understand the future at that level—so we blurt out such a sweeping, vague line. Otherwise, how could one thing truly devour everything? Take the internet: for so many years the internet has claimed it will “devour everything.” Has it really devoured everything now? So is it the internet devouring everything, or is it the big model devouring everything? Both are devouring—yet not a single thing is left?

So I’d rather change the wording: it is a very important foundation layer. Without this foundation, the whole world can’t develop—just like the internet can’t exist without that underlying backbone network, and electricity can’t exist without power plants. I agree with that.

But once you have the foundation, that’s where the lively part begins.

Take electricity. When electricity is first generated, what application do people notice first? The light bulb. Thomas Edison lit the first one, and then it kept shining—shining—shining. If the world had ended there, with only one light bulb, I could completely say: the power plant is the core of the entire world, and power plants devour everything.

But that’s not how it turned out. Later came the engine, to drive machines. Then you realize that once something at the bottom level like electricity exists, countless electrical appliances grow on top of it to use it. Washing machines are for washing clothes, TVs are for watching TV, vacuum cleaners are for vacuuming—they are all applications of electricity. Without electricity, none of these things would exist. But if you say “electricity devours everything,” I don’t believe it.

The same is true for big models. What it provides is basic intelligence. But this intelligence must be put into some specific “machine,” some specific “device” designed for particular scenarios, before it can work and truly change the world.

Claude Code is for writing code. Claude Design is for design. VoiceDrop is for writing articles. It’s the same big model—once you put it into different devices, it solves completely different problems.

Just having electricity, or just having water—without a washing machine—clothes still can’t be washed. Imagine this: the power plant produces massive amounts of electricity, and electricity is extremely powerful. But then what? Without a washing machine, can this pile of electricity wash the clothes by itself?

Intelligence is great, but most things in the world require multiple elements coming together in order to work. Just as a washing machine needs electricity, water, and even elements like a drum assembled together. A big model may be able to replace many things in the software domain, but in the world there aren’t many application scenarios that only require a single element.

Let me give an example right in front of us. Now we have big models, but just having big models isn’t enough. On top of them, there needs to be a layer called Harness—this layer has only emerged recently. It is what establishes the relationship with code generation, and only then can truly usable things be formed. A big model by itself can’t write code. Of course, the core of Claude Code—honestly—I could write in about fifty-something lines. If it’s longer, I just add a few more lines and it can run and write programs. But you have to see this: without that outer layer, even big models alone are still hard to use—that is, the big model’s intelligence isn’t combined with the code execution capability provided by the operating system. Relying on the big model to do math calculations isn’t economical, and in some cases it’s even impossible.

The core value of this interface layer is to help us take that intelligence—like electricity and water—and put it into a specific application scenario, turning it into a machine that can solve a specific problem.

At this point, of course, I still don’t completely disbelieve the logic behind “devouring everything.”

What it mainly refers to is existing software. Up to now, we’ve already piled up a huge layer of software—things put together from many rules, forms, buttons, and workflows. There are a lot of filters, fixed templates, a bunch of backend operations, and many SaaS detection features. And there are also all kinds of “M” things we’re familiar with—whether it’s CRM or HIS (hospital information system)—all sorts of so-called “systems,” “software,” and other similar things, all of them in abundance.

I think this layer of existing software, large language models will indeed devour a lot of it.

Why? Because these software were originally things that computers can execute—clear instructions that are solidified and repeatedly executed. That’s what we call software. And this is precisely what large language models are best at chewing through.

But.

Within this layer, besides software, there are many other things. Customer information. Execution capability—for example, when you book an airline ticket, the real capability that actually moves a plane and transports people from here to there. And trust. Many things in the physical world. I don’t think these will be devoured.

After it devours that layer, it actually opens up a larger space—above it, new types of software.

New types of software will definitely have a streaming interface. They may not need to solidify so many rules like before. Once you hand all those rules to the AI, think about it: previously, we managed to build things like a Salesforce-style CRM, which was already the peak of human capability—we expended every bit of effort possible. But if this part becomes relatively easier to solve, then what everyone does next is to unlock even more imagination and more possibilities above it—and that part is exactly what we still haven’t seen.

The mistake we often make is right here. When a new technology arrives, because we can’t see the bigger road after it, we can only stare at the portion right in front of us. A leaf blocks the view, preventing us from seeing Mount Tai.

Don’t say these are just trend judgments. I remember in 2004, a group of friends got together and complained that the internet could never produce any company bigger than Sina, Sohu, or NetEase—that the internet was basically ending and they would monopolize everything. But only after a few years, what happened? Everything flipped upside down. We would have to be embarrassed to death by our own short-sightedness back then.

So my position is this: is the big model important? Yes—it’s a foundation and the main focus of the most recent push. But once it becomes stable and can be provided continuously, it needs all kinds of “machines” and all kinds of “devices” on top of it to solve specific problems. That thick layer—where it’s used and how it’s used—is what will be the mainstream of the second wave of this wave of change.

The four words “devour everything” are too imprecise. Is there any thing, any kind of social form, any kind of technology in the world that has truly devoured everything?

In the places where it devours, to find the opportunities within—that is the truly important thing.

LAYOUT REFERENCE (source): total_lines=55, non_empty_lines=27, blank_lines=28