Refining the Domain Model Thesis

Refining the Domain Model Thesis | Generated by AI

Home 2026.07

Question: Is the argument “domain models are wrong; foundation models + cross-domain capability win” correct? Is Baichuan / Wang Xiaochuan’s medical AI strategy a mistake?

Answer:

Your intuition is pointing at a real trend, but the conclusion needs some refinement.

The key distinction is:

“Domain model” as a standalone foundation model vs “domain specialization built on top of frontier models.”

The first one is increasingly hard to win. The second one is exactly where many valuable companies will exist.

1. “Medical model” cannot beat GPT-class foundation models by data alone

Your point:

医疗那点文本数据有 10% 吗？可能很少

This is directionally correct.

A medical startup usually does not have:

trillions of tokens
huge multimodal datasets
massive RL infrastructure
frontier reasoning research

Compared with OpenAI / Google / Anthropic / DeepSeek:

Foundation model:
    pretraining
        ↓
    world knowledge
        ↓
    reasoning ability
        ↓
    coding/math/language capability

Medical AI:
    foundation model
        ↓
    medical fine-tuning
        ↓
    RAG
        ↓
    clinical workflow integration
        ↓
    hospital deployment

The medical layer is usually the second half.

A medical company saying:

“We have 100 million medical documents, therefore our model wins”

is similar to saying:

“We have many books about programming, therefore we can beat NVIDIA/CUDA.”

The bottleneck is not just knowledge.

It is:

reasoning
planning
tool use
multimodal understanding
long context
agent ability

2. But “classification is wrong” is only half true

AI itself is becoming more general.

Example:

A doctor today:

patient symptoms
       |
       v
medical knowledge
       |
       v
diagnosis
       |
       v
treatment plan

Future AI:

multimodal foundation model

+
medical knowledge graph

+
clinical database

+
hospital systems

+
medical tools

+
robotics/lab systems

The intelligence is centralized.

The applications are specialized.

This is similar to:

Linux kernel
    |
    +-- Android
    +-- Ubuntu
    +-- Kubernetes nodes
    +-- embedded devices

Nobody says:

“Operating systems are useless because Linux exists.”

The application layer still creates enormous value.

3. Where Baichuan’s strategic problem may be

The article’s strongest argument is not:

“Medical AI is wrong.”

It is:

“Medical AI before owning frontier model capability is risky.”

That is a much stronger argument.

Look at the timeline:

2023

Everyone thought:

Build LLM
    |
    v
Find applications

2024-2025

Reality became:

Strong model
    |
    +--> coding
    +--> agents
    +--> enterprise automation
    +--> research assistant
    +--> healthcare

Coding exploded because coding is:

digital
measurable
fast feedback
low regulation
global market

Medical is:

slow adoption
regulatory barriers
liability problems
expensive deployment

So the order matters.

4. Why Anthropic won with coding

This is the most important part.

Anthropic did not say:

“We are a coding company.”

They built a better general model.

Then coding became the wedge.

Claude

     |
     +--- coding agent
     |
     +--- enterprise assistant
     |
     +--- research
     |
     +--- healthcare

The model creates markets.

Not the reverse.

This is the biggest difference between Internet era and AI era.

Internet:

find users
    |
build product
    |
technology supports it

AI:

improve intelligence
    |
new capabilities emerge
    |
new products appear

5. Was Wang Xiaochuan “doesn’t understand LLM”?

I would phrase it more carefully.

He clearly understands:

NLP
search
language technology
product strategy
entrepreneurship

His background is actually very relevant.

But LLM changed the game.

Search era:

information retrieval
+
ranking
+
NLP

LLM era:

pretraining scaling
+
RL
+
post-training
+
agent loops
+
compute economics

The winner profile changed.

A great search engineer does not automatically become a great frontier model builder.

Same as:

A great compiler engineer does not automatically build NVIDIA GPUs.

Different bottleneck.

6. The strongest counterargument against the article

The article assumes:

“Whoever has the strongest model will win medical AI.”

Not necessarily.

Look at software history:

Microsoft did not invent the internet.

Oracle did not invent databases.

Salesforce did not invent relational databases.

The infrastructure layer and application layer can both create huge companies.

A medical AI company could win through:

GPT/Claude/DeepSeek
        +
medical data
        +
doctor workflow
        +
hospital integration
        +
regulatory approval
        +
distribution

The moat is not the model.

The moat is the system.

My current view

Baichuan abandoning frontier models completely

Probably risky.

Because:

weak model
+
medical data
=
limited ceiling

Medical AI itself

Not wrong.

Actually probably inevitable.

But the winning architecture is more likely:

Frontier foundation model

          +

Medical intelligence layer

          +

Agent workflow

          +

Hospital ecosystem

not:

small medical GPT
+
medical documents
=
future OpenAI

Your analogy with the 1970s-1990s is quite good:

When programming exploded, the biggest winners were not companies making “banking programming languages” or “hospital programming languages.”

They built:

CPUs
operating systems
compilers
databases
networks

Then every industry was built on top.

AI is likely similar:

The scarce layer today is general intelligence infrastructure.

But once that layer stabilizes, domain systems will still be enormous. The mistake is confusing domain expertise with model intelligence.

Back Donate