Mira Murati’s Enterprise AI Strategy | Generated by AI
Question: Is Mira Murati’s Thinking Machines bet on open/customizable foundation models + enterprise fine-tuning likely wrong because enterprise data is too small, and the real competition is pushing intelligence/agents like GLM or Anthropic?
Answer:
Your intuition is pointing at a real tension in the AI market, but the conclusion is slightly too strong.
The key distinction:
“Enterprise fine-tuning” ≠ “training intelligence from enterprise data.”
There are at least four different layers:
Internet-scale pretraining
|
v
Base model intelligence
|
v
Post-training / alignment / reasoning
|
v
Enterprise adaptation
|
+--> RAG
+--> tool use
+--> workflows
+--> fine-tuning
+--> agents
The mistake is assuming enterprise fine-tuning is trying to create a smarter model.
Usually it is not.
1. You are right: enterprise data is tiny compared with internet data
For general intelligence:
Internet corpus:
trillions of tokens
Single enterprise:
millions ~ billions of tokens
Even a huge company like Microsoft has:
- internal documents
- emails
- code
- tickets
- meeting transcripts
but compared with:
- Common Crawl
- GitHub
- books
- papers
- Wikipedia
- web forums
it is small.
So:
enterprise data -> smarter GPT
is probably false.
A bank’s documents will not create a better Claude.
2. But enterprises don’t buy “smarter GPT”
They buy:
GPT-5 + company knowledge + company actions
Example:
A bank does not need:
“a model that knows banking better”
It needs:
Customer request
|
v
Agent
|
+--> search internal policy
|
+--> query transaction system
|
+--> call compliance API
|
+--> generate response
The value is not the weights.
The value is the integration layer.
3. Fine-tuning is becoming less important
I agree with you here.
The 2023 idea:
“Every company will fine-tune its own Llama”
was probably exaggerated.
Why?
Because models improve quickly.
Example:
Company trains:
Llama-3.1 fine-tuned Jan 2025
Six months later:
Claude/GPT/Qwen new model
beats it.
The company now has:
- old model
- expensive maintenance
- unclear benefit
So many enterprises move toward:
strong foundation model
+
RAG
+
tools
+
agent framework
4. The real battlefield is agent capability
This is where your point is strongest.
The frontier is:
Can the model reliably complete long tasks?
Not:
Can it memorize more enterprise PDFs?
The competition:
Anthropic
Focus:
Claude
+
computer use
+
tool use
+
coding agents
OpenAI
Focus:
reasoning models
+
agent infrastructure
+
operator-like systems
GLM / Chinese labs
Focus:
open models
+
efficient reasoning
+
agent ecosystems
The winner probably needs:
better base intelligence
+
better reasoning
+
better planning
+
better tool execution
+
lower inference cost
5. Then why is Thinking Machines interesting?
The bet is not necessarily:
“enterprise data makes models smarter”
The bet may be:
“companies want control over AI systems”
For example:
Open/custom models allow:
Company
|
+-- choose model
|
+-- inspect behavior
|
+-- customize safety
|
+-- deploy private
|
+-- integrate agents
This matters for:
- banks
- governments
- healthcare
- defense
- large enterprises
They often cannot send everything to OpenAI/Anthropic.
6. The future probably looks like Linux + cloud
A possible analogy:
1990s:
Every company does not build its own OS.
But:
Linux kernel
+
company customization
+
enterprise software
became huge.
AI may become:
foundation model
+
company agent layer
+
private data/tools
The model itself becomes commodity.
The agent system becomes differentiation.
My prediction
The market splits:
Frontier labs win
training intelligence
reasoning
multimodal capability
agent algorithms
(OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek, etc.)
Enterprise AI companies win
deployment
integration
workflow automation
security
evaluation
Thinking Machines is betting on the second layer.
The risky part is if they are too focused on:
fine-tune your own model
The stronger thesis is:
customizable agent systems built on frontier/open models
because, as you said, the scarce resource is not enterprise text.
The scarce resource is making models reliably do useful work.