Cooking as AI Training Metaphor

Cooking as AI Training Metaphor | Generated by AI

Home 2026.07

Question: Is the idea of cooking as a metaphor for data/model training and deep conversations about AI, languages, and science a good analogy?

Answer:

Yes, this analogy is actually very close to how modern AI engineering feels.

Cooking:

Buy different ingredients:
- vegetables
- meat
- spices
- fruits

↓

Machine learning:

Collect different data:
- books
- code
- conversations
- scientific papers
- images
- videos

The quality is not only about having “more ingredients”. It is about:

Ingredient diversity = data diversity

A soup with only potatoes is boring.

A model trained only on one domain becomes narrow:

only code -> good coder, weak general reasoning
only math -> good equations, weak world knowledge
only internet text -> broad, but noisy

The magic comes from mixing:

code + math + science + literature + conversations + multilingual data

similar to a good hotpot.

Container size = model capacity

Your analogy:

不同容器大小，水量如何，不同做法

is basically:

model size
+
dataset size
+
training recipe

A small pot:

7B model
8GB VRAM
small dataset

cannot cook a huge feast.

A giant pot:

405B model
H100 cluster
trillions tokens

can absorb much more complexity.

But a giant pot with bad ingredients is still bad.

Cooking method = training recipe

Same ingredients:

beef + vegetables + water

can become:

soup
stir fry
barbecue
hotpot

Same data:

FineWeb
GitHub
books
papers

can produce different models:

GPT-style base model
coding model
reasoning model
chat model

The “chef skill” matters.

This is why Karpathy often emphasizes:

data + architecture + optimization + evaluation

not just parameter count.

“无他，手熟而已” is actually very ML

A lot of frontier AI work is accumulated craft:

train
 ↓
observe loss curve
 ↓
change data
 ↓
change tokenizer
 ↓
change LR schedule
 ↓
change architecture
 ↓
repeat

The final recipe looks simple:

model.fit(data)

but behind it are thousands of small decisions.

Like a chef saying:

“Just cook it.”

But the difference between a normal dish and a Michelin dish is 20 years of small adjustments.

And your second point:

超喜欢和人聊天，只是不讲枯燥车轱辘讲，用日语讲，用英语讲，讲 AI，谈天说地，从王垠讲到牛顿

This is actually the same idea as a good language model.

A good conversation is not:

question -> answer -> question -> answer

It is:

human experience
       +
history
       +
science
       +
language
       +
humor
       +
personal curiosity

A person who can jump:

王垠
 ↓
programming philosophy
 ↓
AI agents
 ↓
Karpathy
 ↓
Newton
 ↓
physics
 ↓
Japanese culture

is doing high-dimensional knowledge navigation.

That is also what we want from models:

not just retrieval, but connecting distant concepts.

Your cooking analogy is close to the core idea of intelligence:

Intelligence is not memorizing ingredients. It is learning how ingredients combine into new dishes.

The interesting future is probably not “bigger pot forever”, but better chefs + better ingredients + better recipes.

Back Donate