Building a Token Factory | Generated by AI
Question: How to build a “token factory” — a small AI factory where engineers each have 10–20 GPUs/computers and operate as super-agents running many experiments?
Answer:
What you are describing is closer to an AI research factory / agentic lab than a traditional software company.
The core idea:
Human engineers become “research directors”. Their agents become thousands of virtual researchers. GPUs become the experimental production line. Tokens are the output.
Think of how semiconductor fabs work:
- TSMC has machines → produces wafers
- AI factory has GPUs → produces tokens, models, experiments, data
A small version could look like this:
AI Factory
Founder / Research Lead
|
-----------------------------------
| | |
Agent Team A Agent Team B Agent Team C
Data Agent Training Agent Eval Agent
Coding Agent Debug Agent Benchmark Agent
Paper Agent Infra Agent Deployment Agent
| | |
10 GPUs 10 GPUs 10 GPUs
|
v
Experiments / Models / Datasets / Agents
1. The unit of production is not “employee”
Traditional company:
1 engineer
|
v
write code
|
v
ship feature
AI factory:
1 engineer
|
v
creates agent workforce
20 agents
|
v
100 experiments/day
|
v
best ideas survive
The engineer becomes like a principal scientist managing a colony.
2. Each engineer needs a “personal AI lab”
Example:
Engineer workstation:
RTX 5090 x 4
+
server GPU access
+
agent runtime
+
experiment database
Software:
~/factory/
agents/
researcher.py
coder.py
critic.py
evaluator.py
experiments/
exp001/
exp002/
models/
datasets/
logs/
Every experiment automatically records:
experiment_id
model_version
dataset
hyperparameters
GPU hours
result
leaderboard score
3. The key is massive parallel experimentation
Humans are bad at trying 100 ideas.
Agents are good.
Example:
You want improve a small LLM.
Human says:
“Try changing attention.”
Agent swarm:
Agent 1:
replace attention with MLA
Agent 2:
increase RoPE theta
Agent 3:
change tokenizer
Agent 4:
synthetic data generation
Agent 5:
curriculum learning
Agent 6:
optimizer experiments
Agent 7:
architecture search
After 24 hours:
Experiment #431
beats baseline +7%
merge
This is how AlphaGo-style research works.
4. Build the “factory operating system”
The most important component is not GPUs.
It is orchestration.
Something like:
Task Queue
|
v
Agent Scheduler
|
-------------------------
| | |
GPU #1 GPU #2 GPU #30
|
v
Experiment Runner
|
v
Results Database
|
v
AI Reviewer
Possible stack:
Ray
Kubernetes
Slurm
Docker
vLLM
PyTorch
MLflow
Weights & Biases
PostgreSQL
But honestly a small company can start simpler:
Python
+
SSH
+
Docker
+
SQLite
+
Git
5. The “super-agent employee”
A future AI engineer may look like:
Morning:
Human:
"Improve our 7B model for coding."
Agent:
08:00
reads papers
08:30
creates 50 experiment plans
09:00
launches training jobs
12:00
evaluates benchmarks
15:00
writes report
18:00
human reviews top 3 ideas
The human spends time on:
- choosing direction
- judging quality
- talking to customers
- architecture decisions
6. Small OPC version (one-person company)
Actually one person can build a miniature version.
Example:
Hardware:
Your PC:
RTX 4070
+
2 x RTX 3090 server
+
cloud GPU when needed
Agents:
Research Agent
|
+-- Paper reading
Coding Agent
|
+-- Implementation
Training Agent
|
+-- GPU jobs
Critic Agent
|
+-- Finds problems
Marketing Agent
|
+-- Writes content
Your daily output:
10 experiments
3 blog posts
2 open-source contributions
1 product iteration
7. The biggest bottleneck will change
Today:
GPU shortage
Future:
attention shortage
Everybody will have agents.
The advantage becomes:
Who has the best:
- experiment loop
- evaluation system
- taste
- data pipeline
- customer feedback
Your idea is very close to what many frontier labs are moving toward:
Old:
100 researchers + 100 GPUs
New:
10 researchers + 1000 agents + 1000 GPUs
The company becomes a learning machine.
A good name for this concept would be:
“AI factory = a self-improving experiment engine that converts compute into intelligence.”
The interesting part is that an OPC with strong engineering background can actually build a tiny version today. The first milestone is not 100 GPUs — it is making one agent reliably create, run, evaluate, and improve experiments without you touching every step.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋