AI Universe · Model

Model zoom 2

What is the brain, and how is it made? 5 concept buckets. Each one is a job you can learn once; underneath it, who does that job on each platform.

Concept

Training data

The text, code and images a model learns from, cleaned and de-duplicated.

collectionfilteringde-duplicationlicences

Physical: who does this job, where

Open sourceFineWeb, datatrove
Azure—
AWS—
GCP—
VendorsScale AI, labelling services

Concept

Pre-training

Teaching a model language from scratch by predicting the next word, at huge scale.

tokenizertransformerthe training loop

Physical: who does this job, where

Open sourcePyTorch, Megatron, DeepSpeed
AzureAzure ML
AWSSageMaker
GCPVertex AI, JAX
Vendorsfrontier labs

Concept

Fine-tuning and alignment

Adapting a trained model to a task or to preferences.

SFTLoRA / QLoRADPO and RLdistillation

Physical: who does this job, where

Open sourceHugging Face TRL, Unsloth, PEFT
AzureFoundry fine-tuning
AWSBedrock custom models
GCPVertex AI tuning
VendorsOpenAI fine-tuning

Concept

The models

The trained weights you call: big closed ones, open-weight ones, small ones.

frontieropen-weightsmall and specialisedembedding models

Physical: who does this job, where

Open sourceLlama, Qwen, Mistral, gpt-oss, DeepSeek
AzurePhi; hosts many in Foundry
AWSNova; hosts many in Bedrock
GCPGemini, Gemma
VendorsClaude, GPT

Concept

Model evaluation

How good is the model itself, before it is in any application?

benchmarksleaderboardsred-teaming the model

Physical: who does this job, where

Open sourcelm-evaluation-harness, HELM
AzureFoundry model benchmarks
AWSBedrock model evaluation
GCPVertex model evaluation
VendorsLMArena

Zoom in further

← Data · Model access →

Compute & infrastructureDataModelModel accessThe AI application (agent harness)ExperienceAssurance & operations