01.05

Pretraining vs post-training

Pretraining installs what the model knows and post-training installs how it behaves; reasoning models let you buy quality with extra thinking, and none of it makes the output deterministic.

• Making AI Product Decisions
• 16 min
• Free

Who's hiring for this skill

30 companies have live AI-PM roles that ask for skills taught in this lesson.

• Nebius
• Anthropic
• Scale AI
• Intercom
• Rogo
• Writer
• Snowflake
• Harvey
• Asana
• Gusto
• Fireworks
• Mistral
• OpenAI
• Okta
• CoreWeave
• Cohere
• Coursera
• Reddit
• 1Password
• Deepgram
• Klaviyo
• Lightning Ai
• Luma AI
• Snorkel AI
• Speak
• Thinking Machines
• Drata
• Eve
• LangChain
• Plaid

Concepts in this subchapter

• Pretraining
• Post-Training
• Supervised Fine-Tuning
• RLHF
• Verifiable Rewards
• Reasoning Model
• Reasoning Tokens
• Reasoning Effort
• Knowledge Cutoff
• Nondeterminism

Source video: How language model post-training is done today, Interconnects AI.

Keep going

• Previous: 01.04 Context windows and attentionFree
• Next: 01.06 The AI-or-not decisionFree
• 01 Making AI Product Decisions9 lessons
• The full curriculum
• AI PM Knowledge Graph219 concepts