Day 23 - How LLMs Learn and Respond: Training vs Inference
Context:
I wanted to understand the flow of data in Large Language Models (LLMs) — how they learn during training and respond during inference.
What I Learned:
- Training Data:
- Used during the training phase of the LLM.
- Includes grammar, facts, reasoning, code, articles, books, etc.
- Stored as the model’s built-in knowledge.
- Inference Data (Input):
- The user’s prompt or question during inference time.
- No data is stored or learned from the user’s prompt.
- Output Data:
- The model’s response to the user’s input during inference time.
- No data is stored or reused.
Key Insight:
Training builds the brain. Inference is the conversation.
Better inputs (prompt engineering) → Better outputs.
Why It Matters for QA / AI Testing:
- Helps testers understand why models can’t “learn” from prompts in real-time.
- Emphasizes the importance of prompt engineering for accurate and useful responses.
- Critical for designing tests that separate training limitations from inference behavior.
My Takeaway:
LLMs learn during training, not during conversation. Prompt quality is the key to unlocking better outputs.
