1.1.2. LLMs as Probabilistic Suggestion Engines
💡 First Principle: A large language model turns your context into a probability distribution over what comes next, then samples from it. "Probabilistic" is the key word: the same prompt can yield different suggestions, and a confident-looking answer is not a guaranteed-correct answer.
This is the engine under every Copilot surface. The model was trained on a vast corpus of public code and text, so it has learned statistical patterns — which token tends to follow which, given the surrounding code. It does not look anything up in a database, and it has no built-in notion of "true." It produces what is plausible given its training and your context.
Three consequences fall straight out of this, and all three are tested:
- Non-determinism. Ask twice, you may get two different completions. This is by design, not a bug. (A traditional formatter or linter is deterministic; Copilot is not.)
- Hallucination and error. The model can invent a plausible-looking API that does not exist, or suggest insecure code, because "plausible" and "correct" are different targets.
- Quality follows the training data. Well-represented languages like JavaScript or Python tend to get stronger suggestions than niche or proprietary languages, simply because the model saw more of them.
❌ Common Mistake: Assuming that because Copilot was "trained on real code," its output is automatically correct or secure. Training teaches patterns, not guarantees. The developer owns verification.
A useful intuition: Copilot is autocomplete's ambitious cousin, but it predicts meaning-shaped text, not just the next word. That power is exactly why it can be wrong in subtle, fluent ways — the errors read as confidently as the correct answers.
⚠️ Exam Trap: When a question implies Copilot will "always" produce a specific output, or that it "knows" the latest version of a library, be suspicious. LLMs have a fixed context window and a knowledge cutoff. Copilot only reasons over the context it is handed.
Reflection Question: Why can two developers send what looks like the same prompt and receive different suggestions — and why is that a property of how the model works rather than a malfunction?
