LLM Limitations

Introduction

LLMs are powerful but flawed. Knowing their limitations is what separates a hobbyist from a production AI engineer.

Beginner analogy: Like a confident but occasionally wrong intern — brilliant in many areas, but you must double-check critical work.

Understanding the topic

Core concepts to understand:

• Hallucinations — confident, fluent, sometimes wrong.
• Knowledge cutoff — no awareness of events after training date.
• No real reasoning — pattern-match, not deduction (mostly).
• Context limits — can't 'remember' beyond the window.
• Bias & safety — reflects training-data biases.
• Latency & cost — slow + expensive at scale.

Syntax reference

Visual workflow / architecture:

Strengths             │  Limitations
──────────────────────┼──────────────────────────
Fluent text           │  Hallucinations
Few-shot generalising │  Knowledge cutoff
Code generation       │  Math & arithmetic
Summarisation         │  Long-range memory
Translation           │  Bias & safety
                      │  Cost & latency

Real-world use

Air Canada was sued (and lost) when its AI chatbot invented a refund policy. NYC's MyCity chatbot told business owners to break the law. Knowing limitations isn't optional.

Best practices

• Always validate critical output.
• Use RAG for factual freshness.
• Add guardrails for safety-critical domains.
• Keep humans in the loop for high-stakes calls.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is an LLM hallucination?
• Q2. How do you mitigate the knowledge-cutoff problem?
• Q3. Why are LLMs bad at math?
• Q4. Give an example of an LLM safety incident.