Why Base LLMs Aren't Enough | Beyond the Prompt

Lesson 1: Why Base LLMs Aren't Enough

Blackboard

Five problems you will hit

10 min

Concept 1

Domain knowledge gaps

A base LLM is trained on broad internet data. It will lack specialised knowledge — a farming AI that needs to judge crop quality, a biotech tool categorising reviews, or a legal assistant where every word carries weight. The model knows a little about everything but not enough about your thing.

Concept 2

Outdated and hard to control

LLMs have a knowledge cutoff — they cannot follow trends or recent events without help. They are also difficult to steer: Microsoft's Tay bot went rogue in 16 hours, and even well-funded labs struggle to prevent controversial outputs. Control is an unsolved problem.

Concept 3

Task-specific underperformance

A model may be broadly capable but not precise enough for your task. A biotech company's idea of a 'negative' review might differ from the general sentiment. Enterprise applications need high precision, low latency, and alignment to company-specific definitions.

Concept 4

Lack of sources

Base LLMs do not cite their work. In fields like medicine, law, research, and education, an unsourced answer is an untrusted answer. Worse, when asked for sources, models often hallucinate fake citations that look convincing.

Pause and Think

Think of one task where a general-purpose chatbot gave you a confidently wrong or unhelpful answer. What was missing?

Flashcards

Key Terms

Click or press Enter to flip a card. Review these before the practice question.

Additional Resources

Article — Why LLMs Still Hallucinate (Duke)
Duke Libraries piece arguing hallucination is baked into how generative models work, not something that will get patched out. Written for a general audience — worth passing to teammates who think the problem is temporary.
Paper — LLM Limitations Survey
Big survey mapping out where LLMs break — reasoning, knowledge, alignment. The value is in the taxonomy and the long reference list, not any one finding.
Glossary — Large language model (LLM)
A short definition with related terms.
Glossary — Knowledge cutoff
Why base models do not know recent events.

Practice Question

A biotech company finds that their LLM classifies slightly negative reviews as 'neutral' instead of 'negative'. What is the most likely cause?

Does not count for certificate

Choose one answer, use the hint if needed, then check your answer.

Use the hint before checking if you need a nudge.

Choose one answer, then click Check Answer.