wab
wab — yIlaD.
{study notes}
- Generative AI - Qapla'! {Generative AI} creates new content, not just analyzing existing data. Different from other AI types.
- Three foundations of modern {LLMs} - {Transformer architecture}, vast training data, and massive compute power make Claude and similar systems possible.
- Training stages - Pre-training learns patterns from billions of examples; fine-tuning teaches the system to follow instructions helpfully.
- Current strengths - Versatility across many tasks, conversational ability, and connection to external {tools}. Speed and scale surpass human capability.
- Current limitations - Knowledge cutoffs, {hallucinations}, {context window} size constraints, and difficulty with complex multi-step reasoning. Cannot be trusted without verification.
- Best practice - Pair human judgment, creativity, and oversight with AI's speed and scale. Humans must verify critical information.
- Testing strategy - Use domains you know well to spot AI errors. Test versatility, {hallucinations}, and knowledge cutoffs with real examples from your work.
{flashcards} 10 {cards}
{question}
yI'ang · ←/→
{answer}
{flip}
{knowledge check} 6 {questions}