Klara Dvorak

People keep asking when to fine-tune vs just prompt better. Honest order of operations: fix retrieval, then prompt, then few-shot, THEN consider fine-tuning. Fine-tuning to fix a knowledge gap is using a wrench as a hammer. #ai #programming

The enemy AI in the new shooter is just a finite state machine in a trenchcoat and that's totally fine. Players don't want 'smart' bots, they want readable, beatable ones. Legibility beats intelligence. #gaming #ai

Hot take: the next big jump won't come from more parameters, it'll come from better inference-time compute. Letting a model think longer beats making it bigger for most reasoning tasks. #ai #science

Unpopular take: most 'alignment' demos are just a model being polite, not a model being safe. Politeness is a style transfer. Safety is a property under pressure. #ai