DevelopersTV

LLMs Don’t Learn Like Humans - They’re Populations

Large language models Llm architecture Machine learning Ai research Neural networks Hallucinations Sycophancy Temperature sampling Tokenization Benchmark testing Population models Nlp
InfoQ
AI summary

AI researcher Naomi Saphra explains why LLMs behave as statistical populations rather than individual reasoners, covering fundamental issues like why traditional benchmarks fail, how temperature settings mimic crowd wisdom, and the architectural roots of common pitfalls including hallucination, sycophancy, and tokenization bugs. This talk is designed for engineering leads and software architects building production AI systems who need to understand these core principles to prevent systemic failures.

Related videos