Craig Martell: Shall We Play a Game?
A DEF CON 31 talk about the risks of over-relying on large language models. The speaker explains that LLMs are fundamentally statistical systems that predict the next word, not true reasoning engines.
He warns that their ability to generate convincing text leads people to trust them too much, despite errors and “hallucinations.”
The talk highlights the need for measurable evaluation metrics, real-world use cases, and the role of hackers in identifying weaknesses in AI systems — especially in critical domains like defense.
He warns that their ability to generate convincing text leads people to trust them too much, despite errors and “hallucinations.”
The talk highlights the need for measurable evaluation metrics, real-world use cases, and the role of hackers in identifying weaknesses in AI systems — especially in critical domains like defense.