Making Sense with Sam Harris #312 — The Trouble with AI
358 snips
Mar 7, 2023 
Stuart Russell, a UC Berkeley professor and author of 'Human Compatible,' and Gary Marcus, a renowned scientist and author, delve into the complexities of artificial intelligence. They explore the limitations of current AI technologies, especially ChatGPT, and the ethical dilemmas surrounding artificial general intelligence. The duo discusses the risks of misinformation, the need for human values in AI systems, and the urgent call for regulations to protect democracy and public safety amid evolving tech. They reveal how business models can exacerbate misinformation crises.
AI Snips
Chapters
Books
Transcript
Episode notes
Go Program Weakness
- Superhuman Go programs, despite initial dominance, are beatable by humans exploiting specific weaknesses.
- This reveals a failure of these programs to truly understand the game's concepts.
Expressive Power Matters
- Deep learning models, like circuits, struggle with expressive power. This leads to overcomplicated representations and limits their ability to generalize.
- They haven't truly learned underlying concepts, making them vulnerable to specific exploits.
AI's Lack of Understanding
- Current AI systems, particularly those based on deep learning, do not possess genuine understanding or world models.
- They excel at approximating text but lack symbolic representation and reasoning abilities.
