The Most Interesting Thing in A.I. A Different Approach to AI - with Nicholas Thompson and Gary Marcus
26 snips
May 14, 2025 
In a thought-provoking discussion, Gary Marcus, a prominent cognitive scientist and entrepreneur known for his critical perspective on AI, joins Nicholas Thompson. They dive into the shortcomings of today’s AI models and advocate for a balanced approach to development. Topics include the limitations of scaling laws, the need for neurosymbolic AI, and the financial challenges facing leading companies. They also explore the potential risks of AI in democracy and discuss future capabilities in tasks like cooking and coding, emphasizing the importance of ethical practices.
AI Snips
Chapters
Books
Transcript
Episode notes
ChatGPT's Early to Improved Answers
- At ChatGPT's launch, it gave amusing but flawed responses like the biking to Maui question.
- Now, responses have improved with more accurate reasoning, showing fast learning and adjustment.
Scaling Laws Are Breaking Down
- Scaling laws in AI, which suggest more data and compute means better models, are failing now.
- This challenges the belief that simply increasing scale is the path to artificial general intelligence.
Integrate Neural Data with Symbolic Reasoning
- Extract abstract concepts from neural networks to feed symbolic systems for deeper reasoning.
- Develop techniques to enable smooth interaction between learned knowledge and symbolic rules.
