Scraping Bits #122 - Tom Burns: Digital Neurosience, Fully Autonomous Agents, AI Safety + Control
Dec 2, 2024
In this talk, Tom Burns, a postdoctoral researcher at Cornell specializing in neuroscience-inspired AI, dives into the fascinating parallels between human brain processes and AI. He discusses the importance of geometry in AI safety and memory, and how emotional factors influence decision-making in both humans and AI systems. Tom addresses ethical concerns surrounding generative AI and the necessity for dialogue among stakeholders to ensure responsible governance. The potential of hydrogen-based computing compared to traditional silicon is also explored, hinting at a promising future.
AI Snips
Chapters
Books
Transcript
Episode notes
Attention and Memory
- Memory items are represented by interconnected neuron groups.
- This maps to the attention mechanism in transformer models, focusing attention on objects and recalling memories.
Digital Neuroscience
- Understanding AI systems involves interpreting their actions, similar to neuroscience.
- This "digital neuroscience" explores causal relationships and how external factors impact AI.
AI's False Information
- An Australian mayor sued OpenAI because their model fabricated scandals about him.
- This highlights AI's potential for generating false information and the need for better safety measures.
