Workshop organized by the Real World Risk Institute. The workshop is an intense 10-day online program, and the 18th edition took place from July 10-21, 2023.
This video discusses the capabilities and limitations of large language models like GPT, the challenges of setting constraints on AI systems, and the potential risks and consequences of AI decision-making. The video talks about:
- The concept of a “stochastic parrot” in language processing and machine learning.
- How language processing systems like GPT use data from the web to generate responses.
- Attempts to “trick” GPT with questions requiring nuanced understanding.
- The simple operation of GPT in predicting the next word in a sequence.
- The use of language models as a new interface to computers.
- The integration of GPT with Wolfram Alpha for computations and informed responses.
- The similarity between writing good prompts for GPT and expository writing.
- The training data for GPT, includes nonsense, fiction, and factual information.
- The problem of the “self-licking lollipop” in information sources.
- The concept of “necessarily human work” requires human choice and input.
- The potential for AI to make decisions and the challenges of setting constraints.
- A thought experiment called “promptocracy” for AI decision-making.
- The actuation layer of AI and the difficulty of setting constraints.
- The phenomenon of computational irreducibility and trade-offs in AI computation.
- The potential risks of AI decision-making and the need for understanding large language models.
