Gang Fang's Blog

This talk has been phenomenally interesting and here are some notes:

• Geoffrey: currently LLMs are probably a higher form of intelligence because, to the contrary of popular belief, it is statistically efficient. He got this conclusion when he compared the ratio between the amount of knowledge packed into a LLM model and size of it to that of human: an individual human knows a lot less even though a single brain is bigger. And this higher form of intelligence only evolves after human intelligence because of its energy intensity (i.e., we have to build it). An implication of this is there might not be many inspirations neuroscience can draw from AI because AI is less and less like a brain, even though in the early days, AI drew a lot of ideas from neuroscience.
• David Siegel: there needs to be a theory of intelligence. Intelligence can have different forms: analog and digital. These different forms will provide different perspectives and facilitate the development of the understanding or theory of intelligence.
• Demis: AI systems are increasingly engineering heavy and the paradigm is one of engineering science: we analyze the engineered artifacts and seek to understand what they do. A lot of work needs to be done on interpretability and our understanding of the models.
• Pietro Perona: the AI scientist idea. We need to build machines that not only rely on correlations to predict, but also understand causations and be able to carry out experiments like a human scientist. Also, we might need embodied intelligence to enable machines to act in the physical world.
• Ilya: re theory, it won’t be like physics, but it will be a focus on chaos and complexity.
  • A research idea is to to perform psychophysical experiments on open-source models
• One stated problem with current approach is implicit knowledge can’t be captured because language only contains explicit knowledge. However, I am doubting this argument because reasoning can be put into language and be captured. The implicit knowledge is more like intuition, which is ironically what the current models are good at.