Goals

• Understand recent key advances in AI capabilities
• Understand how AI could solve global problems
• Understand existing risks and possible risks of advanced AI

Visualizing the Deep Learning Revolution

https://medium.com/@richardcngo/visualizing-the-deep-learning-revolution-722098eb9c5

• Ideas
  • Deep Learning driven revolution in AI capabilities over the last decade
  • Driven by scaling up a few simple algorithms, not by scientific understanding of deep learning.
  • Could not be predicted; whoever predicted, also predict existential risks.
• Vision
  • Image recognition - better than human performance on many datasets
  • Image generation - bigger models, more data, also new algorithms like GANs, transformers, diffusion models
  • Video generation - sora
• Games
  • Video games - from atari to dota and minecraft.
  • Strategy games - superhuman level at go by self play, human level at Diplomacy board game
• Language based tasks
  • Text generation - couldn’t put together a sentence in 2011. now it can tell you your fav receipe as a rhyme
  • Common sense QA - GPT-4 capable of giving correct answers to variety of common sense questions
  • Planning and strategic thinking - can generate plans but limited ability to execute on them
• Science
  • Coding - Codex, Copilot, AlphaCode - top 54% in coding competitions
  • Math - MINERVA model, AlphaTensor discovered a faster algo for matrix multiplication
  • Life sciences - AlphaFold 2 outperformed the rest in predicting the protein structure
• Significant no. of breakthroughs happened in recently
• Becoming difficult to have benchmarks to track progress because they don’t remain unsolved
• Predictions on progress has become harder, progress happening faster than expert predictions
• Compute used in training is growing up dramatically, compute seems to be the most accurate factor in predicting the progress
• AGI may come soon? In a decade? The possibility that AGI might not be aligned with our values and that might be catastrophic.

Four Background Claims

https://intelligence.org/2015/07/24/four-background-claims/

• Why is the mission of ensuring the creation of smarter-than-human AI has a positive impact is important? Four claims underlie this premise