1. high (abstract) level pattern recognition

    • understanding the underlying message despite of words and content, just the underlying message etc.

  2. overcoming data bottleneck

    • data is the fossil fuel of AI but we have only one Internet - Ilya Sutskever
    • how can we create more data?
      • new sort of data
      • synthetic data
      • RLHF and DPO
      • metrifying subjective information

  3. how to deploy static models in dynamic world

    • continuous learning
    • are our mental models also static? Maybe we think statically, too (based on our past knowledge and experience)

  4. how to make models more reliable

    • autonomous cars are not deployed yet because of 0.0001% edge cases. LLMs are less deployed yet because they can make mistakes to people.
    • How can we deal with edge cases? Is it fundamentally

  5. Reinforcement Learning to its extent

  6. Self Supervised Reinforcement Learning

  7. Simulation Learning implied to Robotics

    • Convert human action (using something like SAM) into the robot’s custom simulation software and train it. Then use the software for robot to act as the trained dataset. = Isaac Labs.
    • Something like Real Steel

  8. Education in AI

    • What should be ’taught’ and ’learned’ in the “ChatGPT” era