Yann LeCun and Andrew Ng, Why the 6-month AI Pause is a Bad Idea

07-04-2023

  • url: https://www.youtube.com/watch?v=BY9KV8uCtj4
  • talks related to https://futureoflife.org/open-letter/pause-giant-ai-experiments/
  • why reduce innovation, why reduce progress of knowledge
  • tradeoff between good and bad use cases in ML
    • bias, fairness, concentration of power
    • healthcare
  • concerns on AGI
    • future systems will become much more controllable, factual, non-toxic
      • started with general language models and now we have instruction models, which add more safeguards
      • design objective for these systems aligned with human values
  • considerations around economy / regulations / politics
    • small number of companies creating models
    • regulation is required, not stopping R&D or product development
    • maximize positive effects, minimize negative effects
    • AGI escape is an overhyped risk
      • governments have the ability to shut down research labs, servers, there is greater control
    • too early to impose bans when we are still premature in the development of this technology
      • can't build seatbelts if cars don't exist
    • release technology in a controlled way
      • difficult to identify safety concerns before they occur, with regulations and safety steps put in place after the fact
    • there should be more open research, but there needs to be a distinction between research development and product development
  • comparison to history
    • printing press
      • Ottoman empire stopped usage, became intellectual backwater
      • for societies that adopted it, dissemination of enlightenment, science, democracy
    • research on recombinant DNA in the 1960s
      • conference put in place various levels of containment policies
  • AGI
    • understanding and hype around AGI differs across domains
      • Moravec's paradox
        • we don't have level 5 self-driving cars
          • takes a human around 20 hours
        • we have performant language models
          • trained on a trillion words which would take humans 22000 years reading 8 hours a day
  • references