Sam Altman — AI for the Next Era (Greylock)

OpenAI’s Sam Altman in conversation with Reid Hoffman at Greylock, September 2022.

Why it was kept

Kyle’s notes: a serious search competitor to Google; language as the interface; multimodal models; a platform shift not seen since mobile; skepticism of startups differentiating on their own models, versus companies that tune a base model and build a data flywheel; alignment; models that educate themselves rather than freezing at training time; ‘I doubt we’ll be using transformers in 5 years’; and Altman’s favourite optimistic sci-fi universes.

As Kyle captured it

• OpenAI CEO Sam Altman | AI for the Next Era with Sam Altman at OpenAI and Reid Hoffman at Greylock
  • With the language models we have, there will be a serious competitor to Google in search for the first time.
  • Human-level chatbot that actually works; these trends were good, they’re just too early
  • We’ll get multi-modal models in not too much longer
  • Sam Altman: A language interface; you say what you want, and the computer just does it for you. You see this with DALL E 2 and GitHub Co-Pilot
  • You haven’t really seen a platform shift like this since mobile.
  • Skeptical of the startups that are trying to differentiate their own models
  • Instead, the best companies are those that take the base model, tune it, and have a unique data flywheel that improves all of that
  • People are going to be surprised if they think its not going to be generating net new knowledge for humanity; its not going to add to the what people know or cure cancer, and that will turn out to be wrong.
  • Big believer that the only driver of progress is the structure that enables scientific progress, and the scientific progress itself.
  • Alignment research
  • Once the model can get sophisticated enough to understand complex concepts, like racism, you can just tell the model “don’t be racist.”
  • Language models will go much further than people expect
  • True multi-modal models, not just text or images, but fluidly flowing between things
  • Right now models get stuck in time at the point it got trained, and don’t really learn. But we’re going to get models that can educate themselves.
  • “I doubt we’ll be using transformers in 5 years, I think we’ll come up with something much better.”
  • That’s how to make better predictions. Don’t pay attention to the AI for everything. Sometimes you can’t plan the greatness. #Forecasting Trends
  • Best utopian universes so far? Historical Futurism
    • Star Trek: Loves the universes where we turn our focus to exploration and understanding the universe as much as we can
    • The Last Question: incredible short story
    • Iain M. Banks, The Culture
    • There aren’t very many sci-fi universes that you look at and say “all of this is great,” but the corner of sci-fi that is optimistic futurism is great
    • Childhood’s End - from the 60’s; aliens come and take our kids
  • How will users interact with foundation models in the future?
    • “I don’t think we’ll be doing prompt engineering in 5 years.”
    • Either with text or voice, you’ll just interface in language and the computer will do what you want it to