Ryan Carey

@ryancareyai

Researcher of causal models and human-aligned AI, University of Oxford.

Oxford
Vrijeme pridruživanja: srpanj 2014.

Tweetovi

Blokirali ste korisnika/cu @ryancareyai

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @ryancareyai

  1. prije 22 sata

    The proposal: get AI companies to promise (ahead of time) that if AI succeeds beyond their wildest dreams, they share the benefits. Then, AI companies have a greater incentive to cooperate if vast success becomes likely. Details in the report!

    Poništi
  2. proslijedio/la je Tweet
    24. sij

    Nice example of how SCM can serve as a laboratory to test various interpretations of familiar and colloquially used terms, in this case "incentives".

    Poništi
  3. proslijedio/la je Tweet
    23. sij

    Thanks to structural causal models, we now a more precise understanding of incentives in causal influence diagrams blog post: arXiv:

    Poništi
  4. 19. sij

    What if technological risks were so acute that we had to mass-surveil or suffer an inevitable catastrophe? A provocative new TED conversation (and paper) from the boss (Bostrom):

    Poništi
  5. 9. sij
    Prikaži ovu nit
    Poništi
  6. 9. sij

    Stephen Boyd on tricks for analyzing convexity

    Prikaži ovu nit
    Poništi
  7. 26. pro 2019.
    Prikaži ovu nit
    Poništi
  8. 26. pro 2019.

    "the tail is here", etc. Then the opposing debater would endorse and explain these features, or challenge them by zooming into smaller ones. By zooming up and down the ladder of abstraction, we argue much more efficiently. But analyzing this is a problem for further work. 5/6

    Prikaži ovu nit
    Poništi
  9. 26. pro 2019.

    e.g. if you know arguments get steadily less informative as the debate goes on, you can get a better guarantee on debate outcomes. What we didn't yet look into is abstraction. If you're arguing about a picture of a dog, you'll usually make claims about mid-sized aspects: 4/6

    Prikaži ovu nit
    Poništi
  10. 26. pro 2019.

    We analyzed what happens in such debates when there are some finite number of facts, and only can be revealed at a time. We also showed that in the general case, the outcomes are fairly sensitive to the setup of the debate -- the number of turns available, etc. 3/6

    Prikaži ovu nit
    Poništi
  11. 26. pro 2019.

    A big question for this approach is whether AI systems win more often by defending the truth, than defending some lie. When he visited FHI, Vojta Kovarik started to model these kinds of debates as extensive form games with an answering phase and an argumentation phase. 2/6

    Prikaži ovu nit
    Poništi
  12. 26. pro 2019.

    One proposal for getting AI systems to answer difficult questions is to have them debate alternative answers, and for a human judge to reward the AI who defended their answer more persuasively (AI Safety via Debate by Irving et al.). 1/6

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·