Amanda Askell

@AmandaAskell

Research scientist in ethics & policy at OpenAI. Personal account. All correct views are my own. All incorrect views are me steelmanning someone else.

Vrijeme pridruživanja: srpanj 2016.

Tweetovi

Blokirali ste korisnika/cu @AmandaAskell

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

  1. Prikvačeni tweet
    19. velj 2019.

    In order to align AI systems to human values, we have to resolve a lot of uncertainties about people. This means we need more social science in AI safety. Hopefully of interest to social scientists who would consider working on questions in AI safety.

    Poništi
  2. 22. sij

    If you open with "these billionaires have more money than the bottom 50% of people", it seems like your conclusion should be "more of that money should go to the bottom 50% of people" rather than just "billionaires should pay more in taxes to rich countries".

    Poništi
  3. proslijedio/la je Tweet
    15. sij

    I authored this policy submission with my coworkers at , , and David Lansky, arguing that copying data for training AI systems is protected as fair use under current law, but asking for more clarity on the issue.

    Prikaži ovu nit
    Poništi
  4. 6. sij
    Poništi
  5. 13. pro 2019.

    We've all heard of "damning with faint praise" but "praising with faint damnation" might be more common. It's quite a compliment if your greatest critics can only find minor mistakes with what you've done.

    Poništi
  6. 5. pro 2019.
    Poništi
  7. proslijedio/la je Tweet
    5. pro 2019.

    A surprising deep learning mystery: Contrary to conventional wisdom, performance of unregularized CNNs, ResNets, and transformers is non-monotonic: improves, then gets worse, then improves again with increasing model size, data size, or training time.

    Poništi
  8. 2. pro 2019.

    I think philosophical debates are most useful when empirical methods have been exhausted. If you can work out a way of poking the world into answering your question, that's often easier than trying to work it out from the armchair.

    Poništi
  9. 29. stu 2019.

    I really want to publish an article on this effect in a newspaper, just to mess with people. (Quote is from Michael Crichton's talk, "Why Speculate?")

    Poništi
  10. 27. stu 2019.

    Denying and exaggerating the harms of climate change are both damaging. "Happily, there is a plenty of middle ground between climate apocalypse and climate denial."

    Poništi
  11. proslijedio/la je Tweet
    21. stu 2019.

    We're releasing Safety Gym, environments and tools to evaluate reinforcement learning with safety constraints: Aims to ultimately help agents satisfy real-world safety requirements while training (eg not driving off a cliff, not writing abusive content).

    Poništi
  12. proslijedio/la je Tweet
    18. stu 2019.

    Anonymous answers: What’s the thing people most overrate in their career? Read here: Part of one answer:

    Poništi
  13. 15. stu 2019.

    My initial response to this 2009 article was annoyance. Then I was like "Hang on, why would I be annoyed by someone giving their own money to charity? I guess it makes me feel like I'm being called out. But I'd only feel that way if I thought I should be doing the same thing..."

    Poništi
  14. proslijedio/la je Tweet
    7. stu 2019.

    New analysis (and graph with a mind-bogglingly large log scale) -- quantitative macrohistory of AI! We're uncertain when this new era will end, but society should equip academics and others with the necessary tools (more compute plz) to make breakthroughs in this era.

    Poništi
  15. 4. stu 2019.

    I don't think credit or blame are heritable, which makes me pretty happy to admit to the mediocrity and immorality of my ancestors.

    Poništi
  16. 3. stu 2019.

    It seems like it would be useful to have more red teaming services for individuals. E.g. you can agree to have your credit card company occasionally try common financial scams on you and tell you the results, so that you're more vigilant and aware of potential vulnerabilities.

    Poništi
  17. 1. stu 2019.

    I have much more confidence in media sources if they adhere to basic norms like "cite and link to primary sources and not just to other news outlets". I was reading stories about an Oxfam report from Time, The Guardian, and CNN. Not one of them linked to the report in question.

    Poništi
  18. 30. lis 2019.

    Some people seem to think that any harms dealt out to people who are committing crimes, people with criminal records, or people who've simply done something shitty is totally justified. I find such total disregard for the idea of proportional justice extremely disturbing.

    Poništi
  19. 28. lis 2019.

    Costumes where you don't actually have a costume but try to hide the fact using some clever response ("I'm knot theory, can't you tell from my shoelaces?") are *out*. Costumes that have clearly taken all of your spare time over the course of several weeks to create are *in*.

    Poništi
  20. 28. lis 2019.

    I'm quite enjoying this 80,000 hours interview with Bruce Schneier. Though this part of the interview did make me go "wait, not everyone does this?"

    Poništi
  21. proslijedio/la je Tweet
    15. lis 2019.

    We’re all used to robots that fail when their environment changes unpredictably. Our robotic system is adaptable enough to handle unexpected situations not seen during training, such as being prodded by a stuffed giraffe:

    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:

    ·