Guillermo Valle

@guillefix

I {learn things and make things that}*

Oxford, England
Vrijeme pridruživanja: prosinac 2009.

Tweetovi

Blokirali ste korisnika/cu @guillefix

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

  1. Prikvačeni tweet
    5. svi 2019.

    My mobile-friendly version of the poster now allows arbitrary levels of *ackshually*s

    Poništi
  2. prije 11 sati

    I don't know what kind of *punk this is, but I like it

    Poništi
  3. prije 14 sati

    Although "direct mainpulation" interfaces are more intuitive, it makes sense that the optimal interface separates the input and output because our sensory/actuator organs are separate and very different.

    Poništi
  4. proslijedio/la je Tweet
    29. sij

    Averaging two models results in some beautiful combinations! It's fast too! It feels like but with entire models, instead of just the parameters and latent vectors of a single model. More samples on my site:

    Prikaži ovu nit
    Poništi
  5. 2. velj

    hmmm so you better not be late with your engineered cat girl genes, seems like it's our only hope.

    Poništi
  6. 2. velj

    this was a great talk on the topic of AI and games, with an overview of the field and open opportunities!

    Poništi
  7. 1. velj

    Is there *no* way to order by year in Google Scholar on mobile?

    Poništi
  8. 1. velj

    This is basically like the idea I've been thinking about of how children use toys to more eaisly learn language models. Need to see how they implemented it

    Poništi
  9. 1. velj

    Alternatively could have something like AIdungeon but where people can choose to branch off from any point in the stories people generate publicly on the site, and have an ever growing tree of parallel universes/alternative stories, with humans exploring the most interesting ones

    Prikaži ovu nit
    Poništi
  10. 1. velj

    Has anyone thought of an artbreeder for GPT2? if there was a way to express the generated text in a latent vector.. I guess we need to wait till good GAN-based language models.

    Prikaži ovu nit
    Poništi
  11. 1. velj

    hmm maybe the safest bet to achieve immortality today would be to spend all day in VR with complete full body tracking, and save *all* your data in a safe hard drive, waiting for the day an ML model fine-tuned on your data will be indistinguishable from you.

    Poništi
  12. 31. sij
    Poništi
  13. 31. sij

    I want this as an avatar in Neos; if i knew more 3d modelling..

    Poništi
  14. 31. sij

    Anyone know why most website, e.g. Google are designed to hinder scrapping? Is it to avoid overloading the servers, for business reasons (data is valuable), something else?

    Poništi
  15. 30. sij

    *rejoices in market* with this I can spawn an instance with 8x1080Ti for less than 1/20 of the price of a 8xV100 in google cloud, and I can earn money with my own gpus too

    Poništi
  16. 30. sij

    Toys are literally modelling tools, that's the idea

    Prikaži ovu nit
    Poništi
  17. 30. sij

    In short. Learn world models with toys via RL say u want to learn world model of social interaction, it's much easier to do by playing with dolls, than all in your head, as the toys already encode lots of useful knowledge: object and appearance permanence, physics, etc.

    Prikaži ovu nit
    Poništi
  18. 30. sij

    Yes! finally a chatbot with GPT2 tech. check Connor's video or the blog for nice explanations Is there code or some demo I can test already?

    Poništi
  19. 29. sij

    I can imagine scenarios where the scooping is overall negative, but they seem generally contrived

    Prikaži ovu nit
    Poništi
  20. 29. sij

    can someone convince why I should or shouldn't share every research idea I have on twitter? like if someone scoops me, and does it faster, that's just advancing science faster right?

    Prikaži ovu nit
    Poništi
  21. 29. sij

    hmm I've been self-brainstorming on why children play how they do and have looped around ideas on data augmentation by imagination, learning equivariances, and world models, goal policies by RL, and other stuff, I think I have converged into an idea which sounds actually good.

    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:

    ·