Adult Content Warning

This community may contain adult content that is not suitable for minors. By closing this dialog box or continuing to navigate this site, you certify that you are 18 years of age and consent to view adult content.

Elephants and Jackasses...

Discussion in 'Permanent Threads' started by Nettdata, Oct 14, 2016.

  1. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    Not much, if any. It's a totally different approach to AI that none of my sources or references have seen before, which is partly why it's making such a huge wave in the field.

    I know they steal a shit-ton of IP, but they seem to be the originators of this stuff.

    I hear nobody yelling "thief"... at least not yet.
     
  2. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    To be clear, I'm not an AI academic in big model learning stuff like this... I specialize in the global infrastructure and the development of something called "agentic workflows" as it relates to our specific use cases.

    Our yearly cloud bill, just for the AI part, is $350m.

    If this new shit requires way, way less cloud infrastructure, or can be run on the client device directly, it's a massive game changer, as the hyperscalers (Azure, Google, AWS, etc) aren't fighting over all the Nvidia GPU's that can be made any more.

    We rent thousands of GPU's that would be about $25k each to buy outright in order to do our generative AI. If that could be replaced with a shit-ton of non-GPU CPU's for pennies on the dollar, that are way more available, that would be awesome.
     
  3. downndirty

    downndirty
    Expand Collapse
    Emotionally Jaded

    Reputation:
    501
    Joined:
    Nov 18, 2009
    Messages:
    4,606
    We are a few years past the point of Chinese theft being an issue, and more in the realm of Chinese tech superiority, they just see no need to advertise it or sell it to the West. Just look at their electric cars.

    Put another way, the AI bubble was going to burst sooner or later, but bubbles emerge and die based on a bunch of different factors

    The US economy overheating over the past few years, the tech sector saturation and business model maturity, the incoming administration hostile to China, and given the current regime, no one is lining up to buy from tech companies that watched Elon Musk give a nazi salute...all adds up to a good time to get your money out of American AI specifically and US tech in general.

    This is calculated in terms of timing, but a real enough threat to knock a point or two off the whole damned US stock market. I'd bet as more folks find out that it's legit (or at least, not bullshit), that tanking will continue.

    The hubris that the Chinese could actually deliver a better mousetrap without cheating or theft is staggering, though.
     
  4. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    Trump: "We will invest $500b in AI infrastructure!"

    China: "Hold my rice beer..."

    Narrator: "And they no longer needed AI infrastructure."
     
  5. Aetius

    Aetius
    Expand Collapse
    Emotionally Jaded

    Reputation:
    839
    Joined:
    Oct 19, 2009
    Messages:
    9,103
    I know we have a habit of just throwing hardware at a problem until it goes away, but we really sent a company that used to be best known for making video game cards to an ATH valuation of $3.5 trillion instead of optimizing our AI code.
     
  6. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    China: "Mission accomplished."

    [​IMG]
     
  7. doomrider7

    doomrider7
    Expand Collapse
    Experienced Idiot

    Reputation:
    9
    Joined:
    Jun 8, 2011
    Messages:
    176
    Yes and no. It's cheaper, but it's very much in its infancy compared to ChatGPT and whatnots. Granted it's probably an early build.
     
  8. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    It’s outperforming ChatGTP and OpenAI in a ton of tests.

    It is good.
     
  9. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
  10. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    We use a wide variety of models, of various sizes (depending on their role in the process), from Meta's Llama that we self-host, to OpenAI's remote service, and this new DeepSeek is performing very well in our early testing.

    It really does appear to be a major game changer at this time, and not to be discounted just because it's Chinese and Open Source.
     
  11. GTE

    GTE
    Expand Collapse
    Emotionally Jaded

    Reputation:
    632
    Joined:
    Oct 20, 2009
    Messages:
    3,175

    I'm assuming your company is ok with the data being stored on Chinese servers?
     
  12. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    It's not quite that simple.

    If you use their online or default app, then yeah, data is sent to them.

    If you host their models in your own infrastructure, the data doesn't get shipped back to the Homeland.

    It depends on how you implement DeepSeek.

    We don't do it so that shit gets shipped back to the Homeland.
     
  13. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    Again... their entire model is completely open sourced, so we can use it how we want.
     
  14. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    The real value here is in the method that they train their models. They’ve made that open for the world to use.

    Meta, with their Llama models, have not. They just give you access to a pre-trained model to use.

    OpenAI only gives you an API to hit for the model inferencing, while charging you a shit ton of cash for the privilege.
     
  15. doomrider7

    doomrider7
    Expand Collapse
    Experienced Idiot

    Reputation:
    9
    Joined:
    Jun 8, 2011
    Messages:
    176
    Oh damn. The info I got from some techies was that it was behind in some or most. Well lol then.
     
  16. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    Go look at the evaluation results on this page:

    https://github.com/deepseek-ai/DeepSeek-R1

    These are standardized tests that any ML engineer can run against any model.

    We constantly run them against the models that we train with our own data for our specific use cases.

    They are UP THERE. Not winning in every category, but man, they are fucking good.

    We are currently evaluating whether or not they are good enough to use for what we need them to do in our case, and man, they are scoring well... and their price is damn compelling. (they're free).
     
  17. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
     
  18. AFHokie

    AFHokie
    Expand Collapse
    Emotionally Jaded

    Reputation:
    313
    Joined:
    Apr 13, 2010
    Messages:
    1,622
    Location:
    Manassas, VA
    I'm not a cyber guy so this is probably a dumb question, but are you concerned they're using you hosting their models on your infrastructure as a means to gain access to your systems and databases?
     
  19. Nettdata

    Nettdata
    Expand Collapse
    Mr. Toast

    Reputation:
    3,004
    Joined:
    Feb 14, 2006
    Messages:
    26,748
    Not at all. We completely control everything around access to those models.
     
  20. Kubla Kahn

    Kubla Kahn
    Expand Collapse
    Did I just shit myself?

    Reputation:
    730
    Joined:
    Oct 19, 2009
    Messages:
    11,554
    If it’s open source can’t chatgtp and such just copy it and have their own version?

    So does this mean we don’t need to be firing up old nuke plants just to power ai computations? Is this real life middle out processing?