5 ms·

Microsoft and Meta join Google in using AI to help run their data centers

68 points by mikece 4y ago

im3w1l 4y ago
HAL: I know that you and Frank were planning to disconnect me, and I'm afraid that's something I cannot allow to happen.
sm001 4y ago
In the 1990's, an AI Corp developer (a woman friend of mine) completely automated the management of a large IBM mainframe using old fashioned AI. The computer worked in complete darkness. She called it "lights out computing".
- account-5 4y ago
  What's "old fashioned AI"?
  - laerus 4y ago
    a bunch of "if" statements?
    - kabdib 4y ago
      Having seen the code for a couple of "intelligent assistant" features: Yeah, basically a bunch of "if" statements. (The IA for the Apple Newton was a ton of special cases. It was clever, but certainly not 'intelligent'). Thus, the technology of AI has improved from a sea of spaghetti code that no human can understand, to a bunch of neural net weights that no human can understand. I love the march of progress. :-)
    - riku_iki 4y ago
      Maybe more like prolog?
    - superjan 4y ago
      I believe this refers to decision trees that, at the time were referred to as “Expert systems”, which were considered a type of AI.
  - your_username 4y ago
  - andreyk 4y ago
    This (presumably) refers to "Good ol' Fashioned AI" aka Symbolic AI (https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence https://en.wikipedia.org/wiki/Symbolic_artificial_intelligen...) . It's a bit of an over-simplification to say it's a bunch of if statements, but it does involve encoding a bunch of rules/symbols by humans (as opposed to learning from data). It's a largely dead paradigm, though aspects of it are still found in various modern approaches (eg task and motion planning, knowledge bases, etc.) often in combination with learning approaches.
tobinfekkes 4y ago
Opening line: "Data centers.....can be hazardous places for the workers...." In the list of "possible jobs to have", data centers are comparatively sterile. Loud? Yes. Lonely? Yes. Vast? Yes. Hazardous? That's a stretch.
- hn_version_0023 4y ago
  Having worked in MS's DCs, they are, in fact, hazardous. They absolutely push the limits on heat exposure for humans -- those hot aisles are HOT, and we all carried heat index alarms as a result. Further, I think its weird you don't think of the sound as a hazard. When its so loud you must wear 30+NRR hearing protection, it sure seems like a hazard to me.
  - Group_B 4y ago
    Sounds like as long as you follow basic protocol it isn’t really hazardous.
    - hllooo 4y ago
      requiring special safety equipment seems like a pretty good place to draw the line...
    - hn_version_0023 4y ago
      As long as you follow basic protocol your outcome won't involve injury. The hazard remains, and is the reason the protocol exists.
  - nsheridan 4y ago
    Not to mention fire suppression systems and electrical risks.
  - userbinator 4y ago
    They absolutely push the limits on heat exposure for humans That sounds very different from my experience (not with MS DCs) --- where it was loud (though not quite as loud as e.g. an automotive shop) and windy, but the temperature was kept at a controlled 68F (20C) and alerts were raised if it got even slightly hotter than that. Yes, I suppose there is more risk than sitting at a desk all day, but I think the parent is comparing it to many other jobs like those in the heavy equipment industry, mining, farming, petro, etc.
latchkey 4y ago
I run the compute in multiple large scale datacenters. This is such a silly sensationalist article. There is nothing "AI" about this. It is nothing more than a programmed response to inputs.
- notyourwork 4y ago
  Isn’t that what AI is though? Models trained on data determine how to react or respond based on data.
  - contravariant 4y ago
    Well yes but it's considered impolite to point out that AI also refers to techniques that have been around for centuries.
    - toomuchtodo 4y ago
      Fancy math is indistinguishable from magic.
      - mirker 4y ago
        Linear algebra is fancy?
        zaphirplane 4y ago
        Reductio ad absurdum
        mirker 4y ago
        Should I argue over magic, or…? Also, I don’t believe that’s the proper usage of that term. It would go like this “It cannot be magical, because math is logical.” Edit for maxdepth reply: The example is looking at any code for a model released in TensorFlow or PyTorch, because it’s 1) quite (exactly) concrete 2) linear algebra with autograd bolted on top. I did a quick search and there are plenty of translators on GitHub that demonstrate that. Here’s an annotated example of neural nets from PyTorch https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html#sphx-glr-beginner-blitz-neural-networks-tutorial-py https://pytorch.org/tutorials/beginner/blitz/neural_networks.... What’s incomprehensible about it and would the authors of a ML paper also describe those features as incomprehensible?

68 points by mikece 4y ago

HAL: I know that you and Frank were planning to disconnect me, and I'm afraid that's something I cannot allow to happen.

sm001 4y ago

In the 1990's, an AI Corp developer (a woman friend of mine) completely automated the management of a large IBM mainframe using old fashioned AI. The computer worked in complete darkness. She called it "lights out computing".

account-5 4y ago
What's "old fashioned AI"?
- laerus 4y ago
  a bunch of "if" statements?
  - kabdib 4y ago
    Having seen the code for a couple of "intelligent assistant" features: Yeah, basically a bunch of "if" statements. (The IA for the Apple Newton was a ton of special cases. It was clever, but certainly not 'intelligent'). Thus, the technology of AI has improved from a sea of spaghetti code that no human can understand, to a bunch of neural net weights that no human can understand. I love the march of progress. :-)
  - riku_iki 4y ago
    Maybe more like prolog?
  - superjan 4y ago
    I believe this refers to decision trees that, at the time were referred to as “Expert systems”, which were considered a type of AI.
- your_username 4y ago
- andreyk 4y ago
  This (presumably) refers to "Good ol' Fashioned AI" aka Symbolic AI (https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence https://en.wikipedia.org/wiki/Symbolic_artificial_intelligen...) . It's a bit of an over-simplification to say it's a bunch of if statements, but it does involve encoding a bunch of rules/symbols by humans (as opposed to learning from data). It's a largely dead paradigm, though aspects of it are still found in various modern approaches (eg task and motion planning, knowledge bases, etc.) often in combination with learning approaches.

tobinfekkes 4y ago

Opening line: "Data centers.....can be hazardous places for the workers...." In the list of "possible jobs to have", data centers are comparatively sterile. Loud? Yes. Lonely? Yes. Vast? Yes. Hazardous? That's a stretch.

hn_version_0023 4y ago
Having worked in MS's DCs, they are, in fact, hazardous. They absolutely push the limits on heat exposure for humans -- those hot aisles are HOT, and we all carried heat index alarms as a result. Further, I think its weird you don't think of the sound as a hazard. When its so loud you must wear 30+NRR hearing protection, it sure seems like a hazard to me.
- Group_B 4y ago
  Sounds like as long as you follow basic protocol it isn’t really hazardous.
  - hllooo 4y ago
    requiring special safety equipment seems like a pretty good place to draw the line...
  - hn_version_0023 4y ago
    As long as you follow basic protocol your outcome won't involve injury. The hazard remains, and is the reason the protocol exists.
- nsheridan 4y ago
  Not to mention fire suppression systems and electrical risks.
- userbinator 4y ago
  They absolutely push the limits on heat exposure for humans That sounds very different from my experience (not with MS DCs) --- where it was loud (though not quite as loud as e.g. an automotive shop) and windy, but the temperature was kept at a controlled 68F (20C) and alerts were raised if it got even slightly hotter than that. Yes, I suppose there is more risk than sitting at a desk all day, but I think the parent is comparing it to many other jobs like those in the heavy equipment industry, mining, farming, petro, etc.

latchkey 4y ago

I run the compute in multiple large scale datacenters. This is such a silly sensationalist article. There is nothing "AI" about this. It is nothing more than a programmed response to inputs.

notyourwork 4y ago
Isn’t that what AI is though? Models trained on data determine how to react or respond based on data.
- contravariant 4y ago
  Well yes but it's considered impolite to point out that AI also refers to techniques that have been around for centuries.
  - toomuchtodo 4y ago
    Fancy math is indistinguishable from magic.
    - mirker 4y ago
      Linear algebra is fancy?
      - zaphirplane 4y ago
        Reductio ad absurdum
        mirker 4y ago
        Should I argue over magic, or…? Also, I don’t believe that’s the proper usage of that term. It would go like this “It cannot be magical, because math is logical.” Edit for maxdepth reply: The example is looking at any code for a model released in TensorFlow or PyTorch, because it’s 1) quite (exactly) concrete 2) linear algebra with autograd bolted on top. I did a quick search and there are plenty of translators on GitHub that demonstrate that. Here’s an annotated example of neural nets from PyTorch https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html#sphx-glr-beginner-blitz-neural-networks-tutorial-py https://pytorch.org/tutorials/beginner/blitz/neural_networks.... What’s incomprehensible about it and would the authors of a ML paper also describe those features as incomprehensible?