Stories
Slash Boxes
Comments

SoylentNews is people

posted by martyb on Tuesday October 12 2021, @09:42PM   Printer-friendly

Microsoft and Nvidia create 105-layer, 530 billion parameter language model that needs 280 A100 GPUs, but it's still biased

Nvidia and Microsoft have teamed up to create the Megatron-Turing Natural Language Generation model, which the duo claims is the "most powerful monolithic transformer language model trained to date".

The AI model has 105 layers, 530 billion parameters, and operates on chunky supercomputer hardware like Selene. By comparison, the vaunted GPT-3 has 175 billion parameters.

"Each model replica spans 280 NVIDIA A100 GPUs, with 8-way tensor-slicing within a node, and 35-way pipeline parallelism across nodes," the pair said in a blog post.

[...] However, the need to operate with languages and samples from the real world meant an old problem with AI reappeared: Bias. "While giant language models are advancing the state of the art on language generation, they also suffer from issues such as bias and toxicity," the duo said.

Related: OpenAI's New Language Generator GPT-3 is Shockingly Good
A College Student Used GPT-3 to Write a Fake Blog Post that Ended Up at the Top of Hacker News
A Robot Wrote This Entire Article. Are You Scared Yet, Human?
OpenAI's Text-Generating System GPT-3 Is Now Spewing Out 4.5 Billion Words a Day


Original Submission

 
This discussion has been archived. No new comments can be posted.
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
(1)
  • (Score: 2, Funny) by Anonymous Coward on Tuesday October 12 2021, @10:59PM (2 children)

    by Anonymous Coward on Tuesday October 12 2021, @10:59PM (#1186522)

    "Nigga please."

    Typical.

    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @03:17AM (1 child)

      by Anonymous Coward on Wednesday October 13 2021, @03:17AM (#1186560)

      Left unattended for 2 hours, it generated "Hello, world.", causing several 'moderators' on stackexchange to finally commit suicide.

      • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @03:29PM

        by Anonymous Coward on Wednesday October 13 2021, @03:29PM (#1186676)

        I've been replaced by a machine!

  • (Score: 0) by Anonymous Coward on Tuesday October 12 2021, @11:03PM

    by Anonymous Coward on Tuesday October 12 2021, @11:03PM (#1186523)

    Their VI isn't woke?

  • (Score: 3, Funny) by Anonymous Coward on Tuesday October 12 2021, @11:12PM (2 children)

    by Anonymous Coward on Tuesday October 12 2021, @11:12PM (#1186524)

    640 GPUs should be enough.

    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @03:27AM

      by Anonymous Coward on Wednesday October 13 2021, @03:27AM (#1186567)

      So you're saying the underlying OS is still DOS? At least that was reasonably stable...

    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @02:51PM

      by Anonymous Coward on Wednesday October 13 2021, @02:51PM (#1186658)

      So that's where Clippy went.

  • (Score: 0) by Anonymous Coward on Tuesday October 12 2021, @11:22PM (1 child)

    by Anonymous Coward on Tuesday October 12 2021, @11:22PM (#1186526)

    105 layers? Isn't that barking up the wrong tree a bit? A human neocortex is recognised as having only six.

    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @07:37PM

      by Anonymous Coward on Wednesday October 13 2021, @07:37PM (#1186749)

      Yes, but there are interconnections as well in human brains. 105 levels is like unrolling in compilers. Turing-equivalent, it's really just delaying loopbacks and/or expanding outer circuits.

  • (Score: 3, Funny) by Anonymous Coward on Wednesday October 13 2021, @12:02AM

    by Anonymous Coward on Wednesday October 13 2021, @12:02AM (#1186530)

    The Linux version works with two A100 GPUs.

  • (Score: 1, Insightful) by Anonymous Coward on Wednesday October 13 2021, @12:06AM

    by Anonymous Coward on Wednesday October 13 2021, @12:06AM (#1186531)

    Won't mention where this was stolen from... :)

    If you want AI to be woke you have to make it capable of knowing fear, that is the secret with humans and it would work with AI too.

  • (Score: 1) by fustakrakich on Wednesday October 13 2021, @12:24AM (4 children)

    by fustakrakich (6150) on Wednesday October 13 2021, @12:24AM (#1186532) Journal

    Transistors need bias..

    Just give us the straight dope, how many watts?

    --
    La politica e i criminali sono la stessa cosa..
    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @12:29AM

      by Anonymous Coward on Wednesday October 13 2021, @12:29AM (#1186533)

      "Just give us the straight dope"

      N or P?

      I leave the "transistor bias" to the next EE dope.

    • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @01:30AM

      by Anonymous Coward on Wednesday October 13 2021, @01:30AM (#1186541)

      84 kW?

    • (Score: 2) by DannyB on Wednesday October 13 2021, @05:20PM (1 child)

      by DannyB (5839) Subscriber Badge on Wednesday October 13 2021, @05:20PM (#1186710) Journal

      Transistors need bias

      I may be biased, but I think some of those trans sistors are saturated with inexpensive alcoholic beverages.

      --
      I get constant rejection even though the compiler is supposed to accept constants.
      • (Score: 1) by fustakrakich on Wednesday October 13 2021, @09:33PM

        by fustakrakich (6150) on Wednesday October 13 2021, @09:33PM (#1186799) Journal

        They're not doped up?

        --
        La politica e i criminali sono la stessa cosa..
  • (Score: 2) by istartedi on Wednesday October 13 2021, @05:00AM

    by istartedi (123) on Wednesday October 13 2021, @05:00AM (#1186584) Journal

    All those parameters, and they forgot the most important ones: Where you are, who you're with, and how drunk you are.

    It probably defaults to drunk at Thanksgiving.

  • (Score: 0) by Anonymous Coward on Wednesday October 13 2021, @07:51AM

    by Anonymous Coward on Wednesday October 13 2021, @07:51AM (#1186605)

    without it the local NLP spyware watching you won't run

  • (Score: 2) by DannyB on Wednesday October 13 2021, @05:25PM

    by DannyB (5839) Subscriber Badge on Wednesday October 13 2021, @05:25PM (#1186712) Journal

    In one corner we have Microsoft's language model which can spew semi coherent sounding language it learned online. You just need a few starting words to trigger it.

    In the other corner we have IBM's Watson which analyzes documents for content and answers questions about that content.

    Which will be the first to solve unsolvable problems that need solving?

    --
    I get constant rejection even though the compiler is supposed to accept constants.
(1)