Slashdot is powered by your submissions, so send in your scoop

 



Forgot your password?
typodupeerror
×
AI

Nvidia CEO Huang Urges Faster AI Development (goldmansachs.com) 50

At a time when some are calling for a pause on the development of generative AI, Jensen Huang, founder and CEO of NVIDIA, has an argument for accelerating the work: AI advances are going to provide tools to better understand the technology and to make it safer, Huang said in a discussion with Goldman Sachs Asset Management. From a report: "We need to accelerate the development of AI as fast as possible, and the reason for that is because safety requires technology," Huang said in an interview at The Forum with Sung Cho, co-head of Tech Investing for Fundamental Equity in GSAM.

Consider how much safer today's passenger cars are compared with those of earlier generations, Huang suggested, because the technology has advanced. He cited as an example how OpenAI's ChatGPT uses reinforcement learning from human feedback (RLHF) to create guardrails that make its responses more relevant, accurate, and appropriate.

The RLHF is itself an AI model that sits around the core AI model. Huang lists examples of other AI technologies that hold promise for making the models safer and more effective. These range from retrieval augmented generation, in which the model gets information from a defined knowledge base or set of documents, to physics-informed reinforcement learning, which grounds the model in physical principles and constraints. "We need a bunch more technology like that," Huang said.

This discussion has been archived. No new comments can be posted.

Nvidia CEO Huang Urges Faster AI Development

Comments Filter:
  • by fleeped ( 1945926 ) on Friday November 24, 2023 @09:06AM (#64028599)
    News at 11
    • The question is: does he urge faster AI development because he's afraid of an AI capability ceiling that might hurt further investments? "get it while it's hot" sort of situation? I don't think that's going to be the case, but just some food for thought.
      • by Rei ( 128717 ) on Friday November 24, 2023 @10:41AM (#64028845) Homepage

        I find it really telling how people always tend toward assuming somewhere between "the worst" and "outright conspiracies", and never consider aboveboard explanations.

        Don't get me wrong. OF COURSE he has a vested interest in AI training. And OF COURSE one should weigh that in consideration of analyzing his statements. But it's also a fact that most people in the AI space are True Believers(TM).

        Here's Huang seven years ago [youtube.com] predicting a bunch of ways that AI is about to revolutionize industries.

        Here he is in 2017 predicting AI is going to eat software development [technologyreview.com].

        Here he was back in 2019 calling AI the single most powerful force of our time [venturebeat.com]

        NVidia took the lead on being the go-to source for AI training specifically because its leadership saw huge potential in AI and pushed hard to develop both hardware and software to accelerate AI training and inference. Long before AI became a Wall Street darling. While AMD kept thinking of GPUs as that tool to play video games.

        So yes, OBVIOUSLY it's in NVidia's best interest to sell as many cards for AI training and inference as possible. But also, IMHO, Huang most definitely does have strong personal viewpoints on AI as well, and has for quite some time. As far as I can tell, at least since AlexNet started making waves.

        • Not a conspiracy, just typical behaviour. Your forgot his prediction when he was saying cryptocurrency is here to stay [cnbc.com]. Of course his predictions would be anything that's good for business, like every other CEO's "predictions".
          • by Rei ( 128717 )

            There's a difference between talking about something in the middle of it selling you a ton of cards, and only while it's selling you a ton of cards (and sadly, he's right - the shine wore off but there's still an insane amount of resources dedicated to crypto and tons of crypto nuts), versus while the vast majority of the world is ignoring it.

        • by jddimarco ( 1754954 ) on Friday November 24, 2023 @12:47PM (#64029115)
          That's pretty close to the truth. At CS Toronto, we started using Nvidia GPUs for ML in 2009. The reason we picked Nvidia was that they came out with a programming interface for their GPUs (CUDA) that we could use, that worked reasonably well. The reason why they created CUDA was that they wanted to sell expensive "data centre" GPUs (called Tesla at the time, no relation to the car company) for double precision scientific computation. They had the foresight to let CUDA run on their gaming GPUs too. AMD bet instead on OpenCL, which wasn't quite as mature and straightforward to use as CUDA at the time, and I don't think the AMD gaming GPUs were an option at all. Anyway, we bought an Nvidia Tesla and tried CUDA for ML - it worked great: we were seeing up to 30x speedups vs. CPUs. Moreover, we didn't need the double precision, and the gaming cards were a lot cheaper than the Teslas, so instead of a few Teslas, we bought lots of gaming cards. We (Alex Kriz, with lots of help from Ilya) wrote AlexNet for CUDA. The imagenet work was done on Nvidia gaming cards (GTX 580 - I have one of those GPUs in my office for historical purposes, not in use of course). Anyway Nvidia has had a stranglehold on ML ever since. Pytorch and tensorflow run decently on AMD GPUs now, but not better than Nvidia. The irony is that the rise of ML has created huge demand for Nvidia's datacentre cards. They've come up with new datacentre cards that are astoundingly more expensive than the old Tesla cards ever were, and they're selling as many as they can make! You'd think the fact that we made them rich would be worth a couple of free GPUs now and then, but that hasn't happened, we pay as much for Nvidia GPUs as any other University department. But they did give us a free 1st gen DGX1 a few years ago as part of their "pioneers of AI" giveaway, which was nice.
          • Wait you wrote AlexNet? :O

            • Not me personally. I run the computing team at UToronto Computer Science, where AlexNet was built. We just ran the computers and supported the researchers, Alex K. did all the coding (which is why it's called AlexNet), with Ilya's help.
          • by Rei ( 128717 )

            We (Alex Kriz, with lots of help from Ilya) wrote AlexNet

            I wish Slashdot had a "follow" button :)

            • If you're interested in AlexNet and its aftermath, I'm probably not the one you would want to follow. Ilya is probably the best for that: he went on to Google, then co-founded OpenAI, and lead the design of the GPT models. He's been in the news lately. For CS Research computing at Toronto, sure, or random thoughts from me, you can follow me on linkedin (linkedin.com/in/jddimarco) or twitte...I mean X (@jddimarco), or see what I'm saying on slashdot at slashdot.org/~jddimarco But I'm not that interesting. Th
        • The skepticism here may have to do with some rather questionable business practices of Nvidia. Especially with regards to their history of anti-trust practices, shoddy quality control and underperformance for their products, and that hubbub about selling to crypto miners. This has happened within less then five years no less, so it is still fresh on our minds. And of course pumping up previous short-lived tech trends like VR to move products.

          Even if AI really is the next big tech boon, it likely will not ma

        • Nvidia has been selling video cards for AI since at least 2010, so it didn't take a lot of prescience to talk about it in 2017.
          • Nvidia has been selling video cards for AI since at least 2010, so it didn't take a lot of prescience to talk about it in 2017.

            Nvidia has been selling to the HPC market since the introduction of CUDA. The 2000's was filled with hype about big data. AI hype didn't start ramping until AlexNet, and even then, Nvidia was one of the early leaders in seeing the vision of AI (along with with Google and others). The early vision leaders saw that AI wasn't just categorizing cat photos, just like the current vision leaders of generative AI realize that generative AI isn't just ChatGPT.

            • just like the current vision leaders of generative AI realize that generative AI isn't just ChatGPT.

              What vision leader are you talking about? Who thinks that generative AI is just ChatGPT?

    • It's called pickaxes now? I thought it was still called Bitcoin.

      GPU makers sure made a pretty penny with that particular fad too.

  • by stevenm86 ( 780116 ) on Friday November 24, 2023 @09:21AM (#64028623)
    Nothing to do with safety. Dude wants more money, in the wake of setbacks with respect to autonomous cars. Don't need an LLM to get to the bottom of that one. In other news, the sky is blue and water is wet.
    • Re:Please. (Score:4, Informative)

      by Rei ( 128717 ) on Friday November 24, 2023 @09:37AM (#64028671) Homepage

      In other news, the sky is blue and water is wet.

      The color of the sky is contingent upon factors such as the time of day, weather conditions, and geographic location. During sunrise or sunset, for instance, the sky often exhibits hues of red, orange, or pink, challenging the absolute assertion of its blueness. Additionally, in overcast or stormy weather, the sky may appear gray or even a multitude of colors, further complicating the categorical declaration of a consistently blue sky.

      Wetness typically implies a condition of being exposed to moisture resulting from the application of a liquid to a dry surface. However, water (in the liquid state normally implied), by virtue of being a liquid can not correspondingly be "wet" . From a purely semantic standpoint, applying the descriptor "wet" to water may be considered ill-defined since becoming wet, in its conventional understanding, involves a transition of an object from dryness to dampness, a transformation that water inherently lacks due to its liquid nature; water is the thing that wets, not which becomes wet, and the other thing is that which is wet, not the water. One might critique the premise of wetness being ill-defined without a liquid making contact with a dry surface, but the only counterexample readily within reach in common parlance is the phrase "water is wet", thus rendering it a meaningless truism.

      • by vyvepe ( 809573 )
        Nice one :)
      • by Anonymous Coward

        downvoted to conceal the dangerous truths you spread

      • Pwnz0red / Edumacated.

      • by GrahamJ ( 241784 )

        Wetness isn't a transition, it's a state. Something is wet when it has liquid in or on it. Water has liquid in and on it.

        If water isn't wet then it can only become wet when it comes in contact with something dry. Surely dry things do not make water wet.

        • by Rei ( 128717 )

          " water is the thing that wets, not which becomes wet, and the other thing is that which is wet, not the water."

        • by Rei ( 128717 )

          A surface that in its normal state is not wet is absolutely essential to the definition of wet - because otherwise, what would the adjective "wet" even be describing with relation to the object? "Wet cloth" implies the existence of "dry cloth". There's on such thing as dry water.

          Wet isn't an adjective that means "having the properties of a liquid" (that word is "fluid") - it means "being in contact with a liquid". There's a solid dividing line between describing objects' properties within themselves and

          • by GrahamJ ( 241784 )

            "A surface that in its normal state is not wet is absolutely essential to the definition of wet"

            Wet is a binary state so of course the opposite has to exist. That’s like saying that cold must exist because hot does; it’s only true because they’re semantic opposites. Heat and wet are physical properties whereas dry and cold are merely the lack of those properties.

            As such, dry cannot exist without liquids but wet can exist without non-liquid states.

            Obviously in our universe it’s imposs

      • People say "the sky is blue and water is wet" already know all of that. It's only "meaningless" in the sense that:

        a. you're intentionally using trivial / semantic details that everyone already understands to avoid the original point and
        b. (a) is the point being made by using that phrase in the first place
        c. effective communication requires that people make a good faith effort not to do (a)

        With regard to the OP - in context, the phrase means roughly all of: "I already know that the NVIDIA team has addition

  • by Anonymous Coward
    The first strong AI that is developed (one which can bootstrap itself to ever-greater capabilities while being fed hardware resources) will be able to prevent any other AI from being similarly developed, unless it is done in air-gapped secret. While not perfect, I trust the ethics approach in Western cultures far more than the CCP's values.
  • by Growlley ( 6732614 ) on Friday November 24, 2023 @09:45AM (#64028701)
    wah my bonuses and profits!
  • They want to sell more hardware.
  • "NVIDIA delays launch of China focused AI chip" https://www.reuters.com/techno... [reuters.com]

  • by backslashdot ( 95548 ) on Friday November 24, 2023 @10:25AM (#64028815)

    They failed badly in autonomous driving. They are at least 10 years behind Tesla, which is really bad because Tesla's autonomous driving capability isn't even 10 years old.

    • by Rei ( 128717 )

      Well... yes and no.

      In terms of self-driving programs... NVidia never had one. They've always just provided hardware for others.

      In terms of inference, Tesla used to run on AMD GPUs. They switched to their own internally-developed system optimized solely for inference of their models, because it let them get better inference performance for their dollar, and in particular a lot lower power consumption. But many others still run on NVidia hardware

      However, in the datacentre, for training, even Tesla is still h

      • In terms of self-driving programs... NVidia never had one. They've always just provided hardware for others.

        Really? Then what's this? https://www.nvidia.com/en-us/s... [nvidia.com]

        • That's a partnering program. Nvidia helps you to get into the space, by providing an API and other resources.
          So it's "sort of" a self driving program, but it's only half of one. It's the beginnings of one.
          They also have APIs for lots of car-related stuff, so they are obviously doing lots of work in the automobile software space, but it's all at the API level.

        • by Rei ( 128717 )

          That's a self-driving program in the same way that CUDA + PyTorch + Transformers is a LLM.

      • by Rei ( 128717 )

        ** Still heavily dependent on NVidia

    • by Ichijo ( 607641 )

      Nor is it [wikipedia.org] fully autonomous (SAE Level 5), despite the promise [theverge.com] made 7 years ago in 2016. It's always just a year away.

      • Nor is it [wikipedia.org] fully autonomous (SAE Level 5), despite the promise [theverge.com] made 7 years ago in 2016. It's always just a year away.

        Level 5 is a dream for Tesla (and for everyone else, too) right now. Currently, Tesla is still struggling to achieve Level 3 in a way that the lawyers will allow them to market as Level 3.

    • by m00sh ( 2538182 )

      They failed badly in autonomous driving. They are at least 10 years behind Tesla, which is really bad because Tesla's autonomous driving capability isn't even 10 years old.

      They didn't fail at autonomous driving.

      Autonomous driving itself failed.

      If autonomous driving was successful, nVidia would be there with all the tools and products.

    • They failed badly in autonomous driving. They are at least 10 years behind Tesla, which is really bad because Tesla's autonomous driving capability isn't even 10 years old.

      In terms of AV software, Tesla is not even close to being the gold standard. Waymo started earlier and its software is more capable. There are others that are also better than Tesla.

      Tesla is also unique in being vertically integrated. They sell the compute hardware, the sensor rigs, the software, and the entire car with everything integrated. Nvidia sells mostly the compute hardware, and although it develops software, its current and likely future sales focus is on compute hardware. It's not clear if T

  • by xack ( 5304745 ) on Friday November 24, 2023 @11:03AM (#64028895)
    Starting with power connectors that don't melt. And shrink the card sizes too, there are already memes of 5090s being the size of window units.
  • by m00sh ( 2538182 ) on Friday November 24, 2023 @11:59AM (#64029025)

    You want faster AI development, lower prices and enable everyone to have more RAM and compute.

    Instead of crippling cheap GPUs with low RAM that can't run AI as much as possible!

  • Ok yes I do own a few shares but no we don't need a big AI push.

    I bought N stock because other people think AI is "gOINg to ChAnGE EVrytHInG!" so their stock was going up.

    I'm still waiting for real AI and not this pattern matching Tom foolery.

  • What "Urges faster development..." really means.

  • ...He won't get my money, but I agree with the viewpoint he presents.

Avoid strange women and temporary variables.

Working...