Artificial Intelligence - Page 41

Microsoft has publicly come out and denounced AI-generation tools being used to create deepfake images that are then used to commit crimes such as fraud, abuse, manipulation.

Unfortunately, the demographics that are victims of this form of abuse are children and the elderly, and according to a recent blog post by Microsoft Vice Chair and President Brad Smith, the US government needs to step in and implement new regulations that hold the creators of deepfake content with nefarious purposes accountable for their actions.

Smith explains that AI-generated deepfakes are realistic and extremely easy for anyone to make. Unfortunately, due to their accessibility, the technology, while being built with the intention to conduct research and assist in people's workflows/projects, is increasingly being used to commit fraud, abuse, and other crimes. Smith not only called up regulators for new laws to protect victims of AI deepfakes but also the private sector to acknowledge its responsibility to "prevent the misuse of AI."

Samsung has been hitting problem after problem with developing new HBM memory chips for the ever-expanding HBM market, where it still needs another 2-4 months before the South Korean giant will get approved by NVIDIA to use its new HBM3E memory on its AI GPUs.

In a new Bloomberg report, Samsung has made "important headway in its comeback, including winning the long-awaited approval" from NVIDIA for its HBM3 memory to be used on NVIDIA's leading AI GPUs.

Samsung is now anticipating approval for its next-gen HBM3 memory in the next 2-4 months, according to people familiar with the matter who "asked not to be identified discussing internal developments".

Meta's long-term vision for AGI (artificial general intelligence) involved the use of 600,000 x NVIDIA H100 AI GPUs, something NVIDIA CEO Jensen Huang teased Meta CEO Mark Zuckerberg at SIGGRAPH 2024 this week Check it out:

In the video, NVIDIA CEO Jensen Huang says that Meta is "coming up on 100K H100s" to which Zuck replies that yeah, they're "good customers" for NVIDIA and "that's why you invited me to this Q&A", said Zuck in return to Jensen.

600,000 x NVIDIA H100 AI GPUs at an average cost of $30,000 each combines for a total of $15 billion in AI GPU purchases from Meta to NVIDIA alone. We've seen gigantic GPU clusters from Elon Musk and his xAI startup, with its new Memphis Supercluster needing some insane portable power generators just to get 32,000 x NVIDIA H100 AI GPUs operational in the cluster.

NVIDIA CEO Jensen Huang has confirmed that engineering samples of its next-gen Blackwell AI GPUs will be sent out "all over the world".

The news comes from NVIDIA CEO Jensen Huang himself during a chat at SIGGRAPH 2024 this week, with Jensen saying: "This week, we are sending out engineering samples of Blackwell all over the world".

SIGGRAPH is more tailored towards the "software side of things" so NVIDIA didn't go into the nitty-gritty of its AI hardware, but it did tease that the engineering samples of Blackwell will be headed out this week. That's a very good sign, as it shows that Blackwell AI GPUs are very, very close to being in customers' hands.

SK hynix has announced that its new fifth-generation HBM3E memory is expected to make up for over half its HBM shipments in 2024.

During the company's Q2 2024 earnings call on July 25, SK hynix vice president and chief financial officer Kim Woo-hyun said: "We significantly expanded HBM3E shipments in the second quarter as demand was in full swing. In the third quarter, HBM3E shipments will significantly exceed HBM3 shipments, and HBM3E to account for more than half of our total HBM shipments in 2024, we expect".

He continued: "We have already provided HBM3E 12-layer product samples to major customers and will start volume production in the third quarter as planned. With a full product portfolio from HBM2E to HBM3E 12-speed, SK hynix plans to continue its competitive advantage in the HBM market".

NVIDIA CEO Jensen Huang has said that the next wave of AI is "physical AI" which will require three computer systems to make happen: AI, robotics, and Omniverse.

At the recent SIGGRAPH 2024 event, NVIDIA CEO Jensen Huang discussed the next wave of AI, something he calls "physical AI" and that it currently has a three-computer problem, or a three-body problem (shout out to the Netflix series "3 Body Problem" which is fantastic).

Jensen said: "Generative AI, the first wave of, it, of course, is all the pioneers. And we know many of the pioneers: OpenAI, Anthropic, Google, Microsoft, a whole bunch of amazing doing this. X is doing this. xAI is doing this. Amazing companies doing this. The next wave of AI, we did talk about, which is enterprise".

This is a sign of the end times or the latest example of AI moving at a pace that is almost impossible for people to predict. With the recent launch of Meta's open-source and powerful Llama 3.1 AI model, the company hasn't skipped a beat and is currently rolling out its new AI Studio tool in the US.

What is AI Studio? Well, it's described as "a place for people to create, share, and discover AIs to chat with," with no tech skills required. Integrated into Instagram, Messenger, and WhatsApp, these AIs are custom chatbots with a twist. They're your online friends, community, specialists, or even digital twins that can be trained to become you - respond to messages, post content, and even "generate memes."

"You can use a wide variety of prompt templates or start from scratch to make an AI that teaches you how to cook, helps you with your Instagram captions, or generates memes to make your friends laugh - the possibilities are endless," Meta writes in the announcement post.

"Everybody will have an AI assistant," NVIDIA CEO Jensen Huang said at SIGGRAPH 2024. "Every single company, every single job within the company, will have AI assistance." This is a bold statement, to be sure, but not a surprising one considering the state of the industry.

SIGGRAPH is a professional graphics conference, and yes, this year, AI was not only on the menu but also included in every dish. From new microservices for 3D modeling to physics, materials, and robotics, generative AI is driving innovation. At SIGGRAPH, NVIDIA also announced that the world's largest advertising company was using generative AI as part of the Omniverse to create content for Coca-Cola - arguably the gold standard for brand advertising.

So, where does the AI assistant fit in? At this year's show, NVIDIA discussed the concept of digital agents, which are digital AIs trained on specific data. For example, an AI modeled after everything you've ever written, said, or done at work (that is measurable) could then become a personal AI assistant.

Apple has said it has been using Google TPU chips to train its AI software infrastructure, which will power its upcoming suite of Apple Intelligence, AI tools, and features.

In a new research paper from Apple, the company detailed the hardware and software infrastructure of its AI tools and features without any mention of NVIDIA hardware whatsoever. Apple said in its research paper that to train its new AI models, it used two different TPUs from Google that are organized in large clusters of chips.

Apple used 2048 x TPUv5p chips from Google for the AI model that will work on the iPhone and other devices, while the company used 8192 x TPUv4 chips for its server AI model. NVIDIA doesn't design TPUs but instead makes GPUs for gaming, workstations, AI GPUs, and more.

The power required to run complex, massive clusters of high-performance AI GPUs continues to skyrocket with the power of the AI chips themselves, but new research has a reduction in energy consumption required by AI processing by at least 1000x.

In a new peer-reviewed paper, a group of engineering researchers at the University of Minnesota Twin Cities have showed an AI efficiency-boosting technology, which is in lamens terms a shortcut in the regular practice of AI computations that massively reduces energy consumption for those workloads.

AI computing sees data transferred between components processing it (logic) and where data is stored (memory and storage). The moving around of this data back and forth is the main factor in power consumption being 200x higher than the energy used in the computation, according to this research.

NVIDIA is the company powering the tech industry's massive push into artificial intelligence-powered systems, as the green team is making the incredible hardware that makes it possible for these impressive tools and features to exist.

NVIDIA's dominance in this market was achieved by providing the best hardware for training AI systems, which briefly made the green team the most valuable company on the planet, taking the crown of long-standing tech giants such as Amazon and Microsoft. NVIDIA has since moved down to third place but remains the dependent player in the world of AI-focused microprocessors. With the push into AI many developers want to continue training their creations but don't necessarily want to rely on or keep fueling the massive beast that is NVIDIA.

A new report from The Information provides an example: Microsoft and OpenAI are in talks with several chip designs to create a new AI chip to rival NVIDIA. One of those companies was Broadcom, which is ranked the 13th most valuable company in the world for its solutions in semiconductors and software infrastructure.

X, formerly known as Twitter and the digital echo chamber for Elon Musk and his politics, also has a chatbot and powerful AI called Grok. Created by xAI, alongside hundreds of thousands of high-powered NVIDIA GPUs, Grok is described as an AI with a "rebellious streak" that will deliver candid, unfiltered responses.

X recently updated its terms and settings for all users. By default, it uses all X data for training. Grok now has access to everybody's posts, including yours, if you're on X. This move follows Meta and is understandable, given that massive amounts of raw data are a key ingredient for training and creating complex AI models like Grok.

Several AI companies and models have been under fire lately, with reports indicating that some have been scraping YouTube and other public forums to train AI. According to a Microsoft executive, if it's online, it's free to scrape. So, yes, X, Elon, and Zuckerberg are not alone in looking to social media platforms for AI training. The good news is that you can opt-out.

Meta has been training on its new Llama 3 405B model on a cluster of 16,384 x NVIDIA H100 80GB AI GPUs. Half of the issues during its 54-day training run were caused by the onboard HBM3 memory.

Meta released a new study detailing its Llama 3 405B model training, which took 54 days with the 16,384 NVIDIA H100 AI GPU cluster. During that time, 419 unexpected component failures occurred, with an average of one failure every 3 hours. In half of those failures, GPUs or their onboard HBM3 memory were the blame.

In a system with truck loads of components like CPUs, motherboards, RAM, SSDs, GPUs, power systems, cooling systems, a supercomputer is exotic and ultimately powerful, but it's completely normal for issues to happen every few hours. But, it's how developers work on those issues and get the system to remain operational no matter what local breakdowns are happening.

AMD has introduced Amuse 2.0, a new AI image generation tool currently in Beta. Amuse is a fully local experience, meaning it doesn't require plugging into the cloud. It requires either an AMD Ryzen AI 300 Series processor or a Radeon RX 7000 Series graphics card to run. Thanks to their AI-based XDNA architecture, it also runs on systems with AMD's mobile Ryzen 8040 processors.

This is important because Amuse includes AMD XDNA Super Resolution integration, which upscales lower-resolution images on AMD mobile devices. According to AMD, it "increases output size by 2X at the end of the image generation stage."

As seen in other AI image generation tools, Amuse uses Stable Diffusion models from Stability AI to create its images. Amuse takes the widely used and available Stable Diffusion AI models and creates a "painless, easy to use and optimized end-user experience" for its customers. In creating Amuse, AMD partnered with the New Zealand-based TensorStack to help develop the user-friendly UI.

Amazon is currently letting around half a dozen engineers test out a "closely guarded" new server design through its paces, according to the latest reports.

In a new article from RReuters, the outlet reports that the server in question was "packed" with Amazon's artificial intelligence (AI) chips that will compete with the likes of NVIDIA and its market-leading AI GPUs. The news is coming directly from Amazon executive Rami Sinno, said on Friday to Reuters during a visit to the Amazon AI chip lab.

Amazon is developing its own AI processors to limit its future reliance on more expensive NVIDIA AI GPU offerings and the so-called NVIDIA tax. Amazon uses AI all across its Amazon Web Services (AWS) with the company planning to spend $100 billion on data centers used for AI in the future.

The US Department of Energy (DoE) is working on its next-generation Discovery supercomputer, which will be a whopping 3-5x faster than its existing Frontier supercomputer.

The DoE has put out requests for proposals for its new Discovery supercomputer, with interested parties having until August 30, 2024 to submit their proposals. The next-generation Discovery supercomputer would be delivered to the Oak Ridge Leadership Facility (OLCF) at the Oak Ridge National Laboratory (ORNL) in Tennessee, by early 2028.

Discovery will succeed the Frontier supercomputer, which is the world's fastest supercomputer on the biannual Top500 list, which lists the world's fastest supercomputers -- of which Frontier is #1 -- for the fifth consecutive time in May 2024.

TSMC has reportedly received a ton of "super rush orders" from mainland Chinese customers, who are looking to stock up on chips so fast they're willing to pay a 40% premium on them.

The Chinese companies buying chips made at TSMC are expecting a big impact from the 2024 US presidential election to have (more) negative effects on China's access to the latest chips, according to CNyes.

In a new report picked up by the site, "people familiar with the matter" said that TSMC's excellent gross profit margin for Q2 2024 was "unexpectedly" driven by a large number of orders from mainland China, many of which were super-urgent orders.

Artificial intelligence-powered tools and applications are certainly impressive in what they can generate, but where these AI companies get the data to train these impressive models remains ambiguous or completely closed off from public knowledge.

A new report revealed only last week that Apple, NVIDIA, Anthropic, and others used a public dataset containing hundreds of thousands of YouTube video transcripts to train their AI models. While Apple, NVIDIA, and others weren't the ones to download the transcripts, the data was still used to train AI models, which strictly violates YouTube's Terms of Service (ToS). Earlier in the year, YouTube's CEO stated that any data downloaded from its platform is a violation of its ToS.

Now, a new report from 404 Media states that the popular AI video generator company Runway trained its Gen-3 Alpha model on thousands of YouTube videos without obtaining permission from the creator or YouTube. The report also states the company used pirated content for AI model training. 404 Media was sent a spreadsheet that lists how many videos were taken from a specific source, and judging from the list, the sources are extensive and cover a large variety of channels.

NVIDIA AI Foundry is a new enterprise service where NVIDIA will help companies build AI 'Supermodels' - custom Llama 3.1 generative AI models built with NVIDIA software and expertise. These AI 'Supermodels' are custom-trained for specific industry use cases using proprietary and synthetic data from Llama 3.1 405B and the NVIDIA Nemotron Reward model.

Hardware-wise, it's powered by the NVIDIA DGX Cloud AI platform, which allows it to access the most advanced AI hardware, architecture, and technologies. It's an NVIDIA AI Foundry, so that's to be expected. Meta's open-source Llama 3.1 405B model was announced earlier this week, offering capabilities and power previously limited to closed-source models. This new partnership between NVIDIA and Meta is a big deal for companies everywhere.

"Meta's openly available Llama 3.1 models mark a pivotal moment for the adoption of generative AI within the world's enterprises," said Jensen Huang, founder and CEO of NVIDIA.

iOS 18 is Apple's biggest overhaul to its operating system in years. It will be the first update to introduce AI-powered features to the Cupertino company's product lineup.

The new operating system update coming to the iPhone, Mac, and iPad contains Apple's version of artificial intelligence, which it's calling Apple Intelligence. Last week, Apple rolled out the public beta for iOS 18, but it doesn't include the suite of Apple Intelligence features, leaving users confused as Apple said it would be included in iOS 18. However, it will likely first go to developer beta and then to a public opt-in beta before being pushed out to the wider public.

So, when will you be able to try Apple's new AI features? I can see Apple limiting access to its AI features all the way up until the launch of the iPhone 16, as the company won't want to prematurely release its biggest selling point for its new generation of devices only to see its customers go out and buy now-discounted previous generations that are still AI compatible.