Artificial Intelligence - Page 30

Marvell has just unveiled its new custom HBM compute architecture, enabling XPUs to have even higher levels of compute performance and memory density.

The new HBM compute architecture is available on all of its custom silicon partners, with Marvell collaborating with its cloud customers and leading HBM manufacturers SK hynix, Samsung, and Micron to develop custom HBM solutions for next-generation XPUs.

Will Chu, Senior Vice President and General Manager of the Custom, Compute and Storage Group at Marvell, said: "The leading cloud data center operators have scaled with custom infrastructure. Enhancing XPUs by tailoring HBM for specific performance, power, and total cost of ownership is the latest step in a new paradigm in the way AI accelerators are designed and delivered. We're very grateful to work with leading memory designers to accelerate this revolution and, help cloud data center operators continue to scale their XPUs and infrastructure for the AI era".

OpenAI has just created a more expensive version of its ChatGPT chatbot, with a new "reasoning" model series out of preview status with o1, costing $200 per month.

The new $200 per month subscription tier includes unlimited access to OpenAI o1, GPT-4o, and the new Advanced Voice mode. OpenAI is including a version of o1 that is exclusive to ChatGPT Pro subscribers, which provides more compute power to generate the best answer possible, even with the hardest questions or prompts.

The AI startup has compared its new o1-preview, with people paying $200 per month to expect a more powerful, faster, and more accurate model that is better at coding and math. The new o1 "reasoning" model is also able to respond to images, with OpenAI promising that it's been trained to be more concice, leading to faster response times than the o1-preview model.

NVIDIA CEO Jensen Huang attended the AI Vision event in Thailand last week, announcing a "sovereign AI" project with local AI company Siam.AI Cloud, the country's first Thai NVIDIA Cloud Partner.

In a new post from the Bangkok Post, we're learning that Siam.AI Cloud also announced comprehensive partnerships with leading institutions to drive the national digital transformation with "sovereign AI". NVIDIA CEO Jensen Huang has talked about sovereign AI many times in the past (in the links below) which refers to a nation's ability to provide AI using its own infrastructure, data, workforce, and business networks.

Jensen said: "Thailand recognizes the importance of having AI infrastructure and data. Siam AI Cloud is the first AI cloud in Thailand, the first AI language in Thailand, with the OpenThai GPT local large language model, and this is an extraordinary moment. And frankly, there are only 20 countries in the world that have done this".

UnitedHealthcare CEO Brain Thompson was killed in Midtown Manhattan earlier this week, and with the suspect currently on the run, the public has begun digging into a potential motive for the grim assassination.

With police having yet to make an arrest, publications have begun looking into UnitedHealthcare, and one of the points leading to a possible motive that has gained some attention is a lawsuit filed against UnitedHealthcare in November 2023. The lawsuit filed by two now-deceased individuals, but still ongoing and yet to reach a final conclusion, alleges UnitedHealthcare pressed employees to use an algorithm designed to evaluate how long a patient would be staying based on their condition and issue denials for coverage. Moreover, the suit alleges internally, UnitedHealthcare knew the algorithm had an error rate of approximately 90%.

Notably, the lawsuit specifically states that UnitedHealth and its subsidiary, NaviHealth, are using a computer algorithm reportedly called nH Predict to "systematically deny claims" for coverage for beneficiaries struggling to recover from illnesses while in nursing homes. A UnitedHealth spokesperson responded to the suit, stating the company doesn't use its NaviHealth predict tool to make coverage determinations and that the "lawsuit has no merit and we will defend ourselves vigorously."

TSMC is reportedly in discussions with NVIDIA to make its Blackwell AI GPUs in the USA, producing the world-leading AI chips at its new Fab 21 plant in Phoenix, Arizona.

In a new report from Reuters, we're learning that TSMC is "already making preparations to start production" in early 2025, according to Reuters' sources. NVIDIA unveiled its new Blackwell AI GPUs in March, and are now making their way into tech companies' hands in the form of B200 and GB200, with GB200 NVL72 AI server cabinets insanely popular.

If the agreement between TSMC and NVIDIA is made, it would see the Taiwanese contract manufacturer securing another high-profile client for its US semiconductor manufacturing. TSMC is already preparing to make next-gen chips for the likes of Apple and AMD on US soil, so if NVIDIA joins the party... it's going to be a rather big deal to have some of the most advanced AI chips made in the United States.

Elon Musk's xAI startup has promised to expand its Colossus AI supercomputer by 10x its computational power, with over 1 million GPUs in total... yes, 1,000,000+ AI GPUs.

xAI built Colossus in just 3 months earlier this year, with the world's largest AI supercomputer featuring over 100,000 x NVIDIA IA GPUs operating at once, training Grok for X, and more. The expansion project has already started, with the first step being increasing the size of the facility in Memphis, Tennessee.

NVIDIA, Dell, and Supermicro are going to establish operations in Memphis in order to support xAI's extension of Colossus, while FT reports that the chamber of commerce saying that it would establish an "xAI special operations team" in order to "provide round-the-clock concierge service to the company".

An OpenAI safety researcher has shared a message on her Substack saying she is quitting her position at the company as she believes her goal of implementing humanity-protecting policies into the development of AI can be better achieved externally.

OpenAI has seen a selection of pivotal staff members leave the company recently, and now another has been added to the list. Rosie Campbell joined OpenAI in 2021 with the goal of implementing safety policies for AI development, and now, according to a Substack post, the AI safety researcher is departing the company, citing several internal changes such as workplace culture and the ability to perform what Campbell believes is the most fundamental part of her job - AI safety.

Campbell wrote in the Substack post that she was a member of OpenAI's Policy Research team, where she worked closely with Miles Brundage, a senior staffer who worked at OpenAI's Artificial General Intelligence (AGI) team, a team dedicated to making sure the world is prepared for AGI when it's achieved. Notably, Brundage left OpenAI in October and published a letter on Substack citing concerns with OpenAI's internal policies regarding AGI safety and writing there are "gaps" in the company's readiness policy.

OpenAI has just announced its new "Shipmas" period, with new features, products, and demos for the next 12 days starting December 5.

The ChatGPT creator is expected to debut its much-anticipated text-to-video service codenamed Sora, as well as a new reasoning model according to sources "familiar with OpenAI's plans" according to The Verge. OpenAI CEO Sam Altman confirmed the 12 days of "Shipmas" event on-stage at The New York Times' DealBook conference on Wednesday morning, but didn't elaborate.

Leading up to the launch, OpenAI staffers were teasing some of the upcoming releases on X, with one of them posting "What's on your Christmas list?" while another posted "Got back just in time to put up the shipmas tree". Sora boss Bill Peebles responded to a staffer who said that OpenAI is "unbelievably back" to which he replied with a single word: "Correct".

Google has just introduced its new generative AI video model, Veo, beating OpenAI's text-to-video service Sora, launching Veo into private preview on Google's in-house Vertex AI platform.

Google's new Veo model can generate "high-quality" 1080p resolution videos in multiple different visual and cinematic styles, all from text or image-based prompts. The search giant unveiled its text-to-video model a few months ago with generated clips that would be "beyond a minute" in length, but the company didn't specify... but now, these videos that were made by Veo are pretty astounding.

The latest version of Google Imagen 3 text-to-image generative will be online and available for all Google Cloud customers on Vertex "starting next week" says the company, which will see an expansion of its US-first release on Google's AI Text Kitchen in August 2024.

Amazon Web Services (AWS) has teased its next-gen Trainium3 AI accelerator at re:Invent on Tuesday, promising 4x higher performance than its current-gen Trainium2 chip.

The new Trainium3 AI accelerator is due in late 2025, with Gadi Hutt, director of product and customer engineering for AWS' Annapurna Labs team, expects the new AI accelerator to be the very first dedicated machine learning accelerator built on a 3nm process node (at TSMC) and hit a 40% improvement in efficiency over Trainium2.

Amazon hasn't been too clear on the exact performance of its Trainium3, but the 4x performance improvement figure is based on AWS' complete "UltraServer" configuration, which The Register reports is still in development. The outlet works out that the Trainium2 UltraServer features 64 accelerators, capable of 83.2 petaFLOPS of compute performance (unknown precision).

Meta is shifting into the warm arms of nuclear power for its AI training, with the company posting a new blog on its sustainability website saying that "we believe nuclear energy will play a pivotal role in the transition to a cleaner, more reliable, and diversified electric grid".

In the new post, Meta announces that it is going to release a request for proposals (RFP) to find nuclear energy developers to help them on their nuclear-powered journey. Meta aims to add 1-4 GW of new nuclear eneration capacity in the United States to be delivered "starting in the early 2030s".

Meta explains: "We are looking to identify developers that can help accelerate the availability of new nuclear generators and create sufficient scale to achieve material cost reductions by deploying multiple units, both to provide for Meta's future energy needs and to advance broader industry decarbonization. We believe working with partners who will ultimately permit, design, engineer, finance, construct, and operate these power plants will ensure the long-term thinking necessary to accelerate nuclear technology".

NVIDIA's next-generation Rubin AI GPU architecture release rumored to be pulled up by 6 months, TSMC 3nm process expected, with ultra-fast next-gen HBM4 memory.

The new Rubin AI GPU architecture is the successor to the Blackwell GPU architecture, which is being used in the current fleet of B200 and GB200 chips, as well as the future GB300 series AI GPU that we're hearing more and more about lately. In a new report from UDN, we're hearing that NVIDIA is already working with supply chain partners in Taiwan on the Rubin AI GPU architecture and its new R100-powered AI servers.

Rubin was originally scheduled for 2026, but sources of UDN say that the company has launched the development of Rubin early, so that the AI boom can continue from one AI GPU chip to another (Blackwell to Rubin, and so on).

NVIDIA's beefed-up GB300 AI servers are expected to hit the market in mid-2025, rolling out with even more performance, faster (and more) 12-Hi HBM3E memory, and more.

In a new report from the UDN, we're learning that supply chain manufacturers have already started the process for next-gen NVIDIA GB200 AI servers, which will have massive power consumption increases over the already power-hungry GB200 AI servers.

We heard not too long ago in October 2024 that NVIDIA was reportedly rebranding its upcoming "Blackwell Ultra" AI GPUs to the B300 series, with B300 and GB300 chips using TSMC's new CoWoS-L advanced packaging. The B200 Ultra was reportedly renamed to the B300, while the GB200 Ultra has been renamed to GB300, while B200A Ultra and GB200A Ultra are now B300A and GB300A, respectively.

Elon Musk has reportedly directly approached NVIDIA CEO Jensen Huang, offering a premium price to get priority access to its new GB200 AI servers, with a hefty $1.08 billion order.

In a new report from DigiTimes, industry sources have said that Elon Musk's xAI startup wants its hands on NVIDIA GB200 AI servers, and it doesn't want to wait: with CEO Elon Musk stepping in and making a call to NVIDIA CEO Jensen Huang, waving $1.08 billion around for priority access to the most powerful AI GPU silicon on the planet.

xAI's huge $1.08 billion order for NVIDIA GB200 AI GPUs will be manufactured by NVIDIA's key partner, Foxconn, and should be delivered in January 2025.

Panasonic was originally founded in 1918 by Kōnosuke Matsushita under the name Matsushita Electric Housewares Manufacturing Works, and now the founder has been resurrected by the modern-day Panasonic.

Panasonic collaborated with the University of Tokyo to resurrect Matsushita who died in 1989. Engineers fed 3,000 recordings of Matsushita into an AI, along with any relevant writings, lectures, and interviews. Panasonic's Peace and Happiness through Prosperity (PHP) Institute, a think tank originally founded by Matsushita, plugged all of the relevant data into an AI and trained it to character an AI character that is designed to replicate Matsushita's way of thinking and speaking style.

What's the goal of this? Panasonic wants to use the AI replica of the company's founder as a consultant, querying it in difficult situations to see what Matsushita would do based on the current circumstances. Notably, Matsushita was renowned for his incredible management philosophies in business, and is incredibly celebrated in Japan for the transformation of what was once a business that sold lamps into what Panasonic is today.

Companies developing artificial intelligence tools require large swaths of data for AI training, and what better way to gather large quantities of data than by scraping it from people using popular applications or programs?

@nixCraft, an author of Cyberciti.biz has claimed Microsoft is participating in this type of scheme with Office, and it's Connected Experiences. According to nixCraft, Redmond's Connected Experiences feature automatically scraps data from Word and Excel files, and that data is used to train Microsoft's AI tools, such as Copilot. According to reports, this feature is turned on automatically, which means user-generated Word documents and Excel files are included in Microsoft's AI training dataset unless the user manually disables the feature.

However, following reports sourcing @nixCraft's claims, Microsoft has since responded, saying customer data within Microsoft 365 apps, which includes Word and Excel, isn't used to train the company's Large Language Models (LLMs), the underlying technology powering AI tools such as Copilot, or ChatGPT. Microsoft also added, "This setting only enables features requiring internet access like co-authoring a document."

Amazon has reportedly developed a new video AI model codenamed Olympus, a new generative artificial intelligence (AI) that can process images and videos (as well as text) according to The Information.

The outlet reports that Amazon's move with its new video AI model will reduce its reliance on Anthropic's chatbot, Claude, which is a popular model that Amazon Web Services (AWS) offers. The Information talked to two sources, who said that the new large language model (LLM) is codenamed Olympis, and will be capable of understanding and processing scenes in images and videos, helping people look for specific scenes for their video.

This can include making the best coffee, raindrops falling to the ground, and so much more with easy text prompts according to The Information's sources. Amazon's new Olympus video AI model could be announced as soon as next week, with the annual AWS re:Invent customer conference according to The Information's source.

IMAX is stepping up its game, announcing a partnership with Dubai-based startup Camb.ai, using its AI speech models to translate IMAX original content, including documentaries, to increase its global reach.

The entertainment industry has been going through some gigantic shifts, with the entertainment and media industry growing 5% to $2.8 trillion in 2023 according to a report published by PwC. The industry is expected to continue its momentum, with a compound annual growth rate of around 4% to $3.4 trillion in the next 5 years.

The fact that non-English language content is growing rapidly (even in English markets) including the US, UK, Australia, and Canada. Netflix reported 90% growth in its viewership of non-English content in the UK over the last 3 years, reports Yahoo Finance.

Dell has just announced its new IR7000 servers that support up to 480kW per rack of computing power with Direct Liquid Cooling. Check it out:

Dell CEO Michael Dell posted on X saying: "Our new IR7000 supports up to 480kW per rack of computing power with Direct Liquid Cooling. The flexible design allows for any combination of Cloud, AI and HPC. The system is engineered to capture nearly 100% of the heat produced within each rack".

Dell has three example configurations of its new IR7000 server racks:

NVIDIA is "working as fast as it can" to certify Samsung's new AI memory chips -- HBM3E -- with CEO Jensen Huang himself telling Bloomberg TV.

Jensen talked with the press after attending a ceremony at the Hong Kong University of Science and Technology, where the NVIDIA CEO was given an honorary doctorate in engineering. Jensen said that NVIDIA is looking at ways to purchase 8-layer and 12-layer HBM3E memory chips from Samsung, while as it stands, the company is acquiring most of its HBM chips used on its AI GPUs, from South Korean rival SK hynix.

Earlier this week during NVIDIA's Q3 2024 post-earnings call with analysts, the CEO didn't mention Samsung when talking about the company's major suppliers -- naming SK hynix and Micron -- but now, the company is "working as fast as it can" to get Samsung HBM3E certified.