Artificial Intelligence - Page 6

In a move that came out of nowhere, OpenAI announced on X that it will shut down its Sora AI video generation app and its associated API. The company did not mention any dates for Sora's official termination, but that information will apparently be available soon.

- The Sora Team via X

It's no secret that investment in data centers and AI infrastructure has been straining energy resources, and that the biggest players in the industry are looking for a clean solution to meet the energy needs of the AI factories of the future. One of those solutions is nuclear power, with Microsoft's Darryl Willis, the Corporate Vice President of the company's Worldwide Energy and Resources Industry, calling it "the essential backbone for this future."

And with the newly announced AI collaboration for nuclear power between Microsoft and NVIDIA, both companies are looking to accelerate the construction of new nuclear power plants for AI. With the help of AI. Yes, that means leveraging cutting-edge AI tools for streamlining the permitting process, which can cost "hundreds of millions of dollars." Plus, the creation of Digital Twins and simulations to enable faster iteration.

Microsoft is clear that this collaboration isn't simply about accelerating the process, but also about enabling engineers and regulators to focus on "building a safe, secure, high-capacity, carbon-free power source that's on-time and on-budget." Here's a more detailed snippet from the announcement.

In a not-so-shocking turn of events, it now looks like Google Search is incorporating generative AI to make questionable changes to the headlines of its search results. Google Search has been a trustworthy search engine for people of all kinds since its inception, with a very straightforward "what you see is what you get" approach. Google CEO Sundar Pichai announced plans to integrate AI into search in 2025, and now it seems they are expanding on it even further.

In a recent discovery by The Verge, it seems like Google is putting its own unwelcome twist on several headlines using the power of our "good" friend, generative AI. The Verge uncovered multiple instances in which Google replaced headlines it had written with new ones that sometimes detracted from the original meaning.

One particularly egregious example was that the headline "I used the 'cheat on everything' AI tool, and it didn't help me cheat on anything" was changed to "Cheat on everything AI tool." This was particularly infuriating, as it almost entirely changes the meaning the article's author was trying to convey. The new AI-generated headline can easily be mistaken for a product endorsement rather than the scathing criticism it was.

NVIDIA GTC 2026 has been a big one for announcements, from a first look at next-gen technologies on the horizon to glimpses at what's in store for the future. Attending this year, TweakTown's Jon Coulter got a chance to walk through the showfloor and spend time with some of the biggest names in memory, storage, GPU, and AI solutions. Naturally, he took plenty of pictures of what was on display.

First, because it's impressive tech that would have been considered impossible or unachievable not too long ago, and second, because this is the sort of hardware and technology that will take a few years, or longer, to make its way to the consumer space.

Naturally, one of the highlights was getting to see a NVIDIA Vera Rubin Superchip prototype in the flesh (or silicon) at SK Hynix's booth, which was signed by NVIDIA CEO Jensen Huang. In addition to featuring NVIDIA's next-gen CPU and GPU architecture for AI, it also sports the latest SOCAMM2 and HBM4 flash memory from SK Hynix.

KIOXIA's open-source AiSAQ (All-in-Storage ANNS with Product Quantization) technology is something we've covered in the past, and it has been a game-changer for AI workloads with Retrieval-Augmented Generation (RAG) pipelines by offloading vectorized data from costly DRAM to a more efficient, cost-effective SSD solution.

This opens the door to scaling, especially for running more complex AI workloads. And when it comes to large-scale workloads and systems, at NVIDIA GTC 2026, KIOXIA demonstrated that, using the NVIDIA cuVS Library with KIOXIA AiSAQ Technology, it successfully scaled high-dimensional vector search to an impressive 4.8 billion vectors on a single server.

With the NVIDIA cuVS Library offering a "significant reduction in index build time by leveraging GPU acceleration," this breakthrough is a big step forward for retrieval augmented generation (RAG) search solutions. And with that, KIOXIA is already looking toward supporting larger-scale event deployments that go beyond 4.8 billion vectors.

NVIDIA GTC 2026, the company's long-running AI conference, kicks off next week in San Jose, and KIOXIA will be there, showcasing its cutting-edge memory and SSD storage solutions built to meet the demands of AI workloads head-on. GTC is all about bringing the AI industry together, from researchers to developers and businesses, with KIOXIA's focus centered on how its flash storage can play a critical role in empowering scalable and efficient AI infrastructures.

KIOXIA's booth will feature a wide range of demonstrations, including the company's groundbreaking AiSAQ (All-in-Storage ANNS with Product Quantization) technology. With the large volumes of data required for modern generative AI systems, this open-source solution offloads large-scale AI vector search from increasingly expensive DRAM to SSD storage - like KIOXIA's CM9 Series Enterprise NVMe SSDs.

This plays into KIOXIA's "Redefining Vector DB Management: Uniting Massive Scalability with GPU-Accelerated Indexing" theater session on Wednesday, March 18, presented by Rory Bolt, a senior fellow and principal architect at the company.

At the Game Developers Conference (GDC) in San Francisco, NVIDIA and ComfyUI announced new updates that will streamline AI video generation on RTX GPUs and the DGX Spark for game developers. Although the use of generative AI in game development is somewhat controversial, these tools are all about 'concept development and storyboarding.' Plus, with the new simplified App View interface from ComfyUI, 4K video generation will work on any rig with GeForce RTX 5090.

ComfyUI App View is designed to be a simple, easy way to generate video, with users only needing to enter a prompt, adjust settings and parameters, and then generate the video. This will sit alongside the more detailed, AI enthusiast-friendly Node View, with users given the option to switch between the two. Optimized for RTX hardware, NVIDIA confirms that performance has been improved by 40% since September 2025, with expanded native support for NVFP4 and FP8 formats.

That part is key because it opens the door to 2.5X faster performance and reduces VRAM requirements by 60% on GeForce RTX 50 Series graphics cards when using NVFP4. And with NVIDIA RTX Video Super Resolution support, AI-generated videos can be upscaled to 4K in a matter of seconds. If you've got a powerful GeForce RTX GPU and are looking for a powerful local AI video generation, then the following NVIDIA Studio Sessions video is for you.

With the emergence of AI-powered tools, there have been genuine concerns about their misuse for nefarious purposes, such as impersonating public figures to sell a product or pushing a specific message online.

These AI "deepfake" videos, which are essentially AI-generated content designed to impersonate an individual, are a real problem that social media platforms will need to tackle. Now YouTube has announced it's upping its defences against AI deepfakes by expanding its likeness detection technology.

YouTube has outlined in a new blog post that creators in the YouTube Partner Program can now submit a short video of themselves along with a government ID to teach the system what they look like.

NVIDIA has announced its financial results for the fourth quarter and fiscal year, which saw the company generate a record $68.1 billion in revenue. This is up 20% from the third quarter and 73% from the same period a year ago, and is driven primarily by NVIDIA's record Data Center segment revenue of $62.3 billion.

This too was up 75% from a year ago, showing that the AI boom and appetite for the company's "accelerated computing and AI" are growing. NVIDIA's Data Center revenue for the full year also increased by 68% to a record $193.7 billion.

"Computing demand is growing exponentially, the agentic AI inflection point has arrived," said Jensen Huang, founder and CEO of NVIDIA. "Grace Blackwell with NVLink is the king of inference today, delivering an order-of-magnitude lower cost per token, and Vera Rubin will extend that leadership even further."

AMD and Meta have announced a "multi-year, multi-generation partnership" that will see Meta deploy 6 gigawatts of AMD AI hardware, including various Instinct GPUs. This includes AMD's Helios rack-scale architecture optimized for Meta's AI workloads, which it plans to utilize to accelerate the deployment of new cutting-edge AI models. This partnership will also see the two companies align their roadmaps covering "silicon, systems, and software."

With the first gigawatt expected to begin deployment later this year, the deal also includes Meta buying a stake in AMD, which could see it own around 10% of the company. This "performance-based warrant for up to 160 million shares of AMD common stock" has multiple milestones attached, and fully vests when Meta deploys all 6 gigawatts of AMD Instinct GPUs.

"We are proud to expand our strategic partnership with Meta as they push the boundaries of AI at unprecedented scale," said Dr. Lisa Su, chair and CEO, AMD. "This multi-year, multi-generation collaboration across Instinct GPUs, EPYC CPUs, and rack-scale AI systems aligns our roadmaps to deliver high-performance, energy-efficient infrastructure optimized for Meta's workloads."

Recently, OpenAI CEO Sam Altman sat down for a lengthy interview with The Indian Express, where he gave a rather strange and cold response to a question about the energy required to train complex AI models. This has become a significant concern in many markets, as AI energy consumption is now dwarfing that of most, if not all, other industries.

And with that, the underlying approach of Sam Altman's bizarre response is more or less the idea that if a human can complete a task in a few seconds or minutes versus an AI model that took vast amounts of energy, more than, say, a small city, to train, what's the point or benefit?

"One of the things that is always unfair in this comparison, where people talk about how much energy it takes to train an AI model relative to how much it costs for one human to do an inference query," Sam Altman says. "It also takes a lot of energy to train a human; it takes 20 years of life, and all of the food that you eat during that time, before you get smart."

NVIDIA CEO Jensen Huang has been hanging out with engineers from Korean semiconductor giant SK hynix, enjoying some chicken and beer in California, where he teased that "a chip that will surprise the world will be unveiled at GTC next month".

In an exclusive interview with Jensen and Korean media outlet KED, the NVIDIA CEO said: "we're at the beginning of a billion-dollar infrastructure project". While enjoying some chicken and beer with NVIDIA and South Korean engineers, Jensen said: "we're one team with Korean semiconductors".

Jensen met with the KED reporter after having dinner with around 30 NVIDIA and SK hynix engineers at 99 Chicken, a Korea-style fried chicken restaurant in Santa Clara, California. The NVIDIA CEO readily accepted the unscheduled interview request, saying: "ask as much as you want".

Mustafa Suleyman, Microsoft's AI CEO, recently sat down with the Financial Times to discuss all things AI, and he had a few words to say about AI one day taking over and making jobs and employment a thing of the past. In fact, when it comes to white-collar jobs that involve sitting at a computer day in and day out, Suleyman has set a timeline of 12 to 18 months for those jobs to become redundant and "fully automated by an AI."

"I think we're going to have a human-level performance on most, if not all, professional tasks," Mustafa Suleyman says. "White-collar work, where you're sitting down at a computer - either being, you know, a lawyer, or an accountant, or a project manager, or a marketing person - most of those tasks will be fully automated by an AI within the next 12 to 18 months."

Suleyman notes that this will be due to AI achieving "human-level performance" on nearly all professional tasks, regardless of industry. He adds that this AI advance has already reached the software engineering sector, where he says "AI-assisted coding" is now a mainstay.

SK hynix will be showcasing its latest 16Gbit (2GB) LPDDR6 SDRAM design at International Solid-State Circuits Conference (ISSCC) 2026 next week.

The new 16Gbit LPDDR6 memory is aiming for 14.4Gbps per I/O pin, aligning with the fastest transfer speed for the LPDDR6 standard, with SK hynix saying its 16Gbit LPDDR6 memory design is built on its new 1c DRAM process, its latest 10nm-class node. The upcoming ISSCC 2026 preview also notes that the paper will be focusing on power saving and signal handling changes to hit 14.4Gbps operation on the new LPDDR6 modules.

Samsung has also reportedly started providing early LPDDR6X memory samples to Qualcomm according to The Bell, with the new standard not detailed yet, but we know that LPDDR6X commercialization will happen in the second half of 2027.

Samsung has been fighting hard on its semiconductor and HBM memory business over the last few years, but now it has officially started commercially deploying its next-gen HBM4 memory, ready for NVIDIA's new Rubin AI chips.

The company explained in a press release that its new HBM4 memory has transfer speeds of 11.7Gbps, but when overclocked like NVIDIA needs it, Samsung's new HBM4 is capable of 13Gbps. Its new leading-edge DRAM is based on a 4nm logic die for maximum performance, fabbed in-house at Samsung Foundry, with its 1c DRAM also in play.

Sang Joon Hwang, Executive Vice President and Head of Memory Development at Samsung Electronics, said: "Instead of taking the conventional path of utilizing existing proven designs, Samsung took the leap and adopted the most advanced nodes like the 1c DRAM and 4nm logic process for HBM4. By leveraging our process competitiveness and design optimization, we are able to secure substantial performance headroom, enabling us to satisfy our customers' escalating demands for higher performance, when they need them".

NVIDIA only just recently asked for 9Gbps HBM4 before pushing for 10-11Gbps, but now rumor has it that NVIDIA has quietly lowered the required HBM4 spec speeds for Rubin, because if they didn't, they probably wouldn't be able to enjoy the full volume needs.

In a new post on X from analyst @Jukan, who posted: "Anyway, just wait and see. There's a rumor going around that NVIDIA lowered the required speed specs for HBM4 because even with Samsung alone, they probably can't meet the full volume NVIDIA needs. Who knows? Maybe Micron could sneak in and supply some HBM4 through this gap. (But personally, I still think SK Hynix is the one who's gonna supply it lol)".

If the rumors are true, it would make sense as according to other recent reports, US-based Micron won't be providing any of its HBM4 to NVIDIA for Rubin, with SK hynix providing 70% of the company's HBM4 needs, and Samsung Electronics with the other 30% of HBM4 supply.

Intel and SoftBank announced their next-generation ZAM memory technology recently, but now the new ZAM memory prototype has been shown off at the recent Intel Connection Japan 2026 event.

The focus on ZAM memory discussion during the event was on how the new Z-angle architecture would help mitigate performance issues and improve thermals using existing cooling technology. Intel Fellow and CTO of Intel Government Technologies, Joshua Fryman, was there alongside Intel Japan CEO Makoto Ono.

ZAM has been limited to just research papers and press releases, but with the new Intel and SAImemory (SoftBank subsidiary) the team is now pushing ahead with ZAM and some prototypes. The biggest difference here compared to HBM and other memories is that ZAM integrates a massive amount of interconnect topology that is routed diagonally throughout the die stack, instead of drilling vertically down. Intel says that the biggest benefit with ZAM is its thermal capabilities.

OpenAI has announced that it is currently testing ads in ChatGPT in the U.S., for users on the Free and Go subscription tiers. The AI company is quick to note that the addition of ads won't "influence the answers ChatGPT gives you" and that your conversations with the AI platform will remain private and won't be used for marketing.

The good news for those who pay for or use Plus, Pro, Business, Enterprise, or Education accounts is that you won't see ads, and ChatGPT will remain unchanged. OpenAI has showcased what the ads will look like and how they will clearly be marked as sponsored. Ads will be related to the subject or topic, with a food-and-recipe example interaction delivering the sort of ad you might see elsewhere online.

OpenAI notes that it's adding ads to ChatGPT to support "broader access" to its features, presumably to cover the costs of hundreds of millions of people interacting with ChatGPT every day.

In a new post tagged 'the future of gaming is AI,' the company, best known for creating gaming peripherals, explains why it's investing over $600 million in AI. And that's primarily about AI tools and technologies for game development, with Razer believing that how generative AI is used is more important than whether AI is used at all.

"The way we see it is that AI is a tool to help game developers make better games, rather than replace human creativity," Razer CEO and Co-founder Min-Liang Tan said during a recent episode of The Verge's Decoder podcast. "As gamers... what we're unhappy with is GenAI slop. When I play a game, I want to be engaged. I want to be immersed. I want to compete. I don't want to see characters with extra fingers or shoddily written storylines."

That comment is in response to the influx of AI-generated images and videos, widely referred to as "AI slop," which are considered inferior to human-created art. For Razer, generative AI in games is more of an extension of NPC behavior, procedural systems, and AI used to "strengthen the craft of making games."

When it comes to images, video, voice, and music, AI generation has reached a point where a wide range of models can produce impressive results. That said, on the creative side of generative AI, most users still connect to cloud-based services. ACE Step 1.5 is an open-source foundation model for generating music, and now it's been optimized to run locally on AMD Ryzen AI processors and AMD Radeon graphics cards.

AMD notes that you can generate full-length music tracks like the 'Country Ballad' example above, iterate, and keep all assets on-device from initial prompt to the final piece of generated music.

"Without per-track fees or upload limits, creators can experiment freely, enabling them to sketch ideas, test arrangements, and explore new sounds without friction," AMD writes. "On-device generation enables immediate iteration, making it easier to refine or discard ideas in real time without relying on an internet connection."