Artificial Intelligence - Page 27

The AI industry was set alight when DeepSeek unveiled its R1 AI model, which the company claims is comparable in performance to OpenAI's latest AI model but was created for a fraction of the price.

The new R1 model was created with just $6 million, which is a tiny drop in the lake of money that is pouring into AI companies all throughout the world. For context, OpenAI said it spent approximately $100 million training its latest model, "o3", and it's estimated the company could spend as much as $500 million to create GPT-5, the company's AI model that's on the horizon.

Having a new AI model be released for a fraction of the costs of known AI models caused a massive pullback from investors, resulting in approximately $1 trillion being wiped away from AI companies and the companies that fuel their progression.

DeepSeek is in some deep trouble after cloud services provider Wiz Research did some digging, finding a publicly available database of the Chinese AI company's secret keys, messages, and more.

The database includes details like chat history, API secrets, and more, with Wiz researchers saying that the discovered database can be fully controlled. This means that they could execute code without any oversight, manipulating the data to serve their needs. Wiz reached out to DeepSeek to alert them, with the company quickly taking it down after restricting public access.

The team at Wiz Research set out to assess DeepSeek's external security posture, and find any potential vulnerabilities, with the team explaining: "Within minutes, we found a publicly accessible ClickHouse database linked to DeepSeek, completely open and unauthenticated, exposing sensitive data. It was hosted at oauth2callback.deepseek.com:9000 and dev.deepseek.com:9000".

It has been a tumultuous week in the AI industry after Chinese company DeepSeek unveiled its R1 model, which caused approximately $1 trillion to be wiped away from Silicon Valley AI companies as DeepSeek said it's model was on par with OpenAI's ChatGPT, but cost just $6 million to create, which is a fraction of the billions of dollars poured into ChatGPT.

DeepSeek has truly shaken up the AI industry and has caused AI heavyweights such as OpenAI, Microsoft, and others to reassess their own AI chatbots and processes. But now accusations are beginning to fly, with OpenAI and Microsoft alleging they have obtained evidence DeepSeek used stolen OpenAI data to train its R1 model. In a recent Financial Times article, OpenAI claimed DeepSeek trained its AI with OpenAI's models, while Microsoft told Bloomberg that it believes it has evidence of an OpenAI developer account being connected to DeepSeek.

Furthermore, the accusations don't stop there, as Microsoft said this OpenAI developer account is linked to stealing large amounts of data from OpenAI. These accusations were backed up by President Trump's artificial intelligence and cryptocurrency advisor David Sacks, who told Fox News, "There's substantial evidence that what DeepSeek did here is distilled the knowledge out of OpenAI's models." Sacks added it's "possible" DeepSeek has engaged in IP theft.

NVIDIA is distancing itself from Samsung and its HBM memory, with CEO Jensen Huang saying the company "can't trust Samsung Electronics' High-Bandwidth Memory (HBM) products and engineers".

In a new report from Korean outlet Hankyung, which reports that NVIDIA CEO Jensen Huang said in front of Samsung executives: "is a customer of Samsung Electronics, not an employee. Stop calling and asking questions. I can't trust Samsung Electronics' high-bandwidth memory (HBM) products and engineers. We cannot trust and do business with them because senior executives change frequently".

However, even just weeks ago at CES 2025, NVIDIA said "there is no doubt about the success of Samsung Electronics' HBM". Samsung has been struggling to get its HBM3E memory off the ground -- well, out of the semiconductor fabs, you know what I mean -- with NVIDIA reportedly working "as fast as it can" to have Samsung's new HBM3E memory certified to use on its world-leading AI GPUs.

The DeepSeek R1 model has caused a massive stir in the tech community as the AI model that allegedly is on par with OpenAI's o3 model but cost just $6 million to create on old NVIDIA GPUs has cast shades of doubt over the billions of dollars invested into AI development across the United States.

The market has reacted to the unveiling of DeepSeek's R1 model, with approximately $1 trillion being wiped away, and NVIDIA seeing its biggest loss in the company's history with a $600 billion wipeout on Monday. Now, OpenAI CEO Sam Altman has responded to the new R1 model in a series of X posts, writing he believes the new R1 model is "impressive," especially considering what the company was able to achieve given the price tag it paid. Moreover, Altman said that OpenAI will "obviously deliver much better models," but it's also "invigorating to have a new competitor" on the market.

Altman continued in a follow-up post, saying he and the company are excited to continue executing "our research roadmap," and added, "more computing is more important now than ever before to succeed at our mission." The latter is likely in reference to additional computation power needed to train sophisticated AI models. This computational power comes in the form of NVIDIA GPUs, and the problem AI companies are now facing is that DeepSeek achieved an AI model close to the level of OpenAI's latest model with just 3% to 5% of the resources.

In some news I'm sure you didn't expect, ex-Intel CEO Pat Gelsinger has come out praising China's new AI wonder child -- DeepSeek -- and that the former Intel boss is buying up NVIDIA stock as it has experienced a sharp drop in its value after DeepSeek came in like an (AI) wrecking ball.

In a new post on X, Gelsinger explained: Wisdom is learning the lessons we thought we already knew. DeepSeek reminds us of three important learnings from computing history:

Gelsinger was the CEO of Intel until December 2024, after he exited the company amongst controversy and now is the chairman of his own IPO-bound startup -- Gloo -- which is a messaging and engagement platform for churches. In his tenure at Intel, Gelsinger was trying to get Intel Gaudi 3 AI accelerators to compete with NVIDIA... and we all know how that went, because it didn't.

DeepSeek has been in the headlines as it has completely upended the entire AI industry, swiping $2 trillion off the US stock market, and now analysts have called the Chinese AI firm the "Temu of AI".

Now we've got Wedbush analysts, led by Dan Ives, saying in a new note that they "view the DeepSeek fear across the tech world as in essence a 'tech AI head fake' that will be short-lived as more details and analysis comes out about DeepSeek's model and China resources".

DeepSeek's new R1 model only cost $5.6 million to train, which is insane considering that that's about the salary of an American expert in AI can command. AI training costs so, so much more on the opposite side of the pond with prices ranging between $100 million up to $1 billion and beyond... so you can see why analysts have their knickers in a twist.

DeepSeek's game-changing R1 model is reportedly running inference workloads on Huawei's latest Ascend 910C AI chips, showing how much China has been working behind the scenes on its issues in the AI industry.

In a new post on X from Alexander Doria, we're finding out that DeepSeek R1 was trained on NVIDIA H800 AI GPUs but inference is on the new homemade Chinese chips from Huawei in the form of the Ascend 910C which we were reporting about last year.

Huawei's in-house Ascend 910C AI chips are a direct rival for NVIDIA H100 AI GPUs, and while we don't know about the specifications of the Ascend 910C, rumor has it mass production of the 910C is meant to begin in Q1 2025... so, any minute now. China can't get its hands on advanced AI chips because of US export restrictions, it seems it doesn't need them as it has changed the AI landscape with DeepSeek R1.

DeepSeek AI is currently making headlines because it's doing more with less, creating complex and powerful open-source large-language models (LLMs) with a fraction of the hardware of some of the most prominent AI players. KIOXIA, a pioneer and leader in flash memory technology, has showcased how AiSAQ's all-in-storage solution for AI exponentially improves performance without the need for expensive DRAM.

At CES, we got to check out some cool new technology, but one AI-based breakthrough blew us away with its potential to shake up the industry. KIOXIA's open-source AiSAQ (All-in-Storage ANNS with Product Quantization) is all about leveraging the cost-effectiveness of SSDs over DRAM for data retrieval. And when it comes to AI, we're talking about a lot of data.

Today marks the open-source release of AiSAQ, which delivers exponentially scalable performance for retrieval-augmented generation (RAG) without placing index

DeepSeek AI is making headlines because the Chinese startup has developed cutting-edge and cost-effective large-language models using fewer GPUs than systems costing billions from companies like OpenAI, Meta, and Google. With the company's flagship R1 model performing on par with OpenAI's o1, this has spilled over to the financial markets, where NVIDIA's stock dipped to the point where it saw the biggest single-day drop in US history.

Despite its market cap dropping by several hundreds of billions, NVIDIA recently released a statement praising DeepSeek as "an excellent AI advancement and a perfect example of Test Time Scaling." DeepSeek's efforts are being widely praised because it didn't require billions of dollars of AI hardware to train.

OpenAI CEO Sam Altman released a short statement earlier today on social media on DeepSeek R1, praising it as an "impressive model" while teasing "much better models" coming soon from the company. He also said that OpenAI will bring a few releases forward in response to DeepSeek's surprising and immediate success.

President Trump will be placing tariffs on the foreign production of semiconductors, as well as having some fresh comments on Chinese AI company DeepSeek and its game-changing ChatGPT competitor.

President Trump said during a gathering of House Republicans at his Doral resort in Miami: "In the very near future we're going to be placing tariffs on foreign production of computer chips, semiconductors and pharmaceuticals to return production of these essential goods to the United States of America".

Trump said: "I view that as a positive, as an asset. The release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser-focused on competing to win. Because we have the greatest scientists in the world, even Chinese leadership told me that. They said you have the most brilliant scientists in the world".

NVIDIA executives would be waking up to quite the headache this morning, with Chinese AI company DeepSeek taking the headlines with its new AI model seemingly destroying OpenAI in one swift release.

In a statement to Futurum Group analyst Daniel Newmann, NVIDIA explained that DeepSeek's new AI model is an excellent AI advancement and a perfect example of Test Time Scaling.

In full, NVIDIA said in a statement: "DeepSeek is an excellent AI advancement and a perfect example of Test Time Scaling. DeepSeek's work illustrates how new models can be created using that technique, leveraging widely available models and compute that is fully export control compliant. Inference requires significant numbers of NVIDIA GPUs and high-performance networking. We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling".

DeepSeek's introduction into the world of AI assistants has shaken the world, seeing NVIDIA stock bleeding out $500 billion with no signs of stopping... and now the Chinese AI startup has unleashed its new multimodal AI model: Janus-Pro 7B.

Janus-Pro 7B is under an MIT license meaning it can be used commercially without any restrictions, with DeepSeek explaining Janus-Pro as a "novel autoregressive framework" capable of analyzing and creating images. The company says that on two AI evaluation benchmarks -- GenEval and DPG-Bench -- the largest Janus-Pro model, Janus-Pro-7B, beats DALL-E 3 as well as other AI models including PixArt-alpha, Emu3-Gen, and Stability AI's Stable Diffusion XL.

DeepSeek explained on Hugging Face: "Janus-Pro surpasses previous unified model and matches or exceeds the performance of task-specific models. The simplicity, high flexibility, and effectiveness of Janus-Pro make it a strong candidate for next-generation unified multimodal models".

Chinese AI company DeepSeek is making some serious waves in the world of AI, with its R1 model being trained for 90-95% less money... making the industry re-consider all those AI GPU purchases... and seeing NVIDIA stock dropping 11% in pre-market trading, bleeding out $384 billion. Ouch.

It's not just NVIDIA but other major tech stocks are down, with AMD also down by 5.3% in pre-market trading. Microsoft is also down by 6% and Meta is down by 4.6%... with hundreds of billions of dollars spent, and hundreds of billions of dollars more to spend on AI GPUs... these companies must be worried about the impact DeepSeek is going to have on the market.

DeepSeek has said its R1 model costs just $5.6 million to train, and when compared to other AI models like ChatGPT o1 from OpenAI which cost hundreds of millions of dollars to train... on hundreds of billions of dollars of the most advanced AI chips on the planet... which are mostly made by NVIDIA.

SK hynix has announced its highest-ever profits in 2024, riding off of the AI demand wave with its market-leading HBM memory.

The South Korean memory giant posted its best-ever yearly earnings in 2024, with 66.1 trillion won ($46.1 billion USD or so) in sales, with revenues increased by a huge 102% thanks to huge HBM memory sales for AI chips. SK hynix celebrated the record-high milestone with distributed bonuses of 1500% worth of what it defines as base pay, which works out to around 75% of the annual salary for an employee at SK hynix.

SK hynix had operating profits of 23.4 trillion won (around $16.3 billion USD or so) which beat the previous record of 20.8 trillion won (around $14.5 billion USD or so).

Chinese AI company DeepSeek has had some rather large success on the Apple App Store, with its AI assistant now the top free app on the App Store, beating out ChatGPT.

In particular AI benchmarks, DeepSeek's R1 model matches or even beats OpenAI's new o1 model, and better yet... DeepSeek has said its R1 model cost just $5.6 million to train. This is compared to the hundreds of millions of dollars that other companies spend training their AI models.

This is all with US sanctions restricting high-powered AI chips from hitting Chinese shores, leaving DeepSeek to play it differently when it came to powering its AI training on the cheap. DeepSeek is using NVIDIA H100 AI GPUs for some of its training, with DeepSeek founder and CEO Liang Wenfeng has recently said that US export restrictions are still bottlenecking the company.

Chinese memory manufacturer Tongfu has announced it has started trial production of its HBM2 process, and could have the memory chips inside of Huawei AI GPUs in the near future.

In a new report from Nikkei, we're learning that Tongfu Microelectronics has started HBM2 trial production, which is a game-changer for the domestic industry in China. Tongfu isn't the only one kicking off HBM production, with CXMT and Wuhan Xinxin both making significant progress in DRAM and HBM memory over the last few months.

AMD is a partner of Tongfu and a shareholder, making it famous in the industry, and you wouldn't know it... but Tongfu Microelectronics is the world's third-largest semiconductor packaging and test service provider, but the adventures into HBM memory is new for the Chinese company.

Chinese AI lab DeepSeek has access to tens of thousands of NVIDIA H100 AI GPUs for training, according to DeepSeek CEO.

The DeepSeek R1 is one of the most advanced AI models on the planet, competing with the likes of OpenAI's new o1 and Meta's Llama AI models. In a new interview with CNBC, Scale AI founder and CEO Alexander Wang said DeepSeek R1 has met, or beats all-top performing AI models in his firm's most challenging AI test.

CNBC's Andrew Ross Sorkin interviewed Wang, talking about DeepSeek's new AI test called "Humanity's Last Exam" with the "hardest questions" pumped into it by "math, physics, biology, chemistry professors" that are relevant to the latest research. After testing all of the latest AI models, Wang's team found that DeepSeek's new model was "actually the top performing, or roughly on par with the best American models, which are o1".

Meta will spend over $65 billion this year alone on its AI plans, according to CEO Mark Zuckerberg, in order to boost its AI power against rivals OpenAI and Google.

The social networking giant will increase its hiring for artificial intelligence roles, and will build a new 2-gigawatt data centter, which will be so big it would cover most of Manhattan in New York City. Meta is lusting after NVIDIA's latest AI GPUs -- spending tens of billions of dollars -- with an aim of over 1.3 million AI GPUs and plans to bring around 1GW of computing power online in 2025.

Zuck said: "This will be a defining year for AI. In 2025, I expect Meta AI will be the leading assistant serving more than 1 billion people, Llama 4 will become the leading state of the art model, and we'll build an AI engineer that will start contributing increasing amounts of code to our R&D efforts. To power this, Meta is building a 2GW+ datacenter that is so large it would cover a significant part of Manhattan".

Over the last few weeks, OpenAI has been vocal about promoting their latest reasoning model, o3, which is confirmed to be ready for the general public.

The new model is equipped with 'simulated reasoning' capabilities, which essentially allows it to reflect on its own thoughts before formulating a response. Along with a range of other improvements, such as improved visual reasoning, coding, and performance in mathematics.

However, free users of the tool will also benefit from the latest improvements in the form of OpenAI's condensed model - o3-mini. Succeeding the previous GPT-3.5 Turbo, o3-mini comes equipped with enhanced reasoning capabilities, faster response times, and reduced computational requirements. You'll also find improved performance on complex tasks, such as those involving mathematics, science, and programming.