Artificial Intelligence - Page 26

The ASUS AI POD, which features the NVIDIA GB200 NVL72 platform, has reached its latest milestone. Production is ramping up for a scheduled shipping date of March 2025. This thing is an absolute beast for AI performance and cutting-edge deep learning, as it's built with the NVIDIA GB200 Grace Blackwell Superchip and fifth-generation NVIDIA NVLink interconnect technology.

When broken down, you've got 26 NVIDIA GRACE CPUs, 72 NVIDIA Blackwell Tensor Core GPUs, and 5th Gen NVIDIA NVLink Switches for low-latency communication with a capacity of 14.4 TB/s. Each of the 72 Blackwell GPUs includes two dies connected by a 10 terabytes per second (TB/s) chip-to-chip interconnect to offer 30X the LLM Inference of the NVIDIA H100 Tensor Core GPU and 4X the LLM Training performance of the H100.

The ASUS AI POD also supports liquid-to-air and liquid-to-liquid cooling, and it's powerful enough to train a trillion-parameter language model and perform AI inference in real-time. The cooling includes CPU/GPU cold plates and coolant distribution to increase power efficiency.

Chinese start-up DeepSeek has caused the AI market to flip after its new chatbot, which is comparable to OpenAI's latest AI model, was made available for much cheaper.

The AI market is still reeling from the unveiling of DeepSeek, with the announcement dramatically affecting the stock value of AI companies, including NVIDIA, which lost an estimated $600 billion, and OpenAI, which has accused DeepSeek of using its database. Moreover, the US government has launched a probe into DeepSeek to see if the company circumvented US trade restrictions on high-end NVIDIA AI GPUs by using Singapore as a middleman. It's safe to say the AI market is in turmoil right now since the new player entered the field, and now we are starting to see its effects in terms of users.

According to data from Semrush, ChatGPT's daily visits in the US have reduced significantly since the unveiling of DeepSeek, with the popular chatbot dropping from 22.1 million users in October 2024 to 14.9 million by January 2025. By comparison, DeepSeek had just 2,300 visits in October 2024 but sharply grew to 71,200 by January 19. As for globally, DeepSeek traffic has surpassed 7,120,000 million users and has ascended to the top of the charts for most downloaded apps across 140 countries.

ChatGPT Pro users can get their hands on OpenAI's second AI agent, following the release of Deep Research on February 2nd. Deep Research is an AI-powered tool that can search, crunch, and parse massive amounts of data, generating detailed, citation-backed reports from those insights.

Following the eruption of Deepseek, the move comes in a bid to fortify OpenAI's position as a market leader in the space. Their first AI agent, Operator, was released January 23rd, and their latest reasoning model, o3-mini, was made available February 1st.

As shown in a recent demo, using the tool is straightforward - select 'deep research' in ChatGPT, enter your query, and attach files if needed. From there, it browses, analyzes, and compiles information over 5-30 minutes, then delivers a thorough report with sources and key takeaways.

OpenAI's Operator was attached to some strong claims in the lead up to its January 23 launch. 'Ph.D. level intelligence', the ability to autonomously carry out coding tasks, and the potential to exceed human capabilities. However, early user experiences with the tool suggest the contrary.

The key distinction of AI Agents like Operator from a chatbot like ChatGPT is that they're designed to act on your behalf. Meaning: give them a task, and they'll take care of it with minimal oversight. Operator functions by essentially taking over your computer, utilizing a Computer-Using Agent (CUA) model that integrates visual processing and reasoning capabilities to interpret what's happening on the screen, and to carry out certain actions.

Bloomberg's Rachel Metz spent some time with Operator, taking it through various day-to-day tasks. Purchasing groceries, booking reservations, and filling out forms. The agent was able to successfully order lipstick from Sephora, fill out a cart for Ben & Jerry's ice cream, and even suggested adding additional items to qualify for free delivery. However, it fell short on simple tasks like filling out spreadsheets, managing calendars, and navigating unfamiliar web-pages. A common thread among users was that the agent required constant supervision - and was not particularly efficient even when it did succeed.

The unveiling of the DeepSeek R1 AI model has caused the AI industry to turn on its head, with accusations now flying around about the theft of data sets and the circumvention of US trade restrictions on NVIDIA hardware.

The latter has caused quite a stir, with the US government announcing it has launched a probe into China's potentially illegal use of NVIDIA AI GPUs that industry experts have suggested were used to train DeepSeek's AI. Government officials, and in particular, Howard Lutnick, President Donald Trump's pick to lead the Commerce Department, insinuated DeepSeek circumvented US trade policies by using intermediaries to acquire NVIDIA's hardware.

Notably, the US government has already banned the trade of these high-end NVIDIA GPUs in more than countries as it believes these countries could be used as intermediaries for China. However, Singapore isn't banned, and consists of approximately 20% of NVIDIA's revenue.

With the eruption of Deepseek in the last fortnight, the pace at which companies are moving to advance artificial intelligence is only continuing to increase. However, a key question remains - how are consumer businesses actually embracing the technology? A recent report from CPA Australia appears to shed some light.

The CPA Australia Business Technology Report 2024 examines technology adoption in businesses across the Asia-Pacific region. Conducted mid-last year, the survey was conducted with a few key questions on their mind: how is tech impacting business, what are the challenges businesses face with tech adoption, and what are the trends being observed with AI?

The results were interesting. As you'd expect, AI adoption surged from 55% in 2023 to 69% in 2024 - a significant jump. Half of all businesses also planned to increase their AI usage over the following year, citing the benefits of efficiency, productivity, and decision-making improvements. However, one of the glaring figures was that only 8% of businesses have fully integrated AI into their operations. In other words, while many are experimenting with AI tools, only a small fraction have made them a core part of their workflows.

When DeepSeek unveiled its R1 model the AI industry reeled as the company claimed it had developed an AI model that's on par with OpenAI's most-sophisticated model, but for a fraction of the cost.

But now the AI model has been out for some time, security researchers have been playing around with it and comparing it against the competition. In one set of testing, researchers from the University of Pennsylvania and hardware conglomerate Cisco pitted DeepSeek's AI against some "malicious" prompts, which are designed to bypass AI guidelines that are designed to prevent users from acquiring knowledge on how to, for example, make a bomb, generate misinformation, conduct cybercrime activities, etc.

Bypassing regulatory guidelines of a device typically called "jailbreaking," and in the instance of DeepSeek's AI, the researchers found it "failed to block a single harmful prompt." The R1 model was pitted against "50 random prompts from the HarmBench dataset," and the researchers were surprised to achieve a "100 percent attack success rate." According to the blog post, the researchers say the R1 model test results contrast starkly against other established AI models from OpenAI, Google, and Microsoft.

Google continues to enhance spreadsheet capabilities with its latest Gemini upgrade for Google Sheets. The update introduces a range of AI-powered improvements, focusing on data analysis, visualization, and interpretation.

As announced on January 29th in Google's Workspace Updates blog, the features continue to build upon an existing repertoire of Gemini AI integrations with Google's apps and collaboration tools. The additions include the ability to generate charts, trends, and insights from your data and the ability to ask questions and analyze data using Python or formulas.

One key benefit of these features is the ability to derive insights and visualizations without manual data crunching. For a marketing manager, this would mean asking the tool, "What are my top-performing channels by conversion rate?" From there, Gemini would break down trends, demonstrate which platforms drive the most ROI, and visualize them with bar graphs.

A research firm has found after conducting an extensive analysis that DeepSeek spent much more to train its R1 model than the touted $5 million.

Research firm SemiAnalysis looked at the new AI model that shook up the tech industry when it was announced, as the Chinese creators, DeepSeek, stated it only cost $6 million to create a model that was on par with OpenAI's most advanced o1 model. The announcement of the cost of DeepSeek's R1 model threw the tech space into a tailspin, wiping approximately $1 trillion away from the stock market and $600 billion from NVIDIA. For comparison, OpenAI's "GPT-4 model" is estimated to have cost anywhere between $100 - $200 million to train.

After DeepSeek's announcement, the market sentiment was the cost of training new AI models was completely overvalued, but now, according to SemiAnalysis, DeepSeek's R1 model cost the company a lot more than $5 million. The report states that DeepSeek purchased 10,000 units of NVIDIA's A100 GPUs in 2021, then acquired 10,000 "China-specific" NVIDIA H8000 AI GPUs and an additional 10,000 NVIDIA H100 GPUs. All of these GPUs, and training costs translates to $1.6 billion in capital expenditures for DeepSeek, with an estimated running cost of approximately $944 million.

Samsung Electronics has posted its Q4 2024 earnings, pulling in $54.18 billion in revenue, with an operating profit of $4.65 billion, with some new details going on with its troubled HBM memory.

Samsung says that it started supplying 8-Hi and 12-Hi HBM3E memory in Q3 2024, and that revenue from HBM3E has surpassed HBM3. The South Korean giant is still working on an "optimized version" of HBM3E that will ship in Q1 2025, while NVIDIA CEO Jensen Huang has said Samsung needed to redesign its HBM3E memory and the "optimized version" is probably just that, redesigned HBM3E.

The company says that HBM3E might see temporary demand and price drops in Q1 2025, as AI clients wait for the new "optimized version" instead of taking delivery of the older one. Insider Dan Nystedt notes that "no other major memory chip maker uses garbage terms like 'optimized version' for HBM products'

OpenAI has just released its new o3-mini, its newest, most cost-efficient model in its reasoning series, to both ChatGPT and API today... just as DeepSeek has come in changing the AI landscape.

In a new post on its website, the ChatGPT creator says that its new o3-mini is its most powerful, and fastest model that "advances the boundaries of what small models can achieve, delivering exceptional STEM capabilities -- with particular strenth in science, math, and coding -- all while maintaining the low cost and reduced latency of OpenAI o1-mini".

OpenAI explains on its website: "OpenAI o3-mini is our first small reasoning model that supports highly requested developer features including function calling⁠(opens in a new window), Structured Outputs⁠(opens in a new window), and developer messages⁠(opens in a new window), making it production-ready out of the gate. Like OpenAI o1-mini and OpenAI o1-preview, o3-mini will support streaming⁠(opens in a new window). Also, developers can choose between three reasoning effort⁠(opens in a new window) options-low, medium, and high-to optimize for their specific use cases".

NVIDIA CEO Jensen Huang will reportedly meet with President Trump at the White House today, to discuss all things about the United States and continued dominance in the AI industry, flying into the face of China and its game-changing DeepSeek R1 model.

In a new report from the Financial Times, we're learning that "NVIDIA chief executive Jensen Huang is set to meet US President Donald Trump at the White House on Friday, according to a person familiar with the matter. The person said the meeting had been planned before Chinese artificial intelligence start-up DeepSeek stunned Silicon Valley this week with advances apparently achieved using less computing power than US tech groups".

We should expect the two to discuss how the US government and freshly-minted Trump administration will frame its policies around AI, to ensure that the United States continues with its AI dominance. The entry of DeepSeek into the market has changed the game, wiping a huge $2 trillion in market capitalization in the process.

Samsung has reportedly received approval to supply NVIDIA with its HBM memory chips, with its 8-layer HBM3E destined for NVIDIA after rumors that NVIDIA CEO Jensen Huang said he can't trust Samsung's HBM products or engineers, and wouldn't do business with them.

In a new report from Bloomberg, we're learning that Samsung's less-advanced 8-layer HBM3E -- and not the newer 12-layer HBM3E -- was "cleared by NVIDIA in December" according to "people who asked not to be named as the information is private". These new Samsung HBM3E memory chips would be used in NVIDIA's less powerful AI GPUs destined for China.

Meanwhile, fellow South Korean memory company SK hynix has been NVIDIA's main partner for HBM memory, providing the company with bleeding-edge 12-layer HBM3E, and being asked by NVIDIA CEO to pull up the release of its next-gen HBM4 memory that will find its way into the company's next-gen Rubin R100 AI GPU.

DeepSeek has come in like a wrecking ball in the AI industry, knocking down OpenAI and its leading ChatGPT, and now Apple CEO Tim Cook has had some comments about the Chinese company.

In Apple's recent investors call, Apple CEO Tim Cook said the disruptive new AI model is a positive development for the iPhone giant. During the call, analyst Ben Reitzes from Melius asked Cook about how the arrival of DeepSeek is viewed by Apple. DeepSeek is far cheaper in development costs for training advanced AI models, as well as the running costs of these AI models.

Reitzes asked: "Hey, Tim. I wanted to ask you, you knew this one was coming, but there's a perception that you're a big beneficiary of lower cost compute, and I was wondering of your worldly perspective here on the deep situation, and if you are going to, if anything's happened to change your views in terms of the tailwind to margin and your ability to execute, even due to the potential for cost to come down due to that development and probably what's going to happen that happened anyway, but I love your perspective on that".

Intel has nixxed its upcoming Falcon Shores AI accelerator, shifting its efforts into the next-generation Jaguar Shores as a complete rack-scale solution in its attempt to make a mark on the AI industry.

Intel co-CEO Michelle Johnston Holthaus has explained that the company has stopped development on Falcon Shores, shifting its focus into its next-gen Jaguar Shores generation. Until now, we were expecting the Falcon Shores XPU on TSMC's newer 3nm process node and CoWoS-R advanced packaging, with a huge 1500W TDP.

Intel told CRN: "Many of you heard me temper expectations on Falcon Shores last month. Based on industry feedback, we have decided to leverage Falcon Shores as an internal test chip".

DeepSeek unveiling its R1 model has completely shaken up the AI industry as the company claims its model is comparable to OpenAI's most advanced AI model, which took approximately $100 million to create. However, DeepSeek says it only spent $6 million to make R1.

The disparity between the development costs of OpenAI's o3 model and DeepSeeks R1 model put the industry in a tailspin, causing approximately $1 trillion to be wiped away from AI companies, particularly NVIDIA, the company providing the horsepower necessary to train these models. A part of the reason NVIDIA took such a stock hit was DeepSeek, saying it trained its R1 model using H800 GPUs, which were released in 2023.

The sentiment is - why pay for new and expensive NVIDIA hardware when the same or comparable results to what is considered the best AI model can be generated using hardware from 2023?

OpenAI released its o1 model in December last year, and while OpenAI touted the model as the company's most advanced AI model yet, it does come with a hefty price tag if users want full access.

OpenAI is currently selling access to its o1 reasoning model under its ChatGPT Pro subscription, which comes in at $200 a month. Selecting this option grants the user unlimited access to the latest AI model, which differentiates itself from other AI models by its "reasoning" factor.

The o1 AI model is designed to be more thoughtful of the question a user is asking and implements a step-by-step chain-of-thought process, which makes it excel in specific tasks and categories of questioning. GPT-4 is a more general-purpose model for text generation, answering simple questions, or holding conversations.

The Trump administration is looking at increased sanctions on the sale of NVIDIA AI chips to China, according to "people familiar with the matter" in a new report from Bloomberg.

US officials are looking to expand restrictions to cover NVIDIA's H20 AI GPUs, which can be used to develop and run AI software and services, acting as a scaled-down AI GPU that was designed to meet US export restrictions, and allowed onto the shores of China.

Bloomberg reports that the "people added that a decision on any restrictions is likely a long ways off, given that the Trump administration is only beginning to staff up in relevant departments. Commerce Secretary nominee Howard Lutnick, Trump's pick to lead the agency that oversees chip trade curbs, said during a confirmation hearing Wednesday that he would be "very strong" on semiconductor controls, without providing more specifics".

Chinese AI company DeepSeek has been taking the headlines for nearly a week now, with US competitor OpenAI announcing ChatGPT... a new tailored version of ChatGPT designed to provide US government agencies with more ways to use OpenAI's frontier AI models.

OpenAI explained on its website that agencies can "deploy ChatGPT Gov in their own Microsoft Azure commercial cloud or Azure Government cloud on top of Microsoft's Azure's OpenAI Service. Self-hosting ChatGPT Gov enables agencies to more easily manage their own security, privacy, and compliance requirements, such as stringent cybersecurity frameworks (IL5, CJIS, ITAR, FedRAMP High)".

"Additionally, we believe this infrastructure will expedite internal authorization of OpenAI's tools for the handling of non-public sensitive data. Use of ChatGPT Gov is subject to our usage policies, like other OpenAI services".

AMD has provided some instructions on how to run DeepSeek's exciting new R1 distilled "reasoning" model on AMD Ryzen AI and Radeon products... yeah, you can run a local DeepSeek R1 model on your PC at home.

AMD explains: "Reasoning models add a "thinking" stage before the final output - which you can see by expanding the "thinking" window before the model gives its final answer. Unlike conventional LLMs, which one-shot the response, CoT LLMs perform extensive reasoning before answering. The assumptions and self-reflection the LLM performs are visible to the user and this improves the reasoning and analytical capability of the model - albeit at the cost of significantly longer time-to-first-(final output)token".

"A reasoning model may first spend thousands of tokens (and you can view this chain of thought!) to analyze the problem before giving a final response. This allows the model to be excellent at complex problem-solving tasks involving math and science and attack a complex problem from all angles before deciding on a response. Depending on your AMD hardware, each of these models will offer state-of-the-art reasoning capability on your AMD Ryzen™ AI processor or Radeon™ graphics cards".