Latest news with #HuggingFace
Yahoo
10 hours ago
- Science
- Yahoo
AI is more likely to create a generation of ‘yes-men on servers' than any scientific breakthroughs, Hugging Face co-founder says
Hugging Face's co-founder, Thomas Wolf, is pouring cold water on the hopes that current AI systems could revolutionize scientific progress. Speaking to Fortune at VivaTech in Paris, Wolf argued that today's large language models excel at producing plausible answers but lack the creativity to ask original scientific questions. Rather than building the next Einstein, Wolf says we may be creating a generation of digital 'yes-men.' Hugging Face's top scientist, Thomas Wolf, says current AI systems are unlikely to make the scientific discoveries some leading labs are hoping for. Speaking to Fortune at Viva Technology in Paris, the Hugging Face co-founder said that while large language models (LLMs) have shown an impressive ability to find answers to questions, they fall short when trying to ask the right ones—something Wolf sees as the more complex part of true scientific progress. 'In science, asking the question is the hard part, it's not finding the answer,' Wolf said. 'Once the question is asked, often the answer is quite obvious, but the tough part is really asking the question, and models are very bad at asking great questions.' Wolf said he came to the conclusion after reading a widely circulated blog post by Anthropic CEO Dario Amodei called Machines of Loving Grace. In it, Amodei argues the world is about to see the 21st century 'compressed' into a few years as AI accelerates science drastically. Wolf said he initially found the piece inspiring but started to doubt Amodei's idealistic vision of the future after the second read. 'It was saying AI is going to solve cancer and it's going to solve mental health problems — it's going to even bring peace into the world, but then I read it again and realized there's something that sounds very wrong about it, and I don't believe that,' he said. For Wolf, the problem isn't that AI lacks knowledge but that it lacks the ability to challenge our existing frame of knowledge. AI models are trained to predict likely continuations, for example, the next word in a sentence, and while today's models excel at mimicking human reasoning, they fall short of any real original thinking. 'Models are just trying to predict the most likely thing,' Wolf explained. 'But in almost all big cases of discovery or art, it's not really the most likely art piece you want to see, but it's the most interesting one.' Using the example of the game of Go, a board game that became a milestone in AI history when DeepMind's AlphaGo defeated world champions in 2016, Wolf argued that while mastering the rules of Go is impressive, the bigger challenge lies in inventing such a complex game in the first place. In science, he said, the equivalent of inventing the game is asking these truly original questions. Wolf first suggested this idea in a blog post titled The Einstein AI Model, published earlier this year. In it, he wrote: 'To create an Einstein in a data center, we don't just need a system that knows all the answers, but rather one that can ask questions nobody else has thought of or dared to ask.' He argues that what we have instead are models that behave like 'yes-men on servers'—endlessly agreeable, but unlikely to challenge assumptions or rethink foundational ideas. This story was originally featured on
WIRED
2 days ago
- Business
- WIRED
How Much Energy Does AI Use? The People Who Know Aren't Saying
Jun 19, 2025 6:00 AM A growing body of research attempts to put a number on energy use and AI—even as the companies behind the most popular models keep their carbon emissions a secret. Photograph: Bloomberg/Getty Images 'People are often curious about how much energy a ChatGPT query uses,' Sam Altman, the CEO of OpenAI, wrote in an aside in a long blog post last week. The average query, Altman wrote, uses 0.34 watt-hours of energy: 'About what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes.' For a company with 800 million weekly active users (and growing), the question of how much energy all these searches are using is becoming an increasingly pressing one. But experts say Altman's figure doesn't mean much without much more public context from OpenAI about how it arrived at this calculation—including the definition of what an 'average' query is, whether or not it includes image generation, and whether or not Altman is including additional energy use, like from training AI models and cooling OpenAI's servers. As a result, Sasha Luccioni, the climate lead at AI company Hugging Face, doesn't put too much stock in Altman's number. 'He could have pulled that out of his ass,' she says. (OpenAI did not respond to a request for more information about how it arrived at this number.) As AI takes over our lives, it's also promising to transform our energy systems, supercharging carbon emissions right as we're trying to fight climate change. Now, a new and growing body of research is attempting to put hard numbers on just how much carbon we're actually emitting with all of our AI use. This effort is complicated by the fact that major players like OpenAi disclose little environmental information. An analysis submitted for peer review this week by Luccioni and three other authors looks at the need for more environmental transparency in AI models. In Luccioni's new analysis, she and her colleagues use data from OpenRouter, a leaderboard of large language model (LLM) traffic, to find that 84 percent of LLM use in May 2025 was for models with zero environmental disclosure. That means that consumers are overwhelmingly choosing models with completely unknown environmental impacts. 'It blows my mind that you can buy a car and know how many miles per gallon it consumes, yet we use all these AI tools every day and we have absolutely no efficiency metrics, emissions factors, nothing,' Luccioni says. 'It's not mandated, it's not regulatory. Given where we are with the climate crisis, it should be top of the agenda for regulators everywhere.' As a result of this lack of transparency, Luccioni says, the public is being exposed to estimates that make no sense but which are taken as gospel. You may have heard, for instance, that the average ChatGPT request takes 10 times as much energy as the average Google search. Luccioni and her colleagues track down this claim to a public remark that John Hennessy, the chairman of Alphabet, the parent company of Google, made in 2023. A claim made by a board member from one company (Google) about the product of another company to which he has no relation (OpenAI) is tenuous at best—yet, Luccioni's analysis finds, this figure has been repeated again and again in press and policy reports. (As I was writing this piece, I got a pitch with this exact statistic.) 'People have taken an off-the-cuff remark and turned it into an actual statistic that's informing policy and the way people look at these things,' Luccioni says. 'The real core issue is that we have no numbers. So even the back-of-the-napkin calculations that people can find, they tend to take them as the gold standard, but that's not the case.' One way to try and take a peek behind the curtain for more accurate information is to work with open source models. Some tech giants, including OpenAI and Anthropic, keep their models proprietary—meaning outside researchers can't independently verify their energy use. But other companies make some parts of their models publicly available, allowing researchers to more accurately gauge their emissions. A study published Thursday in the journal Frontiers of Communication evaluated 14 open-source large language models, including two Meta Llama models and three DeepSeek models, and found that some used as much as 50 percent more energy than other models in the dataset responding to prompts from the researchers. The 1,000 benchmark prompts submitted to the LLMs included questions on topics such as high school history and philosophy; half of the questions were formatted as multiple choice, with only one-word answers available, while half were submitted as open prompts, allowing for a freer format and longer answers. Reasoning models, the researchers found, generated far more thinking tokens—measures of internal reasoning generated in the model while producing its answer, which are a hallmark of more energy use—than more concise models. These models, perhaps unsurprisingly, were also more accurate with complex topics. (They also had trouble with brevity: During the multiple choice phase, for instance, the more complex models would often return answers with multiple tokens, despite explicit instructions to only answer from the range of options provided.) Maximilian Dauner, a PhD student at the Munich University of Applied Sciences and the study's lead author, says he hopes AI use will evolve to think about how to more efficiently use less-energy-intensive models for different queries. He envisions a process where smaller, simpler questions are automatically directed to less-energy-intensive models that will still provide accurate answers. 'Even smaller models can achieve really good results on simpler tasks, and don't have that huge amount of CO 2 emitted during the process,' he says. Some tech companies already do this. Google and Microsoft have previously told WIRED that their search features use smaller models when possible, which can also mean faster responses for users. But generally, model providers have done little to nudge users toward using less energy. How quickly a model answers a question, for instance, has a big impact on its energy use—but that's not explained when AI products are presented to users, says Noman Bashir, the Computing & Climate Impact Fellow at MIT's Climate and Sustainability Consortium. 'The goal is to provide all of this inference the quickest way possible so that you don't leave their platform,' he says. 'If ChatGPT suddenly starts giving you a response after five minutes, you will go to some other tool that is giving you an immediate response.' However, there's a myriad of other considerations to take into account when calculating the energy use of complex AI queries, because it's not just theoretical—the conditions under which queries are actually run out in the real world matter. Bashir points out that physical hardware makes a difference when calculating emissions. Dauner ran his experiments on an Nvidia A100 GPU, but Nvidia's H100 GPU—which was specially designed for AI workloads, and which, according to the company, is becoming increasingly popular—is much more energy-intensive. Physical infrastructure also makes a difference when talking about emissions. Large data centers need cooling systems, light, and networking equipment, which all add on more energy; they often run in diurnal cycles, taking a break at night when queries are lower. They are also hooked up to different types of grids—ones overwhelmingly powered by fossil fuels, versus those powered by renewables—depending on their locations. Bashir compares studies that look at emissions from AI queries without factoring in data center needs to lifting up a car, hitting the gas, and counting revolutions of a wheel as a way of doing a fuel-efficiency test. 'You're not taking into account the fact that this wheel has to carry the car and the passenger,' he says. Perhaps most crucially for our understanding of AI's emissions, open source models like the ones Dauner used in his study represent a fraction of the AI models used by consumers today. Training a model and updating deployed models takes a massive amount of energy—figures that many big companies keep secret. It's unclear, for example, whether the light bulb statistic about ChatGPT from OpenAI's Altman takes into account all the energy used to train the models powering the chatbot. Without more disclosure, the public is simply missing much of the information needed to start understanding just how much this technology is impacting the planet. 'If I had a magic wand, I would make it mandatory for any company putting an AI system into production, anywhere, around the world, in any application, to disclose carbon numbers,' Luccioni says. Paresh Dave contributed reporting.
Fox News
4 days ago
- Fox News
New robots make AI something anyone can try at home
Hugging Face, a well-known name in AI development, is making big moves in the world of robotics. The company has just introduced two open-source humanoid robots, HopeJR and Reachy Mini, designed to make advanced robotics more accessible to everyone, from researchers and developers to students and hobbyists. Sign up for my FREE CyberGuy ReportGet my best tech tips, urgent security alerts, and exclusive deals delivered straight to your inbox. Plus, you'll get instant access to my Ultimate Scam Survival Guide free when you join. HopeJR is Hugging Face's new full-size humanoid robot. It stands out for its impressive 66 actuated degrees of freedom. This means it can walk, move its arms, and perform a wide range of independent gestures, making it a versatile platform for research, experimentation, and even household tasks in the future. Imagine a robot that can help with chores or serve as a hands-on learning tool. HopeJR is designed to be just that. One of the biggest talking points is price. Hugging Face aims to keep HopeJR affordable, with an estimated cost of around $3,000. That is a fraction of what most full-scale humanoid robots cost, opening the door for smaller labs, schools, and dedicated enthusiasts to get involved in advanced robotics development. If you are looking for something more compact, Reachy Mini is Hugging Face's answer. This desktop robot stands about 11 inches tall and is packed with features, including a camera, microphone, speaker, and a full six-degree-of-freedom neck for expressive movement. It can move its head, listen, speak, and interact with AI applications, making it a fun and practical tool for developers and educators alike. Reachy Mini is expected to cost between $250 and $300. It is designed to be a hands-on gateway into robotics, letting users build, customize, and experiment with embodied AI. Plus, it integrates seamlessly with Hugging Face Spaces, giving access to over 500,000 AI apps on the Hugging Face Hub. Both HopeJR and Reachy Mini are fully open source. This means anyone can assemble, rebuild, and understand how these robots work. Hugging Face's CEO, Clem Delangue, emphasized that this approach helps keep robotics open and accessible rather than dominated by a handful of companies with proprietary black-box systems. The community-driven model encourages collaboration and innovation, allowing users to share their improvements and build on each other's work. Hugging Face has not set an exact shipping date for these robots, but the company expects to deliver the first units by the end of the year. There is already a wait list open for those interested in getting their hands on HopeJR or Reachy Mini. This expansion into robotics builds on Hugging Face's recent acquisition of Pollen Robotics, the creators of the original Reachy robot. That partnership gave Hugging Face the expertise needed to accelerate hardware development and bring these new robots to market quickly. If you have ever wanted to dive into robotics or just see what AI can do in the real world, now is a great time to jump in. Hugging Face is making it easier than ever for anyone to experiment, build, and learn with robots that are open, affordable, and packed with potential. Whether you are a student, a developer, or just someone who loves to tinker, HopeJR and Reachy Mini could be your ticket to hands-on experience with the future of AI. What would you create or change in your daily life if you had your own open-source humanoid robot at your fingertips? Let us know by writing us at For more of my tech tips and security alerts, subscribe to my free CyberGuy Report Newsletter by heading to Follow Kurt on his social channels: Answers to the most-asked CyberGuy questions: New from Kurt: Copyright 2025 All rights reserved.
Yahoo
12-06-2025
- Business
- Yahoo
Nvidia collaborates for sovereign LLMs
Nvidia has partnered with various model builders and cloud service providers across Europe and the Middle East to enhance the development of sovereign large language models (LLMs). The collaboration aims to accelerate AI adoption in industries such as manufacturing, robotics, healthcare, finance, energy, and creative sectors. Key partners include Barcelona Supercomputing Center (BSC), Dicta, H Company and Domyn. Other key platers include LightOn, the National Academic Infrastructure for Supercomputing in Sweden (NAISS), KBLab at the National Library of Sweden, the Technology Innovation Institute (TII), University College London plus the University of Ljubljana, and UTTER. These partners are using Nvidia Nemotron techniques to enhance their models, focusing on cost efficiency and accuracy for enterprise AI workloads, including agentic AI. The models support Europe's 24 official languages and reflect local languages and cultures, Nvidia said. Some models, developed by H Company and LightOn in France, Dicta in Israel, Domyn in Italy, in Poland, BSC in Spain, NAISS and KBLab in Sweden, TII in the UAE, and University College London in the UK, specialise in national language and culture. The optimised models will run on AI infrastructure from Nvidia Cloud Partners (NCPs) like Nebius, Nscale, and Fluidstack through the Nvidia DGX Cloud Lepton marketplace. The LLMs will be distilled using Nvidia Nemotron techniques, including neural architecture search, reinforcement learning, and post-training with Nvidia-curated synthetic data. These processes aim to reduce operational costs and improve token generation speed during inference. Developers can deploy these models as Nvidia NIM microservices on AI factories, both on-premises and across cloud platforms, supporting more than 100,000 LLMs hosted on Hugging Face. A new Hugging Face integration with DGX Cloud Lepton will allow companies to fine-tune models on local NCP infrastructure. Perplexity, an AI-powered answer engine processing over 150 million questions weekly, will integrate these models to enhance search query accuracy and AI outputs. Nvidia founder and CEO Jensen Huang said: 'Together with Europe's model builders and cloud providers, we're building an AI ecosystem where intelligence is developed and served locally to provide a foundation for Europe to thrive in the age of AI — transforming every industry across the region.' Recently, Nvidia announced multiple partnerships in the UK to boost AI capabilities, aligning with the start of London Tech Week. "Nvidia collaborates for sovereign LLMs" was originally created and published by Verdict, a GlobalData owned brand. The information on this site has been included in good faith for general informational purposes only. It is not intended to amount to advice on which you should rely, and we give no representation, warranty or guarantee, whether express or implied as to its accuracy or completeness. You must obtain professional or specialist advice before taking, or refraining from, any action on the basis of the content on our site. Sign in to access your portfolio
Entrepreneur
11-06-2025
- Business
- Entrepreneur
Mistral Launches Magistral to Compete in the Reasoning AI Race
While Magistral puts Mistral in closer competition with well-known reasoning AI models, there are still doubts across the industry about how well current LLMs can actually "reason" Opinions expressed by Entrepreneur contributors are their own. You're reading Entrepreneur India, an international franchise of Entrepreneur Media. French artificial intelligence firm Mistral has announced the release of its latest large language model (LLM), Magistral, marking its entry into the growing space of "reasoning" AI models. The new model aims to improve the transparency and traceability of AI-generated outputs, particularly in tasks that require step-by-step logical processing. Unveiled on Tuesday during London Tech Week, Magistral is available through Mistral's platforms and the open-source AI repository Hugging Face. The company has released two versions of the model: Magistral Small, a 24-billion-parameter model licensed as open-source, and a more powerful, proprietary version, Magistral Medium, currently available in limited preview. Mistral describes Magistral as suitable for general-purpose use cases that involve more complex reasoning and demand greater accuracy. The model is designed to provide a visible "chain of thought," which the company says helps users understand how conclusions are reached. This feature may appeal to professionals in law, healthcare, finance, and public services where regulatory compliance and interpretability are key concerns. According to CEO Arthur Mensch, a key distinction of Magistral is its multilingual reasoning capability, especially in European languages. "Historically, we've seen U.S. models reason in English and Chinese models reason in Chinese," he said during a session at London Tech Week. Mensch noted that Magistral is initially focused on European languages, with plans to expand support to other languages over time. The launch comes as more AI companies shift their focus from building larger models to improving how existing models process and present information. Reasoning models are designed to handle more sophisticated tasks by simulating logical steps, rather than generating answers based solely on pattern recognition. This shift also responds to ongoing concerns about the interpretability of AI systems, which often function as black boxes even to their creators. Mistral claims that Magistral Medium can process up to 1,000 tokens per second, potentially offering faster performance than several competing models. It joins a growing list of reasoning-focused models released over the past year, including OpenAI's o1 and o3, Google's Gemini variants, Anthropic's Claude, and DeepSeek's R1. The release also highlights Mistral's continuing emphasis on open-source AI development. The company, founded in Paris in 2023, has received significant backing from investors including Microsoft, DST Global, and General Catalyst. It raised approximately USD 685.7 million million in a Series B round in June 2024, bringing total funding to over USD 1.37 billion and reaching a reported valuation of USD 6.63 billion. Despite its relatively short history, Mistral has seen considerable commercial traction. As per the media reports, the company has secured over USD 114.3 million in contracted sales within 15 months of launching its first commercial offerings. While Magistral puts Mistral in closer competition with well-known reasoning AI models, there are still doubts across the industry about how well current large language models (LLMs) can actually "reason." A recent research paper from Apple, titled The Illusion of Thinking, questions the belief that today's models truly have general reasoning abilities. The researchers found that these models tend to struggle or fail when tasks become too complex, revealing key limitations in their capabilities.
