Latest news with #reasoning
TechCrunch
3 days ago
- Business
- TechCrunch
DeepSeek's distilled new R1 AI model can run on a single GPU
DeepSeek's updated R1 reasoning AI model might be getting the bulk of the AI community's attention this week. But the Chinese AI lab also released a smaller, 'distilled' version of its new R1, DeepSeek-R1-0528-Qwen3-8B, that DeepSeek claims beats comparably-sized models on certain benchmarks. The smaller updated R1, which was built using the Qwen3-8B model Alibaba launched in May as a foundation, performs better than Google's Gemini 2.5 Flash on AIME 2025, a collection of challenging math questions. DeepSeek-R1-0528-Qwen3-8B also nearly matches Microsoft's recently released Phi 4 reasoning plus model on another math skills test, HMMT. So-called distilled models like DeepSeek-R1-0528-Qwen3-8B are generally less capable than their full-sized counterparts. On the plus side, they're far less computationally demanding. According to the cloud platform NodeShift, Qwen3-8B requires a GPU with 40GB-80GB of RAM to run (e.g., an Nvidia H100). The full-sized new R1 needs around a dozen 80GB GPUs. DeepSeek trained DeepSeek-R1-0528-Qwen3-8B by taking text generated by the updated R1 and using it to fine-tune Qwen3-8B. In a dedicated webpage for the model on the AI dev platform Hugging Face, DeepSeek describes DeepSeek-R1-0528-Qwen3-8B as 'for both academic research on reasoning models and industrial development focused on small-scale models.' DeepSeek-R1-0528-Qwen3-8B is available under a permissive MIT license, meaning it can be used commercially without restriction. Several hosts, including LM Studio, already offer the model through an API.
TechCrunch
12-05-2025
- Business
- TechCrunch
Improvements in ‘reasoning' AI models may slow down soon, analysis finds
An analysis by Epoch AI, a nonprofit AI research institute, suggests the AI industry may not be able to eke massive performance gains out of reasoning AI models for much longer. As soon as within a year, progress from reasoning models could slow down, according to the report's findings. Reasoning models such as OpenAI's o3 have led to substantial gains on AI benchmarks in recent months, particularly benchmarks measuring math and programming skills. The models can apply more computing to problems, which can improve their performance, with the downside being that they take longer than conventional models to complete tasks. Reasoning models are developed by first training a conventional model on a massive amount of data, then applying a technique called reinforcement learning, which effectively gives the model 'feedback' on its solutions to difficult problems. So far, frontier AI labs like OpenAI haven't applied an enormous amount of computing power to the reinforcement learning stage of reasoning model training, according to Epoch. That's changing. OpenAI has said that it applied around 10x more computing to train o3 than its predecessor, o1, and Epoch speculates that most of this computing was devoted to reinforcement learning. And OpenAI researcher Dan Roberts recently revealed that the company's future plans call for prioritizing reinforcement learning to use far more computing power, even more than for the initial model training. But there's still an upper bound to how much computing can be applied to reinforcement learning, per Epoch. According to an Epoch AI analysis, reasoning model training scaling may slow down. Image Credits:Epoch AI Josh You, an analyst at Epoch and the author of the analysis, explains that performance gains from standard AI model training are currently quadrupling every year, while performance gains from reinforcement learning are growing tenfold every 3-5 months. The progress of reasoning training will 'probably converge with the overall frontier by 2026,' he continues. Techcrunch event Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you've built — without the big spend. Available through May 9 or while tables last. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you've built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | BOOK NOW Epoch's analysis makes a number of assumptions, and draws in part on public comments from AI company executives. But it also makes the case that scaling reasoning models may prove to be challenging for reasons besides computing, including high overhead costs for research. 'If there's a persistent overhead cost required for research, reasoning models might not scale as far as expected,' writes You. 'Rapid compute scaling is potentially a very important ingredient in reasoning model progress, so it's worth tracking this closely.' Any indication that reasoning models may reach some sort of limit in the near future is likely to worry the AI industry, which has invested enormous resources developing these types of models. Already, studies have shown that reasoning models, which can be incredibly expensive to run, have serious flaws, like a tendency to hallucinate more than certain conventional models.
The Sun
11-05-2025
- Entertainment
- The Sun
The 1% Club players struggle with tough letters question – but could you work out the correct combination?
THE 1% Club players have struggled with a tough letters question - but could you work out the correct combination? Lee Mack, 56, took to the ITV airwaves once again to present another edition of the smash-hit gameshow. 4 4 The 1% Club is a unique format that doesn't test players on their general knowledge like other shows. Instead, it tests them on their logic, reasoning skills and common sense. 100 players are whittled down, question by question, as they are tasked with solving different riddles. They aim to get to the last round where only one percent of the public could answer the final question correctly. The letters on the screen were TI, Man and Cas. After Lee revealed that four people had been knocked out he gave the answer- don't read on if you don't want to know yet. Lee explained that the answer was tle which made the words Tile, Mantle and Castle. He then turned to a player called Rachel who was still in the game and asked: "What would you do with the money if you won?" "I would go to the Galapagos" she replied. The 1% Club players struggle with tough letters question - but could you work out the correct combination- Rachel added that she would take her friend with her because when they were younger they didn't get there because they ran out of money. "Brilliant, Good Luck tonight," responded Lee. Meanwhile, in the show's latest edition, Lee showed the contestants three images that appeared to be a rope, a rocket and some test tubes. He then asked them: "What common three word phrase is represented here?" The 1% Club's Most Difficult Questions The 1% Club sees 100 contestants try and make it to the 1% question and be in with a chance to win a share of the jackpot. Here are just some of the show's most difficult teasers. Players had to compare and contrast three images of butterflies then explain which of the butterflies were exactly the same on both sides. Find the image and answer here. Players were shown groups of six symbols then asked which were in the same order whether you read them from left to right or right to left. Find the image and answer here. Players were asked how many different combinations were there of displaying four digits on one hand. Find the answer here. Peter had recently found his old diary that he'd written in secret code but he couldn't remember how to decipher what he wrote. Players were asked to crack the code and find out what the bold word was. WH89 I GR1W UP I WA92 21 B8 A 5L1RI72. Find the image and answer here. Players were tasked with working out how many eyes they could see in an image, which was made up of letters, symbols and emojis. Find the image and the answer here. A 1% question was based on a grid of numbers going in ascending order from 1 to 49. Starting on 25, the middle square, SEEN took you to square 27. From there, NEW took you to 20. From there, which square would SEWN take you to? Find the image and the answer here. And finally, an easy one - What common food in bold has had its letters rearranged into alphabetical order? ABDER If you really don't know you can find the answer here. As the contestants put their thinking caps on, he quipped: "Looks like a selection of ways my wife has thought to get rid of me." He then revealed that four people had been knocked out and went on to reveal the answer which you can find here. 4 4
Geeky Gadgets
08-05-2025
- Geeky Gadgets
Microsoft Phi-4 AI Models Offer Advanced AI Reasoning for Your Daily Life
What if your next email assistant could not only summarize your inbox but also reason through your schedule conflicts, all without needing an internet connection? Microsoft's latest leap in artificial intelligence, the Phi-4 Reasoning series, promises to make this a reality. With models like Phi-4 Reasoning, Phi-4 Reasoning Plus, and the ultra-compact Phi-4 Mini Reasoning, the tech giant is setting its sights on redefining how AI handles complex reasoning tasks. Unlike traditional AI systems that rely on sheer scale, these models prioritize efficiency and adaptability, making them accessible for everyday devices while maintaining innovative performance. This bold move signals Microsoft's intent to lead the charge in transforming AI from a tool of convenience into a cornerstone of innovation. In this exploration of Microsoft's new reasoning models, you'll uncover how these systems are trained to think critically, why their compact design is a fantastic option, and where they might soon show up in your daily life. From offline functionality in productivity tools like Outlook to on-device optimization for Windows, the Phi-4 Reasoning series is poised to make advanced AI more practical and private than ever before. But this isn't just about better tech—it's about reshaping the boundaries of what artificial intelligence can achieve. As Sam Witteveen delves into the details in the video below, one question looms large: could these models be the key to unlocking the next era of AI-driven innovation? Microsoft's Phi-4 Reasoning Models What Are the New Reasoning Models? The Phi-4 Reasoning series includes three distinct models, each tailored to meet specific reasoning needs: Phi-4 Reasoning: The flagship model, offering robust reasoning capabilities suitable for a wide range of applications. The flagship model, offering robust reasoning capabilities suitable for a wide range of applications. Phi-4 Reasoning Plus: An enhanced version that delivers improved accuracy and adaptability, ideal for more demanding and nuanced tasks. An enhanced version that delivers improved accuracy and adaptability, ideal for more demanding and nuanced tasks. Phi-4 Mini Reasoning: A compact model with only 3.88 billion parameters, designed to maximize efficiency while maintaining strong performance. These models are derived from larger systems such as GPT-4 and DeepSeek R1, inheriting their advanced reasoning capabilities while being optimized for computational efficiency. For example, the Phi-4 Mini Reasoning model demonstrates exceptional performance relative to its size, showcasing Microsoft's commitment to creating smaller, high-performing AI systems that can operate effectively even in resource-constrained environments. How Are These Models Trained? The development of the Phi-4 Reasoning series is underpinned by advanced training techniques that enhance their reasoning abilities while making sure they remain efficient and adaptable. Key methods include: Model Distillation: Smaller models are trained using synthetic datasets generated by larger, more complex systems. This process allows the smaller models to retain the advanced reasoning capabilities of their larger counterparts. Smaller models are trained using synthetic datasets generated by larger, more complex systems. This process allows the smaller models to retain the advanced reasoning capabilities of their larger counterparts. Supervised Fine-Tuning: Carefully curated datasets, particularly those focused on mathematical reasoning and logical problem-solving, are used to refine the models' accuracy and reliability. Carefully curated datasets, particularly those focused on mathematical reasoning and logical problem-solving, are used to refine the models' accuracy and reliability. Alignment Training: This ensures that the models produce outputs that align with user expectations and factual accuracy, improving their practical utility. This ensures that the models produce outputs that align with user expectations and factual accuracy, improving their practical utility. Reinforcement Learning with Verifiable Rewards (RLVR): A feedback-driven approach that rewards models for generating accurate, logical, and contextually appropriate outputs, further enhancing their reasoning skills. By combining these techniques, Microsoft has created models capable of handling complex reasoning tasks while maintaining a high degree of efficiency. This approach ensures that the models are not only powerful but also practical for real-world applications. Microsoft Joins the AI Reasoning Race Watch this video on YouTube. Advance your skills in AI reasoning models by reading more of our detailed content. Performance: How Do They Compare? The Phi-4 Mini Reasoning model exemplifies the balance between size and performance. Despite its smaller parameter count, it competes effectively with larger models such as Quen and DeepSeek. While Quen models are recognized for their compact size and strong reasoning capabilities, Microsoft's Phi-4 Mini Reasoning model offers a unique combination of efficiency and reasoning depth. Benchmarks indicate that smaller models like Phi-4 Mini Reasoning can deliver high-quality reasoning without the computational demands typically associated with larger systems. This demonstrates the potential of compact AI models to provide advanced functionality while reducing resource consumption, making them ideal for deployment in a variety of environments, including local devices. Where Will These Models Be Used? Microsoft envisions a broad range of applications for the Phi-4 Reasoning series across its ecosystem of products and services. Potential use cases include: Outlook and Copilot: Enhancing productivity tools with offline functionality for tasks such as scheduling, summarization, and data analysis, making sure seamless user experiences even without internet connectivity. Enhancing productivity tools with offline functionality for tasks such as scheduling, summarization, and data analysis, making sure seamless user experiences even without internet connectivity. Windows Devices: A specialized version, known as FI Silica, is being developed for local use. This version emphasizes offline and on-device optimization, allowing advanced reasoning capabilities without relying on external servers. By embedding these reasoning models directly into operating systems and applications, Microsoft aims to improve functionality while prioritizing data privacy and efficiency. This approach reduces reliance on external APIs, making sure that users can access advanced AI capabilities in a secure and resource-efficient manner. What's Next for Microsoft's Reasoning Models? Looking ahead, Microsoft is exploring how small reasoning models can contribute to the development of artificial general intelligence (AGI) and more efficient large language models (LLMs). These models are expected to adopt a hybrid approach, combining their reasoning capabilities with external tools for factual data retrieval. This strategy could lead to the creation of more versatile and efficient AI systems, capable of addressing a broader range of tasks while maintaining a focus on reasoning. Microsoft's vision for the future includes integrating these models into a wider array of technologies, paving the way for innovative advancements in AI-driven applications. By focusing on efficiency and adaptability, the Phi-4 Reasoning series could play a pivotal role in shaping the next generation of AI systems, making advanced reasoning an integral part of everyday technology. Media Credit: Sam Witteveen Filed Under: AI, Technology News, Top News Latest Geeky Gadgets Deals Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
The Sun
07-05-2025
- Entertainment
- The Sun
Picture question stumps The 1% Club players – but would you have got it in 30 seconds?
A PICTURE question has stumped The 1% Club players - but would you have got it in 30 seconds? Lee Mack, 56, took to the ITV airwaves once again to present the latest edition of the smash-hit gameshow. 4 4 4 The 1% Club is a unique format that doesn't test players on their general knowledge like other shows. Instead, it tests them on their logic, reasoning skills and common sense. 100 players are whittled down, question by question, as they are tasked with solving different riddles. They aim to get to the last round where only one percent of the public could answer the final question correctly. In the latest edition, Lee showed the contestants three images that appeared to be a rope, a rocket and some test tubes. He then asked them: "What common three word phrase is represented here?" As the contestants put their thinking caps on, he quipped: "Looks like a selection of ways my wife has thought to get rid of me." He then revealed that four people had been knocked out and went on to reveal the answer. "Not Rocket Science," he explained. He then spoke to Daisy, a contestant who used her pass during the round - because she didn't know the answer. 20 players ELIMINATED in brutal 1% Club picture question - would you have guessed correctly? Daisy explained: "I saw it just after the time as well, but yeah I couldn't put it together in that time." "Are you a competitive person?" he asked her. Kilimanjaro and I was like yeah that sounds like a great idea." "Did you do it?" asked Lee. "Yeah, I was the only one that got to the top, as well, cos I got that bad altitude sickness, I didn't know where I was. The 1% Club's Most Difficult Questions The 1% Club sees 100 contestants try and make it to the 1% question and be in with a chance to win a share of the jackpot. Here are just some of the show's most difficult teasers. Players had to compare and contrast three images of butterflies then explain which of the butterflies were exactly the same on both sides. Find the image and answer here. Players were shown groups of six symbols then asked which were in the same order whether you read them from left to right or right to left. Find the image and answer here. Players were asked how many different combinations were there of displaying four digits on one hand. Find the answer here. Peter had recently found his old diary that he'd written in secret code but he couldn't remember how to decipher what he wrote. Players were asked to crack the code and find out what the bold word was. WH89 I GR1W UP I WA92 21 B8 A 5L1RI72. Find the image and answer here. Players were tasked with working out how many eyes they could see in an image, which was made up of letters, symbols and emojis. Find the image and the answer here. A 1% question was based on a grid of numbers going in ascending order from 1 to 49. Starting on 25, the middle square, SEEN took you to square 27. From there, NEW took you to 20. From there, which square would SEWN take you to? Find the image and the answer here. And finally, an easy one - What common food in bold has had its letters rearranged into alphabetical order? ABDER If you really don't know you can find the answer here. "I didn't know my own name, nothing." she replied. Lee said: "You just wandered up there confused to the top?" "Yeah, I thought my guide was trying to kill me at one point, it was awful but yeah I went to the top," replied Daisy. "Well, listen well done Daisy, You're still with us," replied Lee before he went onto the next question. 4
