AI^2的动态

Smaller is better… While the race for larger models continues, Apple and Microsoft are leading a shift towards Small Language Models (SLMs) in the AI landscape. Key points: - Apple and Microsoft have introduced SLMs with 3-14 billion parameters, far fewer than traditional Large Language Models (LLMs). - Despite their size, these SLMs are matching or outperforming LLMs on various benchmarks. - The secret? High-quality, curated training data instead of just increasing parameter count. Why SLMs matter: - Energy efficiency: They consume less power and can run locally on devices. - Privacy: Local processing means better data protection. - Accessibility: Smaller models democratize AI, allowing more organizations to participate in development. - Research potential: SLMs could offer insights into human language acquisition and improve AI efficiency. OpenAI's CEO Sam Altman also suggests, we may be at the end of the era of giant models. The future of AI might just be small, efficient, and more accessible to all. https://lnkd.in/emsvxBRJ

Discover how tech giants like Apple, Google, and Microsoft are scaling down AI language models to make them more efficient and accessible. By integrating these compact models into everyday applications, they are leading the way in this transformative technology. This advancement not only impacts the tech ecosystem but also offers insights into how children learn to talk. For more details on the impact and prospects of this innovation, check out the article on IEEE Spectrum: https://lnkd.in/ey3SPFwj #AI #Tech #Innovation #ArtificialIntelligence #TechGiants #FutureProspects

??Things are moving so fast in the world of #ai. Introducing SLMs ! ???????? Small Langauage Models “Tech giants are shifting focus from large language models to more efficient small language models (SLMs), with Apple and Microsoft introducing models with substantially fewer parameters yet comparable, or even superior, performance in benchmarks.” https://lnkd.in/gvre98xD

The article highlights Alibaba's significant advancements in AI technology, with its large language model, Qwen-72B-Instruct, topping the global ranking of AI developer platform Hugging Face. This achievement demonstrates Alibaba's commitment to open-source development and its rapid progress in AI technology. As a business analyst with expertise in data science and AI, I recognize the potential of large language models to drive business value and appreciate Alibaba's contributions to the field. Key points from the article include: - Alibaba's Qwen-72B-Instruct model achieved an average score of 43.02 across six benchmarks, surpassing other top-ranked models. - The Qwen model was trained on 72 billion parameters, enhancing its performance in math, long-range reasoning, and knowledge. - Alibaba's large language models have been adopted by over 90,000 corporate clients, showcasing their practical applications. - The Hugging Face ranking excludes advanced LLMs developed in a proprietary environment, such as OpenAI's GPT. Overall, this development reflects the growing importance of AI in business and the need for professionals to stay informed about the latest advancements in this field. https://lnkd.in/gMy_u_8E

The old adage 'garbage in garbage out' seems to be very much applicable to AI. Quote: "How did Microsoft cram a capability potentially similar to GPT-3.5, which has at least 175 billion parameters, into such a small model? Its researchers found the answer by using carefully curated, high-quality training data they initially?pulled from textbooks." It seems that if you want a useful AI that doesn't spout misogynistic, racist, violent, or just plain false information, with the potential to melt the planet in the process, you don't just let it loose and try to train it on the entire internet. Training on carefully curated data, turns out might be more ethical both in terms of respecting content rights and the environment. It seems to me that if rights holders of significant bodies of content train AI models on their data, but keep their data private, they're going to have more reliable, energy efficient models that the 'vacuum cleaner' approach won't be able to compete with. #ai #sustainability

Alibaba has launched its Qwen2-Math large language models, claiming they outperform OpenAI's and Google's in solving mathematical problems, contributing to the company's growing AI capabilities. Key Takeaways ? Alibaba introduced Qwen2-Math, a series of maths-specific large language models that reportedly outshine OpenAI’s GPT-4o and other major models in mathematical tasks. ? The Qwen2-Math models are focused on enhancing reasoning capabilities to solve complex arithmetic and mathematical problems. ? Testing on various benchmarks, including GSM8K and the gaokao, showed impressive performance despite existing limitations like English-only support. ? Future plans include developing bilingual and multilingual models to expand the capabilities of Qwen2-Math. ? Alibaba's models are recognized as competitive in the global AI landscape, narrowing the gap with US models. https://lnkd.in/eejs-ZDV

Global tech leaders like Microsoft and Meta are moving towards smaller, more efficient language models to save costs and energy. Microsoft announced the release of small language models with just 3.8 billion parameters, while Meta introduced two versions of its smaller Large Language Model, Llama 3, with 8 billion and 70 billion parameters. These compact models are designed to be more cost-effective and energy-efficient while still providing high performance in AI applications. #AI #SouthKorea https://lnkd.in/gsttVT9e

Apple tested over 20 Large Language Models (LLMs)—including OpenAI's o1 and GPT-4o, Google's Gemma 2, and Meta's Llama 3—to see if they were capable of "true logical reasoning," or whether their ‘intelligence’ was a result of "sophisticated pattern matching" and the results revealed some major weaknesses. LLM’s reasoning abilities are usually tested on the popular benchmark test—GSM8K—but there’s a probability that the?LLMs can only answer questions correctly because they’ve been pre-trained on the answers. Apple’s new benchmark—GSM-Symbolic—tested this by changing variables in the questions (eg. adding irrelevant information/changing names or numbers) and found?every LLM dropped in performance. As a result, they believe there is “no formal reasoning” with LLMs, “their behavior is better explained by sophisticated pattern matching” as?even something small, like changing a name, degraded performance by 10%. If LLMs can only rely on sophisticated pattern matching, rather than genuine logical reasoning, it means they can’t be relied on for AI applications that require consistent, accurate reasoning to help with real-world situations and environments, although it’ll be interesting to see how Apple, with its own AI models, and a major competitor of Google, Meta, and OpenAI (despite their new partnership), will respond to its own disparaging study.

?? Exciting news in the AI world! Microsoft has just unveiled a new class of small language models that pack a powerful punch in a tiny package. The Phi-3 family, including the remarkable Phi-3-mini, is set to revolutionize how we use AI, making it more accessible and efficient. ??These models are not only smaller and trained on less data, but they also outperform their larger counterparts in various benchmarks. This means organizations with limited resources can now leverage AI more effectively, and applications can run directly on devices, ensuring quick responses and enhanced privacy. ??The future is here, and it's all about choice. Whether you need a model for simple tasks or complex reasoning, there's now an AI model tailored to your needs. Check out Microsoft's latest innovation and explore the possibilities for small language models. #AI #MachineLearning #Innovation #MicrosoftAI https://lnkd.in/etN9n-NB

The pace we see in #AI is just breathtaking. I don’t know about you, I struggle to keep up….anyway, now we all have to find the business in it. ?On Tuesday, Microsoft announced a new, freely available lightweight AI language model named Phi-3-mini, which is simpler and less expensive to operate than traditional large language models (LLMs) like OpenAI's GPT-4 Turbo. Its small size is ideal for running locally, which could bring an AI model of similar capability to the free version of ChatGPT to a smartphone without needing an Internet connection to run it.“ https://lnkd.in/ezeApMju