Top AI companies are unveiling language models that redefine an increasing number of business processes and have proven their versatility in many use cases. Innovation in AI and machine learning helps develop LLMs (large language models) and also plays a vital role in language translation as they break down language barriers by providing accurate and contextually relevant translations. Let’s dive into the AI language models developed by the top companies recently.
Microsoft’s Orca 2 LLM
The Orca 2 LLM is a tuned version of the Llama 2 that performs as well or better than models that contain 10x more parameters. Orca 2 uses a synthetic training dataset and a new technique called Prompt Erasure to achieve this feat. Orca 2 models are trained using a teacher-student scheme where a larger, more powerful LLM acts as a teacher for a smaller student's LLM to improve the student's performance to be comparable to that of the larger model. Microsoft's training technique teaches the smaller model multiple reasoning techniques, as well as how to choose the most effective technique for a given task. For this, the teacher is given sophisticated prompts to induce certain thinking. However, in a scheme called Prompt Erasure, the student only receives the prompts for the task and the required response, but not the teacher's prompt. When benchmarked, the 13B Orca 2 outperformed the base 13B Llama 2 by 47.54 percent. The 7B Orca 2 was "better or comparable" to the 70B Llama 2 in reasoning tasks.
Alibaba Cloud Unveils AI Models That Understand Visual Content
Alibaba Cloud has unveiled two open-source artificial intelligence (AI) models that can understand both images and text. These models—Qwen-VL and Qwen-VL-Chat were trained on Alibaba Cloud's Qwen-7B large language model and offer better performance in image recognition and understanding compared to other open-source large vision language models. The launch of these new models demonstrates Alibaba Cloud's commitment to improving the multimodal capabilities of its large language models. By incorporating sensory inputs such as images and sound, Alibaba Cloud aims to explore new applications for researchers and commercial organizations. These models have the potential to transform user interactions with visual content, the report says. One of the key advantages of these models is their ability to generate photo captions for news channels or to help non-Chinese speakers read street signs that are in Chinese. In addition, the models allow visual answering of questions, making shopping more accessible for blind and visually impaired users.
Ola Unveils ‘Made for India’ Krutrim AI
Ola CEO Bhavish Aggarwal took to the stage in Bengaluru to introduce India's own AI model called Krutrim. A new AI model was also showcased during the event, and it was able to respond in various Indian languages. The Krutrim AI model can also code in real time. According to Aggarwal, the new AI model will be able to better understand the Indian context compared to AI models trained mostly on Western data.
Krutrim was introduced in two classifications. The basic LLM model called Krutrim was trained on 2 trillion tokens. The model understands 22 languages and can generate output in 10 Indian languages. The company claims that the new AI Krutrim outperforms the GPT-4 and Llama in terms of Indian performance. The company claims that AI Krutrim has been trained on 20 times more Indian tokens than any other model. The new model can use voice as input, and the model will also respond with voice. The Krutrim team in Bengaluru also mentioned a more powerful model: the Krutrim Pro. This particular model will be multi-modal but will be available next quarter. According to the company, the new model was built from the ground up, and the entire chimney was built by Krutrim.
"Aggarwal says, "India needs a solution that it thinks is a foundation for itself. For India to be an AI-first economy, we need to build the whole stack on India's performance levels, India's cultural relevance and structure costs in India,” adds Aggarwal.
"It is our time to realize the dream of Viksit Bharat and lead the AI vision over the next 25 years. AI is going to define the future paradigm of economy and culture. We are extremely excited to launch India’s first complete AI computing stack, Krutrim, which understands our unique cultural context, connecting our future to our roots. With an India-first cost structure, Krutrim will have the largest representation of Indian data, enabling us to create novel models beyond LLMs across sectors, making India the most productive, efficient, and empowered economy in the world.”
Huawei Unveils Latest AI Large Language Model Pangu 3.0
Huawei Technologies Co unveiled the latest version of its large artificial intelligence language model, Pangu 3.0, as part of its broader effort to use artificial intelligence to accelerate digital transformation across a range of industries. Huawei said its large language model Pangu has already been used in more than 10 industries, such as finance, manufacturing, government affairs, energy, coal mining, healthcare, and railways, and supports the implementation of AI applications in more than 400 business scenarios.
Amazon Plays Catch-Up with Alexa Generative AI
Amazon unveiled a range of new and updated devices and updated its Alexa voice assistant with generative artificial intelligence to lure users to the unprofitable product as competition grows from chatbots such as Google's Bard. Alexa will converse more naturally and lose her robotic tone for almost a decade and answer questions like what time football games start and recipe ideas. It will also be able to compose and recite poems.