TWO Platforms, founded by renowned AI innovator Pranav Mistry, introduces SUTRA, a groundbreaking multilingual generative AI model poised to revolutionise the global AI landscape. SUTRA empowers individuals and enterprises alike with its unparalleled language fluency, real-time information retrieval, and unmatched speed—all while emphasising the democratisation of AI access. 

TWO is a tech company that aims to redefine human-AI interaction through its proprietary multilingual and cost-efficient language model, SUTRA, and its solutions and services. 

What is Sutra? 

SUTRA is a novel multilingual large language model architecture trained by decoupling concept learning from language learning. It functions in over 50 languages and outperforms many multilingual benchmarks with cutting-edge performance. “SUTRA” comes from Sanskrit and means “aphorism” or thread. It illustrates how we combine succinct wisdom from several languages. Because most generative AI models are trained in English, billions of non-English speakers cannot use them due to their neglect of global linguistic diversity. 

Language models are tested using the Massive Multitask Language Understanding (MMLU) benchmark. It was found that modern language-specific LLMs such as HyperClovaX in Korean, Airavata in Hindi, Jais in Arabic, and Rakuten-7B in Japanese are routinely outperformed by the Sutra models. 

In addition to outperforming previous models in Hindi, SUTRA can also handle and recognise Hinglish, a colloquial Hindi and English language spoken by millions of people in India. Sutra models have demonstrated notable gains in MMLU scores in Korean, one of the most important markets; in fact, they have outperformed cutting-edge Korean LLMs, like NAVER’s HyperClova X model.  

SUTRA is superior to several other languages, including Arabic and Japanese. Its strengths include its broad range of multilingual capabilities at a significantly lower cost. 

Versions of SUTRA models 

• SUTRA-Pro - best performant multilingual model 
• SUTRA-Light -lightweight and cost-efficient multilingual model 
• SUTRA-Online -internet-connected multilingual model 

Significance of SUTRA 

According to Pranav Mistry, SUTRA (सूत्र) is a mission to fix the language gap in AI language models. He said, “I genuinely believe the world deserves high-quality and cost-efficient AI language models, free of language barriers. With SUTRA, we are introducing cost-efficient foundational LLMs that excel for हिंदी, ગુજરાતી, বাংলা, العربية, मराठी, తెలుగు, தமிழ், ಕನ್ನಡ, മലയാളം, ଓଡ଼ିଆ, ਪੰਜਾਬੀ, and over 50 other languages.” He added, “SUTRA is AI for all, AI that can understand the nuances of languages and dialects—AI beyond just English.” Furthermore, SUTRA models are trained with an innovative LLM architecture that learns new languages independently, making it multilingual, highly scalable, and cost-efficient.  

Challenges involved 

Because current AI models are primarily trained on English data, they have not adequately catered to emerging markets. Due to the substantial data and training needs, language-specific LLMs such as HyperClova in Korean, Rakuten in Japanese, OpenHaathi in Hindi, and Jais in Arabic are expensive and difficult to scale. A distinct model with specialised training and deployment resources is required for every language. Furthermore, the process becomes more costly because each new base model needs to be adjusted for multiple languages. For example, a model trained in Hindi won’t be able to function in Arabic or Spanish. 

Opportunities of Sutra 

TWO’s official statement says, “We are committed to pioneering AI solutions for non-English markets. Our SUTRA models unlock AI growth opportunities in large economies such as India, Korea, Japan, and the MEA region.” SUTRA is intended to close the linguistic gap in LLMs by developing a multilingual foundational model. This will make AI more widely applicable and efficient by opening up new application opportunities in various industries, including banking, healthcare, and customer service. 

The linguistic diversity of India presents both opportunities and challenges. The team opined that their goal with SUTRA is to eliminate linguistic barriers and provide AI-powered solutions to all regions of the nation. They opined that SUTRA has the power to revolutionise how people and businesses use technology. “Our goal is to create an AI that can engage in meaningful conversations, understand the nuances of different languages, and provide accurate and contextually relevant responses”, they added. 

Comparison with other LLMs 

Some LLMs, such as GPT 3/4 or Llama, have multilingual capabilities. However, they have trouble with complicated multilingual tasks and frequently fall back to English, especially when dealing with languages with lower resources. These large, English-centric models are also ineffective and inappropriate for certain non-English use cases that require low-cost deployment for applications in developing nations such as India.  

Conventional multilingual LLMs lack a way to distinguish between language learning and idea learning and are trained on data that is strongly biased towards English. This results in confusion and decreased performance between languages. 

• Source: TWO Platforms
• Image: Unsplash