December 22, 2023
📰 FEATURE STORY
Can Ola’s Krutrim make it big?
Looks like we’re squarely in the AI Large Language Model (LLM) race now. Millions of dollars are being spent to develop AI foundation models à la ChatGPT and Google’s Bard. Closer to home, some Indian companies are busy with LLMs trained in Indian languages.
The most recent entrant is Ola; yes, the ride-hailing company. It unveiled “India’s first full-stack AI solution” called Krutrim. The LLM is built from scratch and trained in Indian languages and wants to take on the likes of ChatGPT and Bard. While this is clearly just the beginning for Ola, can it make it big in the long run?
Context
First things first, what’s a LLM? LLM is an AI algorithm that uses deep learning techniques and large amounts of data to understand, summarise, generate, and predict new content. Generating text-based content is a type of LLM. Let’s put it another way. Humans have used languages to communicate, which provide words, semantics, and grammar to convey ideas and thoughts. In the world of AI, a LLM serves the same purpose.
All language models are first trained on data sets and then use various techniques to generate new content based on those sets. It’s not exactly a recent phenomenon. Language models go back to the earliest days of AI. The Eliza model debuted at the Massachusetts Institute of Technology (MIT) in 1966.
Since then, significant progress has been made that propelled LLMs forward. One came in 1997 with the introduction of Long Short-Term Memory (LSTM) networks that could handle more data. In 2011, Google Brain gave researchers more powerful computing power and data sets.
Modern LLMs emerged in 2017 and used “transformer models”. They’re essentially neural networks. Here, LLMs understand and generate new content faster with more parameters and transformer models. OpenAI’s ChatGPT-3 is one such model. Some LLMs are called ‘foundation models’, coined by the Stanford Institute for Human-Centered Artificial Intelligence in 2021. A foundation model is so big that it’s the base for further optimisation and particular use cases.
The basic process of how LLMs work involves multiple steps. They’re first trained on large volumes of data. The model then begins identifying and developing patterns between words and concepts. After this, some fine-tuning is done to help the model better identify the different concepts.
Then comes deep learning. The model understands and recognises connections between words and concepts using a self-attention mechanism. This mechanism assigns a score, called a weight, to a given item, called a token. Then it’s ready to go. After providing the LLM with a prompt, the model will generate a response.
Back home, in April, Ola founder Bhavish Aggarwal registered a new company named Krutrim SI Designs Private Ltd. The rumours were that the company would be entering the AI space. Aggarwal is someone who has pushed for India to take the lead in adopting AI technologies. Last week, Krutrim (Sanskrit for “artificial”) was finally unveiled as “India’s own AI.” The base version will go live next month, and the more sophisticated Pro model will be available next quarter.
With India now having a Made in India LLM, can it take on the big guns like OpenAI’s ChatGPT and Google Bard?
VIEW: Not afraid of the competition
When announcing Krutrim, Bhavnish struck a confident tone on its capabilities and comparisons with the competition. He compared it with the Llama 27B model, claiming superior performance on several recognised LLM benchmarks. What gives Krutrim the edge in some ways is proficiency in understanding 20 Indian languages and generating content in 10. This is something of a must for India, given its diversity and is a significant step in linguistic inclusivity.
The Made in India part can’t be underscored. Other LLMs are predominantly trained on Western data. Krutrim is trained on 20 times more Indian tokens than any other model. This sets Krutrim apart. During its launch, it was emphasised that Krutrim was tailored to capture the nuances of India’s heritage and culture. Then there’s the accessibility feature of Krutrim going beyond text and responding to voice inputs.
The other interesting thing about Ola’s venture is they’ll be developing cost-effective silicon systems and platforms tailored to India’s needs. Plus, it’s all done in-house. Even one of its local competitors, Zoho, working on its own AI mode, depends on AMD and NVIDIA for chipsets.
COUNTERVIEW: Tough to dethrone the incumbents
Building such an AI model isn’t cheap. Krutrim raised $24 million from Matrix Partners, also an investor in Ola Electric. Compare that with Google investing billions of dollars in AI companies and research. Microsoft has invested about $10 billion in OpenAI. Both companies have significantly advanced their AI capabilities and implementation. They’ve also got a lot more experience.
While the company is taking a bootstrap approach, we still don’t know the amount being spent on Krutrim and where the next round of investments will come from. While AI is certainly taking off, the question remains on whether Ola will in any way be affected by Krutrim’s investments and costs. There’s also the question of competition at home with the likes of Nilekani Center at AI4Bharat, Zoho, and others. It remains to be seen if they can compete with the global bigwigs.
While Aggarwal did disclose quite a bit on Krutrim, there’s still quite a bit we don’t know. There’s still no specific information on the datasets it’s trained on. Also, how did the company manage to build the model in a relatively short time? Krutrim also hasn’t disclosed the number of Graphic Processing Units (GPUs) needed to train their model, only mentioning it was all in-house. For context, GPT was in the works for over six years and used about 30,000 GPUs.
Reference Links:
- India’s LLM Battle: Ola’s Krutrim vs Zoho – Analytics Magazine
- Ola chief Bhavish Aggarwal floats new company, likely to enter AI space – The Economic Times
- Meet Krutrim, ‘India’s own AI’ developed by Ola CEO Bhavish Aggarwal-led venture – Hindustan Times
- Ola’s Bhavish Aggarwal takes the ‘made for India’ approach with Krutrim AI model – Moneycontrol
- Krutrim can outperform GPT-4 in 6 Indian languages! 10 things to know about Ola’s AI platform – Mint
- Ola Unveils ChatGPT Clone, Calls it Krutrim – Analytics Magazine
What is your opinion on this?
(Only subscribers can participate in polls)
a) Krutrim can make it big.
b) Krutrim will have a tough time making it big.
🕵️ BEYOND ECHO CHAMBERS
For the Right:
Lessons for Modi government on parliamentary democracy — from L K Advani and Sushma Swaraj
For the Left:
How Hamas attack on Israel woke America up to Left’s influence in academics