Sarvam: India’s Ambitious Generative AI Startup

Salutations, Olio aficionados! 👋

Happy Hump Day and welcome to the 103rd edition of Weekly Olio - your trusted source for giggles, wisdom, and a dash of intrigue, courtesy of our tantalizing thought piece (yes, buckle up for Publisher's Parmesan). 🧀

This edition of Weekly Olio is brought to you by Salesforge.

See AI Agent Frank In Action!

Lower costs by letting Agent Frank, Salesforge’s AI SDR, take care of prospecting, crafting messages and booking meetings while your team can focus on closing deals! Now, you can get a personalized email from Agent Frank to test out his capabilities in real-time:

Show me the magic

The Quote󠀢 💭

The Tweet 🐦

Introducing Systems of Agents: These are AI-powered tools that do more than just help—they take action. They can read emails, understand phone calls, and handle documents. Most importantly, they can make decisions on their own. They transform messy business communication into clear, useful information.

The Infographic 💹

India's Generation Z, comprising 377 million people, is reshaping the consumer landscape, driving $860 billion in spending, projected to reach $2 trillion by 2035, according to a Snap Inc. report with Boston Consulting Group. Read more…

The Short Read 📝

Apollo pushes into high-grade debt business long dominated by banks - by FT

Apollo is getting more involved in high-quality lending. In private equity, it was discovered early on that even a modest increase in value could lead to great returns if you used enough borrowed money. The same idea applies to lending: if your borrowing costs are determined by what life insurance or annuity holders expect, then it's more important to safely invest a large amount of money than to aim for the highest possible return on each investment. Read more…

Apollo can gather plenty of money from insurance customers to support loans. However, the challenge for the company is to come up with smart investments, like the big deals such as the Intel transaction, to ensure bigger and safer profits.

The Long Read 📜

AI Detectors Falsely Accuse Students of Cheating - by Jackie Davalos and Leon Yin

Bloomberg has a good roundup of the state of AI detection in grading. It focuses on the false positives, i.e. students who were incorrectly accused of using AI, in part because it's hard to get a good handle on how AI-generated papers get passing grades. (Especially in environments with a late-Soviet approach: students pretending to write papers that teachers pretend to read. That will be especially hard to detect because the sides that find this mutually satisfactory will find each other, and neither side will really complain). Read more…

As with other applications of AI, LLMs in academia are a labor-saving measure that introduces new monitoring costs. There are completely reasonable ways to use them in text composition—they're great for tip-of-the-tongue searches, and a good way to get unstuck if you're lost enough that you don't know what to Google to figure out what you're doing wrong. They are also a tempting substitute to sitting in front of a blank page and trying to whip out an essay. It's either up to schools to figure out better ways to grade or up to students to decide how they'll demonstrate proficiency if doing the work is optional.

Publisher’s Parmesan 🧀

Sarvam: India’s Ambitious Generative AI Startup

Artificial Intelligence (AI) is rapidly evolving, and while India’s AI industry is still in its early stages, one startup, Sarvam, is making waves. Based in Bangalore, Sarvam has attracted attention for raising over $50 million in under six months, backed by investors like Peak XV and Khosla Ventures. In just one year, Sarvam has launched five AI-driven products, placing itself at the forefront of generative AI in India.

However, as the hype grows, so do concerns about the performance and robustness of its products. Let’s break down what Sarvam has achieved so far, the challenges it faces, and whether it can live up to its promise of revolutionizing India’s AI landscape.

Sarvam’s product lineup addresses different use cases in AI, with a focus on Indian languages and markets. Here’s a look at some of their key offerings:

1. Sarvam Agents: A voice-enabled service that helps companies manage customer care and sales calls in 10 Indian languages. It’s designed to provide businesses with multilingual AI agents for better customer engagement.

2. Sarvam 2B: A small language model (as opposed to large language models, or LLMs) launched in 2023, this AI tool is geared towards tasks like translation and summarization in Indian vernacular languages. Small language models use fewer parameters, which are essentially the "settings" that AI models learn to adjust during training. This product is touted as being more adept at Indian language tasks than international models like Meta’s LLaMA.

3. Shuka 1.0: An audio language model designed to handle Indian voice inputs and provide text outputs. It is positioned as India's first open-source audio language model and can be likened to a localized version of Siri. However, concerns have been raised about its transcription accuracy.

4. Sarvam Models: A collection of AI models trained to perform specific tasks, like legal drafting with A1, designed to assist lawyers.

Sarvam’s focus on developing products that cater to Indian languages and local markets is part of its mission to make AI accessible to Indian businesses and consumers. But the startup’s rapid pace of product launches may also be one of its biggest challenges.

The Glitches: A Closer Look at Sarvam’s Products

Despite the excitement surrounding Sarvam, there are fundamental issues with its technology. For instance, Shuka 1.0, the open-source audio model, struggles with basic transcription accuracy. It also finds it hard to differentiate between different speakers in a conversation, a critical function for audio models. Moreover, the model requires users to specify the language of the audio before processing it, which creates problems when the audio contains multiple languages, a common occurrence in India.

Similarly, Sarvam’s voice model was tested on their website but failed to deliver consistent results. Out of 10 attempts, the model only gave the correct response twice. This is particularly troubling for a product designed to handle multilingual voice inputs.

Developers pointed to the frugality of the model’s training process as a key reason for its underperformance. Sarvam claims to have trained its audio model with just 100 hours of data, which is considerably lower than what’s typically needed to develop robust AI models. In the context of generative AI, frugality may save costs, but it comes at the expense of accuracy and functionality, especially in a complex linguistic landscape like India’s.

The Role of Synthetic Data: A Double-Edged Sword?

Another major point of discussion around Sarvam’s AI models is its reliance on synthetic data. Synthetic data is artificially generated using algorithms, and it’s often used to train machine learning models when real-world data is insufficient or too noisy.

Sarvam’s yet-to-be-released text model reportedly uses 100% synthetic data, an unusual approach in the AI field. While synthetic data can be useful, relying entirely on it presents risks. Researchers have observed that models trained exclusively on synthetic data can produce nonsensical results, as they fail to capture the nuances of real-world scenarios.

In India, where linguistic diversity is vast, high-quality, real-world data is difficult to come by for many regional languages. Sarvam’s use of synthetic data could be a solution to this problem, but the risks of overreliance are real. Larger models like Meta’s LLaMA and OpenAI’s video model Sora also use synthetic data, but not to the extent Sarvam does.

Hire Ava, the AI SDR & Get Meetings on Autopilot

Ava automates your entire outbound demand generation process, including:

• Intent-Driven Lead Discovery

• High Quality Emails with Waterfall Personalization

• Follow-Up Management

Free up your sales team to focus on high-value interactions and closing deals, while Ava handles the time-consuming tasks.

Book a demo to see how Ava can 10x your outbound.

Rushing to Market: Is Speed Sacrificing Quality?

Sarvam’s products have been developed in record time, launching multiple models within a year of the company’s founding. While this fast-paced approach has garnered attention from investors and the media, it may be coming at the cost of quality.

The AI models built by Sarvam appear to have been trained on limited data, and the technical issues that have emerged suggest that more time and resources are needed to improve their performance. In a field where data quality is paramount, Sarvam’s rush to market could hurt its reputation if these issues are not resolved quickly.

Additionally, Sarvam’s claim of being R&D-intensive raises questions about whether the company has conducted sufficient research and development to match its ambitious product pipeline. Launching AI models without adequate training or real-world testing risks alienating potential customers, especially in sectors like customer service and legal work where precision is critical.

Sarvam’s Founders and Support Network: A Strong Foundation

Despite these challenges, Sarvam’s founding team and support network provide reasons for optimism. Vivek Raghavan, one of the co-founders, played a key role in developing India’s Aadhaar system and the Unified Payments Interface (UPI), two of the country’s most successful digital infrastructure projects. His expertise in large-scale, data-driven systems gives Sarvam credibility.

Co-founder Pratyush Kumar, formerly of IIT Madras and a co-founder of AI4Bharat, brings academic rigor and technical depth to the company. Additionally, Sarvam is backed by major tech partners like Meta, Microsoft, and Nvidia, as well as Nandan Nilekani, co-founder of Infosys and architect of Aadhaar.

These connections provide Sarvam with a solid foundation, access to cutting-edge technology, and potential avenues for improvement. However, having a strong team and high-profile backers may not be enough to overcome the technical flaws in its products.

The Road Ahead: Can Sarvam Deliver on Its Promise?

Sarvam is positioned to be a trailblazer in India’s generative AI landscape, but it faces significant hurdles. The startup’s reliance on synthetic data, frugality in training models, and the speed at which it has brought products to market have led to concerns about the quality and functionality of its AI models.

To succeed, Sarvam must focus on improving the performance of its products, particularly in the areas of transcription accuracy, multilingual support, and voice recognition. The company’s ambitious goal of building AI models that cater to India’s diverse linguistic landscape is laudable, but it requires more robust data, better training processes, and longer development cycles to reach its full potential.

Ultimately, Sarvam’s journey highlights both the promise and the pitfalls of building AI in a rapidly growing but complex market like India. If the startup can address its technical challenges and capitalize on its strong foundation, it could become a key player in the global AI ecosystem. But for now, the road to that success seems fraught with obstacles.

Olio Jobs 💼

• Clay | New York, NY - Senior Product Designer, Go-To-Market Engineer, Head of Engineering, Data Scientist, Enterprise Growth Strategist

• Retool | San Francisco, CA - Product Marketing Manager, Security Engineer, Software Engineer, Product Designer, Sales Engineer (Enterprise)

• Modern Treasury | San Francisco, CA or New York, NY - AI/ML Ops / Software Engineer (Tech Lead), Revenue Operations Lead, Software Engineer (Senior)

• Pinecone | New York, NY or Remote (US) - Technical Support Engineer, Engineering Director (Experience), Platform Engineer

That’s all for this week. If you enjoyed this edition, we’d really appreciate if you shared it with a friend, family member or colleague.

We’ll be back in your inbox 2 PM IST next Wednesday. Till then, have a productive week!

Peeyush and Mohit

Disclaimer: The views, thoughts, and opinions expressed in the text belong solely to the author, and not necessarily to the author's employer, organization, committee or other group or individual.