“Our chief technology officer Devesh Mishra has dropped all his other responsibilities to focus on building SLMs for Physics Wallah,” company founder and chief executive Alakh Pandey told Mint in an interview.
The Noida-based company aims to have these SLMs ready by the end of 2026.
“The investment is huge and I’m investing blindly. I’m not thinking too hard about it because I don’t want to miss this [AI], because timing here will make a lot of difference,” Pandey said. The company is able to make substantial investments in AI because the rest of Physics Wallah’s projects run “frugally”.
SLMs are trained on significantly fewer parameters, ranging from a few million to a few billion. In comparison, large language models (LLMs) such as ChatGPT are trained with hundreds of billions or even trillions of parameters. SLMs are also cost-efficient because they require significantly less computational power and can be tuned to perform specific tasks—in Physics Wallah’s case, answering questions on different subjects.
“Every business will potentially do this. And if you look at the broader spectrum of AI, they say that anything related to STEM skills—like the subjects PW is potentially using AI for—will likely be impacted by AI,” said Amit Nawka, technology deals partner at PwC India. “Tech-first companies providing education in this space will need to adopt AI.”
Also read | Physics Wallah’s big leap: New funding, higher valuation, IPO next
Physics Wallah, seeking to race ahead of peers, is hiring for its AI and data science teams and has given them complete freedom to build products, without worrying about success or failure.
The company’s latest AI offering, Project Bharat, is part of its larger focus on AI as it prepares to file for an initial public offering this year. Project Bharat attempts to localize video lessons with context and teachings specific to local regions of India.
The goal is twofold: “How do I personalize classes for students while keeping costs down. Secondly, how do I reach out to all of our students with the help of AI, without increasing the number of teachers we already have,” Pandey explained.
But PwC’s Nawka flagged challenges in the monetization of such AI tools.
“For edtechs, the question is whether AI will be integrated as a feature within existing offerings—will it be part of a subscription? Or can it be monetized separately?” he said. “What’s certain is that they have to adopt AI. It’s not a choice. The depth of investment and spending will depend on ROI.”
Human touch
The Physics Wallah team had been working on AI-driven dubbing but wasn’t satisfied with the results. “The problem with AI was that, conceptually, it’s still not strong enough. No LLM is advanced enough to fully understand concepts from physics,” Pandey said.
AI also struggled with local nuances. “Let’s say, in a lecture, I’m giving examples from northern states—maybe talking about Bihar or Lucknow. AI wasn’t able to understand and adapt these references properly.”
Realizing that a human touch was essential, Physics Wallah shifted its approach. “What we did was let the dubbing be handled by a professional dubbing artiste, because they understand how to convey emotions, and maintain academic accuracy. Then, we trained AI on my voice—its modulation, the way I speak,” he explained.
A significant part of the investment is now going into hiring dubbing artistes. Pandey sees this as an initial version of Project Bharat but believes that as they advance with SLMs, roadblocks will reduce.
“Doing dubbing well is going to be expensive. PhysicsWallah needs to see if they’re an education company or a technology company,” according to AI-first venture capital firm SenseAI founder Rahul Agarwalla.
He suggests that the company should use third-party software from companies like unscript.ai that are actively working in that space. “The technology is in flux, which means its very expensive to keep up with and build a great team,” Agarwalla said. “It’s a lot of capital going into something that might be better off going into the company’s core competencies.”
Immortalizing ‘star’ teachers
As the project progresses, PhysicsWallah’s aims to preserve and scale the teaching of “star” educators like Pandey—those with exceptional delivery skills, strong student feedback and high learning outcomes.
“What is happening now in India is that there are very few teachers with exceptional delivery skills. So, the question is whether these teachers can teach the whole of India with the help of AI and dubbing artistes,” he said.
Pandey added that while the project is still in its early stages, it has the potential to scale. “This is a project for the entire country, which is currently being thought of on a very small scale. We don’t know yet where it will go.”
Project Bharat is set to go live with its first batch of students on 5 April. Physics Wallah had initially planned to take it live with 16 languages. Pandey, however, limited it to four languages for the launch.
“This is the year that we’re going to build out Project Bharat. We want to crack the formula on how to make it work, what are the problems we need to iron out, what are the kinds of dubbing artistes that we’ll need and what issues within the AI we need to solve,” Pandey said.
With Project Bharat, the company has loftier ambitions. Physics Wallah doesn’t just want to recreate the content it already puts out but also move into different streams of education. Or, as Pandey puts it, “Taaki main saare subjects bhi padha sakta hun (So that I can teach all subjects).”
AI offerings so far
Physics Wallah already has three AI-driven products in circulation: AI Guru, a multimodal round-the-clock companion that can answer doubts related to coursework; AI Grader, which is able to grade answers of students; and Sahayak, an AI assistant capable focusing on practice and revision.
Initially, the company had wanted to name its suite of AI products Alakh AI. But he was apprehensive citing concerns around AI hallucinations—a phenomenon when an AI model generates incorrect or misleading information while passing it off as correct, essentially, ‘making things up.’
“I said let’s not release it to the children with my name on it because it’s new AI. It can say something controversial. If ‘Alakh AI’ says something incorrect, then the impact would be wrong,” Pandey said. Ultimately, the Physics Wallah team went ahead with AI Guru, but internally, it continues to be referred to as ‘Alakh AI.’
Also read | Physics Wallah ropes in Hornbill Capital for a lesson in IPOs
Initially, AI Guru was able to solve 30-40% of queries with approximately 50% accuracy. “Today, the bot can solve 95% of queries, with an accuracy of 90%,” Pandey said, attributing it to the company’s large database. Last year, AI Guru, the company claimed, solved 50 million doubts for students.
For context, Physics Wallah claims to have a database of 5 million questions related to the Indian Institutes of Technology (IITs), the National Eligibility cum Entrance Test (NEET), the Union Public Service Commission (UPSC) exams as well as major education boards.
Physics Wallah was started as a YouTube channel in 2016 by co-founders Pandey and Prateek Maheshwari, and made a formal beginning as an edtech startup only in 2020. It became a unicorn, or a billion-dollar company, with its first funding round two years ago. The startup has raised over $300 million so far.
Mint exclusively reported previously that the edtech startup is set to raise ₹200-230 crore ($22-26 million) in a secondary funding round led by existing investor WestBridge Capital, at $3.7 billion—up from $2.8 billion just six months ago.
The company has been chasing aggressive growth over the past two years, entering new business segments and making acquisitions to expand into new markets. Its consolidated revenue nearly tripled to ₹2,015 crore in the fiscal year ended 31 March 2024, but its losses surged on the back of a jump in expenses.
Leave a Reply