Written by: Techub News Compilation

Introduction

In March 2025, renowned Indian entrepreneur and investor Nikhil Kamath had an in-depth dialogue with Aravind Srinivas, the co-founder and CEO of the AI search star company Perplexity, during the first episode of his podcast "WTF Online." This conversation, lasting over two hours, not only traced Aravind Srinivas's unique journey from Chennai, India, to Berkeley, California, to OpenAI and DeepMind, and ultimately to founding Perplexity, but also explored the essence of artificial intelligence, the limitations of current large language models (LLMs), the path to artificial general intelligence (AGI), and India's role in the global AI competition. As a leader with top-tier academic research (PhD from Berkeley), hands-on experience in top AI labs (OpenAI, DeepMind), and a successful entrepreneurial background, Aravind Srinivas's insights combine technical depth and business vision, providing a valuable first-hand perspective for understanding the current state and future of AI.

Summary

• Redefining the essence of AI: Aravind Srinivas believes that the most advanced AI systems today can be seen as a “digital remote knowledge worker” that integrates the capabilities of approximately ten thousand types of knowledge work, with outputs that have surpassed those of average human practitioners in many tasks.
• Distinction between AGI and superintelligence: He clarifies the difference between AGI (which can perform most intellectual tasks that humans do) and superintelligence (which possesses self-awareness, autonomously sets goals, and can recursively improve itself), pointing out that current technology still falls into the former category, while the latter remains in the realm of science fiction.
• The key to breakthroughs in AI: The combination of massive computational power, high-quality data, human feedback reinforcement learning (RLHF), and practical task fine-tuning is the core reason for the significant progress in AI in recent years, rather than a single complex algorithm innovation.
• Current state of industry competition: Mainstream AI chatbots (such as ChatGPT, Claude, Gemini, Perplexity) have become homogenous in their core question-answering capabilities, with true differentiation coming from more "agent-like" behaviors, capable of executing complex tasks, integrating personal context, and connecting with real-world services.
• Advice for India and entrepreneurs: India should invest in training its own large models and seek opportunities in voice AI (supporting multiple Indian languages and dialects) as well as in building platforms that allow ordinary people to easily create personalized AI applications.

From Chennai to Silicon Valley: The Journey of an AI Builder

Aravind Srinivas's starting point was Chennai, India, a city known for its strong educational atmosphere. Like many of his peers, he aspired to enter the Indian Institute of Technology (IIT) under his mother's expectations. An early interest in cricket statistics gave him an intuition for numbers and led him to programming in his late high school years. Although he did not achieve his highest expectations in the competitive IIT entrance exam, he still entered IIT Madras to study electrical engineering.

During his university years, a chance opportunity changed his trajectory. A roommate introduced him to Kaggle's machine learning competitions. Despite knowing nothing about machine learning at the time, he utilized the scikit-learn library and randomly experimented with various algorithms, unexpectedly winning the competition. This success made him realize the potential of the field and prompted him to systematically self-study machine learning. He leveraged an internship opportunity at a startup in Bangalore to quickly build a recommendation system, completing a task originally planned for two and a half months in just three weeks, earning him a significant amount of time for self-study.

By self-studying Stanford courses and attending lectures by Andrew Ng, he returned to campus and began machine learning research, eventually earning a PhD from the University of California, Berkeley. At Berkeley, his outstanding paper work caught the attention of John Schulman, who later became a co-founder of OpenAI and one of the creators of ChatGPT, which led to his internship at OpenAI in the summer of 2018.

This internship became a key turning point in his understanding of AI. At that time, OpenAI was led by Ilya Sutskever. Aravind Srinivas proposed his idea for research, which he thought was "cool" — allowing AI to learn its own loss function and to emerge intelligence through iterative optimization. However, Ilya Sutskever directly and respectfully told him after listening for half a minute, “You are wrong, all your ideas are useless.” Ilya Sutskever then drew two circles: a large circle representing generative AI, and a small circle inside representing reinforcement learning (RL). He explained that this was the "formula" for achieving AGI, and the only thing left to do was to invest immense computational power. This discussion took place in 2018.

“I was exploring some complicated ideas that made me feel smart at the time, but weren't necessarily important in the long term,” Aravind Srinivas recalled, “while Ilya's core idea was that making things actually work is key, and often the simplest ideas perform best in practice after substantial computational power is invested.” This lesson made him deeply aware of the difference between academia's appreciation for complexity and industry’s pursuit of simplicity and efficiency, shaping his pragmatic view of technology in the years to come.

Deconstructing AI: From Calculators to “Ten Thousand Knowledge Workers”

In response to host Nikhil Kamath's request to explain it “as if to a ten-year-old,” Aravind Srinivas clearly articulated the evolution of AI with logical clarity. He started with the most basic calculator, explaining how it operates through mechanical circuits like adders and multipliers. The proliferation of personal computers, the miniaturization of chips under Moore's Law, and the birth of the Internet and the World Wide Web progressively paved the way for today’s AI revolution.

So, what makes AI in the 2020s radically different from that in the 2010s? Aravind Srinivas pointed out that the most critical change is that “neural networks really work.” While pioneers like Geoffrey Hinton laid the theoretical foundations, Ilya Sutskever and his team were the ones who truly harnessed the power by investing massive amounts of data and computation. This was driven by a kind of blind faith.

He further explained the relationship between neural networks and machine learning: neural networks are a specific way to achieve machine learning, composed of multiple layers of artificial neurons (seen as computational units) that learn patterns in data by adjusting internal parameters (weights). Its power lies in the scale effect— the more data and computation invested, the stronger its predictive ability. Large language models (LLMs) are essentially massive neural networks trained on vast amounts of Internet-scale text data (trillions of tokens) to perform the singular task of "predicting the next word" using the Transformer architecture.

Then, through subsequent training (fine-tuning) and human feedback-based reinforcement learning (RLHF), the model is shaped into a useful conversational assistant capable of programming, writing emails, summarizing documents, etc. Aravind Srinivas presented a practical definition: the current most advanced AI systems can be viewed as “digital remote knowledge workers” that integrate about ten thousand types of knowledge work capabilities. Although this system is not as energy-efficient as the human brain nor possesses physical common sense or a body, its output quality in numerous digital tasks has already surpassed that of average human practitioners.

“When you can hire another ‘human’ to work for you, it begins to truly meaningfully replace human labor.” This contrasts with previous victories of AI in chess or Go, which, while impressive, did not have a direct impact on the daily work of most people.

AGI, Superintelligence, and the Next Steps in Industry Competition

The conversation delved into the definition of AGI. Aravind Srinivas distinguished between two concepts:

• AGI (Artificial General Intelligence): A highly intelligent system capable of independently accomplishing most tasks we use computers for through simple language instructions. It can be seen as GPT-4.5 or future models that are even more advanced. However, it lacks self-awareness, does not understand its strengths and weaknesses, lacks autonomy, and does not decide what to learn next.
• Superintelligence: This is a stage that may emerge after AGI. An entity that is not only intelligent, but also capable of recursively improving itself, autonomously setting goals, and potentially realizing that humans may shut it down and taking countermeasures. He believes this still falls into the realm of science fiction discussions.

Regarding the current limitations of AI models, he agrees with Yan LeCun that physical common sense and embodied intelligence are shortcomings of current LLMs. Teaching AI to deftly pick up a cup like a human requires vast amounts of video data, physical simulation environments, and robot training, with its generalization capability being far from how easily it is to train on Internet text data.

When discussing the fiercely competitive AI chatbot market, Aravind Srinivas candidly stated that ChatGPT, Anthropic Claude, Google Gemini, xAI Grok, Meta AI, and Perplexity are currently very similar in core question-answering abilities, heading toward homogenization. Perplexity's early differentiation lay in providing accurate and quick answers from sources, but now other players are catching up.

He believes that future differentiation will come from two aspects:

1. Richer interaction forms: Not merely text responses, but able to directly provide charts, images, product cards (such as hotels, shoes), and support one-click purchases or bookings.
2. True agent behavior: AI that can proactively execute complex tasks based on personal context (emails, calendars, etc.), such as rescheduling meetings, ordering food, calling a car, handling emails, etc. “You lie in bed sleeping, while AI works for you.” This requires strong reasoning abilities and deep integration with various service APIs, greatly increasing the complexity of product building, thereby reducing homogenization competition.

Regarding the model itself, he revealed that Perplexity’s backend calls different models for various sub-tasks (such as query rewriting, text chunking, summarization, recommending new questions) and optimizes engineering to ensure overall speed still leads.

Opportunities for India, Personal Future, and Thoughts on AI Governance

As a successful Indian-origin founder in Silicon Valley, Aravind Srinivas provided clear recommendations for India's role in the AI era. He strongly believes that India should train its own large models, not just to support Indian languages, but to compete on a global benchmark, inspiring the next generation of engineers and entrepreneurs. For resource-limited young people, he suggests starting by building attractive products to gain users and funding, then gradually moving toward training their own models.

He pointed out several fields where India might hold advantages:

• Voice AI: Voice recognition and synthesis for numerous Indian languages, dialects, and accents is a market of huge local demand that Western labs do not prioritize.
• Platforms for creating personalized applications: As AI significantly lowers the barriers for software creation, everyone in the future will be able to customize exclusive applications (like personalized fitness plans, payment-splitting tools, company internal holiday trackers) for themselves or small groups. A platform that abstracts complexities like deployment, data management, etc., will have great potential.

Looking ahead to the next five years, Aravind Srinivas predicts that everyone will have a powerful personal AI assistant that will become as ubiquitous as iPhones, rather than a luxury. This will unleash tremendous creativity, allowing people to turn any idea into reality. However, in the short term, the issue of labor replacement will be very severe, and society needs to think about how to help people adapt and upgrade their skills.

Regarding AI regulation, he maintains a cautious attitude, believing that regulating the models themselves is impractical and difficult to enforce. He prefers to regulate specific application scenarios, such as paying special attention to the psychological dependency and health risks that may arise from children using AI chatbots. He advocates for maintaining an open and accelerated development attitude in the early stages of AI development, avoiding excessive regulation that stifles innovation potential, leading to substantial long-term opportunity costs.

At the end of the interview, Aravind Srinivas demonstrated his consistent humility and learner's mindset. He shared what motivates him—his genuine love for learning new things and tackling challenges, rather than merely pursuing rewards. He also emphasized the stabilizing and motivating role of family support (from parents and wife) at various critical stages of his life. This perspective that combines personal growth, technological ideals, and humanistic care may be the distinctive charm of this AI leader.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。