Spent my Sunday morning digesting Nikhil Kamath first WTF Online pod with Perplexity's Aravind Srinivas, and wow - the insights were gold! If you don't have 2.5 hours to spare, I've got you covered.
Even Aravind is trying to wrap his head around the definition of intelligence in the AI age and says that the current systems already encompass the skills of "some 10,000 knowledge worker professions in one system without hard coding". So that's a reality check for every one of us.
Beyond this, he gives a different take on why India should still develop its own models, tries to defend PPLX's moat, says the end of question-answer AI interface is around the corner, and that India's Big IT (TCS, Infy, etc.) might be in for a rough ride.
Also, really loved how Aravind gives some direction to the Indian AI ecosystem, including policymakers. With almost all global AI action and, consequently, mainstream discourse concentrated in Silicon Valley, conversations about the future of AI from an India PoV are rare. Hope to see more of Nikhil's conversations with global Indian founders exploring this intersection.
Future of AI (Next 5 Years) and Interactions:
- Highly personalized AI assistants will become common and affordable
- AI will replace human labor in meaningful ways in certain sectors
- Brand value and person-to-person connections will be even more critical since AI can filter info and ignore sponsored links. WhatsApp, Instagram are well-positioned.
- Google's ad biz model might face challenges with the rise of agents that can bypass ads and prioritize user needs over sponsored content.
Current Chatbot Landscape and Differentiation:
- No genuine differentiation between major chatbots today. Focus is shifting towards better user experiences and more agentic functionalities.
- Question answering in AI will become a commodity. Providing more comprehensive experiences, such as integrating charts, images, product cards, and direct transactional capabilities.
- Perplexity aims to differentiate itself through more agentic behavior and richer experiences beyond just text (e.g., charts, images, inline product cards).
India and AI:
- India should definitely train its own foundational AI models meeting global benchmarks to inspire the next generation of engineers
- Indian startups could start with novel AI products, then move towards building own models, starting with post-training on open-source models and eventually pre-training.
- Voice interfaces and support for Indian languages (speech recognition, synthesis, dialects, grammar) are big opportunities.
- Data centers within India are important for data sovereignty and could attract international companies looking to store Indian data locally.
- Indian IT giants will face pressure on pricing and timelines as AI becomes more capable.
AI Progress:
- Recent surge in capabilities is a combination of unprecedented amounts of compute, high-quality data, and RLHF, along with training on tasks useful for human labor.
- Simpler ideas, when given enough compute, often outperform more complicated ones in AI (OpenAI internship takeaway via Ilya)
- Next gen AI will focus on agentic behavior like booking reservations, managing calendars, and sending emails autonomously.
- Current LLMs may not be the direct path to AGI as they lack the intuitive understanding of the physical world that humans possess, a prerequisite for AGI.
- Access to compute will be a key differentiator, even if models become more broadly accessible through open source.
Data Centers and Compute:
- Data center business might get commoditized in the long run unless providers offer significant value beyond just infrastructure, such as integrated software for managing AI workloads.
- Nvidia's margins & dominance difficult to disrupt due to specialized chips, CUDA, and established relationships with hyperscalers. Alternatives are emerging, particularly at inference layer.
Regulating AI:
- Regulating models is impractical. Focus should be on regulating specific applications with potential for harm. (Eg. AI interactions with children).
- Avoid overly restrictive regulations in the early stages of AI development to allow for innovation and progress.
Aravind's Journey:
- Grew up in Chennai; cricket got him interested in statistics; competitive programmer at IIT-M during EE; Self-taught ML concepts during a BLR internship
- PhD at Berkeley led to an OpenAI internship in 2018. What Ilya said (simpler models, more compute) shaped his understanding of AI development.