Nikhil KamathNikhil Kamath ft. Perplexity CEO, Aravind Srinivas | WTF Online Ep 1.
CHAPTERS
- 0:00 – 4:37
Chennai roots to IIT: early interests in math, cricket stats, and coding
Nikhil opens by locating Aravind’s upbringing in Chennai and uses it as a springboard into his origin story. Aravind explains how cricket statistics nudged him toward numbers, then programming, and the expectations that pushed him toward IIT.
- •Grew up in Chennai; parents still live there
- •Early comfort with numbers via cricket statistics and math
- •Picked up programming around 11th standard
- •Family/IIT expectations and the competitive JEE environment
- •Entered IIT (Electrical), gravitated toward CS and competitive programming
- 4:37 – 7:21
First exposure to machine learning: Kaggle, scikit-learn, and a fast Bangalore internship
Aravind describes stumbling into ML through a Kaggle competition and learning by experimentation. That success led to a Bangalore startup internship where he quickly delivered a recommender-system solution and used the remaining time to self-study.
- •Learned ML via Kaggle contests and scikit-learn experimentation
- •Winning a contest became a confidence/inflection point
- •Bangalore startup internship building recommender systems
- •Finished internship work unusually fast; used time to learn more
- •Self-taught via Andrew Ng/Stanford material and campus ML courses
- 7:21 – 12:13
Berkeley and OpenAI: building fundamentals, humility, and the ‘simple ideas + compute’ lesson
Aravind recounts the path from research at Berkeley to internships at OpenAI/DeepMind and the culture shock of being surrounded by world-class peers. He highlights a formative OpenAI moment with Ilya Sutskever that shifted his thinking from “fancy” ideas to what works at scale.
- •Early Berkeley period without an advisor; intense work routine (Philz Coffee)
- •Paper led to advisor Pieter Abbeel and connections to OpenAI founders
- •OpenAI internship (2018): humbling gap vs. top researchers
- •Ilya’s framing: generative AI + RL, plus massive compute as a recipe
- •Practical takeaway: simplest scalable ideas often beat complicated ones
- 12:13 – 29:04
Defining AI, AGI, and superintelligence in plain language
Nikhil asks for a child-simple explanation of AI and intelligence. Aravind distinguishes narrow task-specific systems from general intelligence, and then separates AGI from the more autonomous, self-improving notion often called superintelligence.
- •AI as computers doing tasks that require ‘intelligence’ in a human-like way
- •Narrow AI (e.g., chess engines) vs. systems that generalize across tasks
- •AGI as a highly capable general digital worker (practical definition)
- •Superintelligence framed as autonomy + recursive self-improvement
- •Why today’s models lack self-awareness and self-directed goal formation
- 29:04 – 34:44
From calculators to modern computing: personal computers, networks, cloud, and the path to AI
To build intuition, Nikhil starts from calculators and asks what happens “under the hood.” Aravind explains circuits and the evolution from mainframes to PCs, then to the internet, mobile, cloud, and finally today’s AI wave.
- •Calculators as circuits (adders/multipliers) parsing input to outputs
- •Mechanical intuition for computation (binary counters, simple circuits)
- •PC revolution: democratization of compute (Moore’s Law + packaging)
- •VisiCalc/spreadsheets as a killer app for early personal computing
- •Internet → web → mobile → cloud as stepping stones to AI ubiquity
- 34:44 – 35:53
Why AI ‘worked’ in the 2020s: neural nets at scale, data quality, and RLHF
Aravind pinpoints the 2010→2020s shift as neural networks finally working at scale, driven by compute and high-quality data. He adds that RLHF and training for useful human tasks turned raw models into products people can actually use.
- •Neural networks ‘working’ became the big practical change vs. 2010 era
- •Scale matters: huge compute + large, curated datasets
- •RLHF (learning from human feedback) improves usefulness and alignment
- •Reasoning improvements supported by step-by-step data and “chain-of-thought” style supervision
- •Chatbot interfaces made advanced AI accessible to the public
- 35:53 – 48:30
Neural networks, machine learning, and LLMs: an intuitive walkthrough (with stock-market pitfalls)
Nikhil asks for a ground-up explanation of neural networks, why they didn’t predict markets well, and how ML differs from neural nets. Aravind explains neural nets as nonlinear functions trained via loss minimization, then positions LLMs as giant neural nets trained to predict the next token, with transformers and post-training turning them into chatbots.
- •Neural nets as layered ‘artificial neurons’ approximating complex nonlinear functions
- •Training loop: prediction vs. target, loss computation, backprop weight updates
- •Why finance prediction fails: irreducible noise + weak signal in limited inputs
- •Machine learning as the umbrella; neural nets as one scalable ML approach
- •LLMs: next-token prediction at internet scale; transformers; pretraining vs. post-training
- 48:30 – 1:05:11
Beyond text: physical common sense, robotics, and why ‘agentic’ requires reasoning
Reacting to Yann LeCun’s skepticism, the conversation shifts to what current models lack—physical grounding and tool use. Aravind explains why robotics generalization is hard, why data is scarce vs. the internet, and why reasoning/planning is key to moving from chat to action.
- •LeCun’s AGI bar: physical common sense, dexterity, tool use
- •Robotics needs repeated trials/simulation; generalization remains fragile
- •Humans benefit from evolutionary ‘priors’ in physical skills
- •Agentic systems require planning and reasoning, not just text completion
- •Learning from video/audio plus building internal “mental models” for new scenarios
- 1:05:11 – 1:12:44
How AI products will differentiate: from similar chatbots to assistants that act
Nikhil asks how Grok, Meta, Google, OpenAI, Anthropic, and Perplexity truly differ. Aravind argues that today’s chatbots are converging, and the next differentiation will come from integrated experiences (cards, charts, booking) and agents that complete tasks end-to-end with personal context and integrations.
- •Most chatbots feel similar because they optimize for the same benchmarks
- •Perplexity’s early differentiation: sourced, research-focused answers
- •Upcoming moat: ‘agentic behavior’—AIs that do tasks (book, email, calendar)
- •Personal context + integrations + voice UX become product differentiators
- •UI beyond text: charts, images, shopping/hotel cards, transaction flows
- 1:12:44 – 1:19:28
Inside Perplexity’s stack: multi-model pipelines, speed, and unit economics pressures
Aravind details how Perplexity routes a single query through multiple models for different sub-tasks, and why latency and infrastructure matter at scale. He also explains why cost-per-query is a moving target as model prices fall while “deep research” and future agentic tasks can raise serving costs.
- •Per-query pipeline: query rewrite, page chunking, summarization/chat, follow-up suggestions
- •Speed advantage comes from infrastructure (serving, runtimes, fast indexing)
- •Streaming answers reduces perceived latency; tail latency matters operationally
- •Costs are dynamic: new open-source models force API price drops
- •Deep research and agentic tasks increase compute cost; prioritize experience over margins
- 1:19:28 – 1:49:39
Competing with Google and Meta: distribution moats, ads, and transaction completion
The discussion turns strategic: why Google remains the default starting point and how ads and distribution reinforce dominance. Aravind argues that to truly disrupt, AI must not only answer but also help users transact natively, while startups must fight UX defaults, Android leverage, and network effects.
- •Google’s moat: default search placement, browser habits, Android/Play Store leverage
- •Research often happens in AI tools but transactions still route through Google/Amazon
- •To beat Google, AI must complete purchases/transactions within the assistant flow
- •Meta’s moat is network effects; hard cold-start for new social platforms
- •TikTok lesson: new content unit + massive spend + retention can overcome skepticism
- 1:49:39 – 1:58:31
India’s role: models, compute, voice-first opportunities, and what young founders should build
Nikhil asks how India can participate meaningfully rather than watching from afar. Aravind advocates training competitive foundation models, but also points to more accessible opportunities like Indian voice and language support, plus a staged path from products to post-training to pretraining and infrastructure.
- •India should train its own globally competitive models (not only local-language focus)
- •Compute access is the biggest constraint; data centers and chips matter
- •Advice for early-stage founders: build product → get users → raise → post-train → pretrain
- •Low-hanging wedge: speech recognition/synthesis for Indian languages, accents, dialects
- •AI impact on Indian IT services: fewer hires, pricing pressure, faster delivery expectations
- 1:58:31 – 2:14:21
Where AI is headed: personal assistants, personalized software, labor displacement, and regulation
Aravind predicts mainstream personal assistants and a surge of personalized apps people build for themselves, enabled by easier software creation and deployment platforms. He also addresses societal downside (job displacement, unhealthy chatbot relationships) and argues regulation should focus more on risky applications than on models themselves.
- •Near-term future: affordable personal assistants for everyone
- •Personalized app creation as a major new platform opportunity; sharing and monetization unclear
- •Economic shift: fewer humans needed for some work; pressure on entry-level jobs
- •Regulation focus: applications and harms (especially kids/companionship dynamics) over model bans
- •Open source spreads capability; compute remains the least democratized resource
- 2:14:21 – 2:16:30
Closing: internship invite, learning mindset, and final reflections on access
Nikhil ends by asking to intern at Perplexity to reduce his ‘FOMO’ and learn firsthand. Aravind agrees, while also emphasizing that access is increasingly online—time spent experimenting with tools and learning from top thinkers matters more than geography.
- •Nikhil’s request to intern at Perplexity for hands-on learning
- •Aravind’s view: physical proximity matters less; the internet contains most knowledge
- •Real learning comes from deep tool usage, understanding failure modes, and conversations
- •Importance of staying curious and asking “basic” questions
- •Friendly wrap-up and plan to meet when Aravind visits India
