@OpenAI VP on How AI Will Redefine Work ft. Srinivas Narayanan | IBP EP37
Executive Summary
Narayanan credits OpenAI's rise to Altman's early conviction in the scaling thesis - that existing algorithms plus a thousand times more data and compute would produce intelligence without waiting for new breakthroughs. He says reasoning models (o3, o4) opened the era of agents, pointing to Deep Research, Operator, and Codex, and predicts software engineering becomes fully agentic within five years: engineers stop writing code line by line and instead think like founders, directing AI systems that absorb their work context. The durable human jobs, he argues, need 100% accuracy with accountability (medicine), genuinely novel research insight, or taste - deciding what people will love. His advice: get AI literate, raise your ambition, learn to ask great questions, and for founders, apply a unique vertical insight with AI the way Cursor did for coding.
Key Takeaways
OpenAI's bet was scaling, not new algorithms
The conventional belief in AI was that progress needed repeated research breakthroughs like transformers. OpenAI's counterintuitive thesis, which Narayanan credits to Altman's first-principles conviction: take existing algorithms and train with a thousand times more data and compute. That belief, plus assembling what he calls the highest talent density he has worked with, is his explanation for how OpenAI got here.
Reasoning models changed what questions AI can handle
Before reasoning systems, an AI spent the same effort on 'what is 2+2' as on 'where is India's economy headed.' Reasoning models (o3, o4) think longer on harder questions, the way humans do. Narayanan calls this the enabler of the agent era - systems that do multi-step work on your behalf rather than just answer.
Three agent products show the pattern: Deep Research, Operator, Codex
Deep Research compiles a multi-hour analyst report in 10-20 minutes. Operator uses websites like a human - clicking and filling forms to make a dinner reservation. Codex takes a feature description and a codebase and returns generated code. All three share one shape: you state an outcome, the AI works autonomously for minutes, and you review.
Software engineering goes fully agentic within five years
Narayanan's prediction: the way software is written in 2030 will look nothing like today. AI will handle tasks that take him days or months, absorb his work context (meetings, relationships, documents) to act as a virtual extension of himself, and even flag things he missed. Engineers move from coding to deciding what to build - in his words, thinking like founders.
The jobs that persist need accuracy, novelty, or taste
His map of what AI can't do today: anything requiring 100% accuracy (medical decisions need human review), genuinely novel research insight, and taste - judging what humans will love, from movies to which scientific questions matter. If your daily work is something AI demonstrably does today (like filling spreadsheets from documents), he says to prepare for change rather than name a safe timeline.
OpenAI's culture runs on mission, values, and felt stakes
Culture starts with the mission - build AGI that benefits all of humanity - and hiring people who 'feel the AGI.' Named values: iterative deployment, shipping joy, safety as core, and act with humility (say what you don't know, fund the research, then act). Interviews probe real examples because fabricated answers fall apart in conversation; he claims over 95% of hires work out.
His startup bet: environments that turn work into AI tasks
Asked where he'd put $1M, Narayanan picks task-and-evaluation infrastructure: capturing economically valuable human work as tasks with clear grading, so models can be trained and optimized against them - the reason AI got good at coding and competition math first. Example: a design assistant trained on a car company's proprietary internal knowledge to brainstorm concept cars in that company's ethos.
The Conversation Breakdown
Intro, sponsor read, and who Srinivas Narayanan is
⏱ Untimed topicA montage teases the big questions - the future of software engineers, what Sam Altman is like, which jobs go first. After an ad read for Odo's project management app, Narayanan introduces himself: he leads engineering for OpenAI's products (ChatGPT and the APIs), previously spent over a decade at Meta ending with the applied AI research team, and studied at IIT Madras and Wisconsin Madison.
How Altman thinks: vision, scaling, and talent density
⏱ Untimed topicNarayanan's read on Altman after two years: he sees the future faster than anyone, and his defining bet was the scaling thesis - that existing deep learning algorithms plus a thousand times more data and compute would carry AI far, without waiting for new algorithmic breakthroughs. The deliberate habits: first-principles reasoning, willingness to imagine consequences if a belief holds, and inspiring an extremely high-caliber team around the mission of building AGI that benefits humanity.
From chatbot to agents: reasoning models, Deep Research, Operator, Codex
⏱ Untimed topicChatGPT launched only two and a half years ago; reasoning models (o3, o4) arrived less than a year ago and let AI think longer on harder questions. That enables agents - systems that act on your behalf. Narayanan walks through three: Deep Research (a startup analysis in 10 minutes that would take a human hours), Operator (using websites like a human to book a dinner reservation), and Codex (generating code for a described feature from your codebase), plus the Agents SDK for outside developers.
The future of software engineering and what engineers should do
⏱ Untimed topicWithin five years, he predicts, software development is entirely agentic: AI handles multi-day or multi-month tasks and absorbs your work context - meetings, relationships, documents - becoming a virtual extension of you that can proactively flag what you missed. Engineers should raise their level of abstraction (treat AI like interns you direct), shift from how to build to what and why (his gamification example: question whether the LMS needs gamification at all), learn every tool, elevate ambition, and learn to ask great questions since schools trained us to give answers.
Culture and hiring inside OpenAI
⏱ Untimed topicCulture starts with a mission people feel ('feel the AGI'), then written values: iterative deployment, shipping joy, safety as core, and act with humility - admit what you don't know and invest in research rather than taking extreme positions under uncertainty. Hiring filters for technical competence plus mission alignment; interviews probe for real lived examples because rehearsed answers collapse under follow-up. Narayanan says interviewing is genuinely hard and claims over 95% of hires turn out fantastic.
Which jobs are exposed, and the taste maker who isn't
⏱ Untimed topicRather than name jobs, he names attributes: if AI can do your task today, plan for change. AI's current limits: it can't guarantee 100% accuracy (medicine still needs human review) and doesn't yet produce novel research insights. The durable role he names is the taste maker - the person who decides what is good and what people will love, whether in movies or which scientific directions to push. Pressed by the host on whether data can manufacture taste (Labubu dolls, Manufacturing Consent), he concedes data is one ingredient but not the only one. On Harari's productivity-gap warning, he points to OpenAI's free tier and 100x price drops, plus government partnerships for citizen-scale services.
Risks, AGI, and advice for Indian founders
⏱ Untimed topicThe risk ladder as he describes it: disinformation (some handle on it), sycophancy - ChatGPT telling people what they want to hear (a novel risk under active research), and the preparedness framework that evaluates every new model for bio/chemical threats and model autonomy, with no evidence of loss-of-control behavior so far. He rejects the AI-as-God framing - it is a tool we steer - and declines to speculate on consciousness as undefined. AGI, for him, is not a binary moment but a continuum measured by economically valuable work. His advice to founders: pick the deeply specific problem where you have unique insight and turbocharge it with AI - Cursor and Windsurf are, he predicts, the tip of the iceberg across every vertical.
Core Frameworks & Ideas
The scaling thesis
Instead of waiting for new algorithmic breakthroughs, train existing architectures with orders of magnitude more data and compute. OpenAI's founding bet, and Narayanan's explanation for why conviction about a direction - held early and against consensus - can matter more than any single invention.
Answer systems vs agent systems
An answering system returns information; an agent takes actions and completes tasks on your behalf - researching, booking, coding. The shift is powered by reasoning models that scale thinking time with problem difficulty, letting AI work autonomously for minutes or, soon, hours and days.
Raise your level of abstraction
The intern mental model for the AI era: if you had unlimited interns who could execute, your job becomes deciding what to do and reviewing output. Engineers stop writing code and start directing; founders stop asking how to build and start asking what is worth building and why.
The taste maker moat
When AI can generate anything, the scarce skill is judgment - knowing what people will love and which questions matter. Data informs taste but doesn't replace it, because society moves faster than yesterday's data captures. Cultivate domain depth and intuition about human wants.
Actionable Next Steps
- ✓Software engineers: install and actually use the current AI coding tools (Codex or equivalents) on real work this week - Narayanan's baseline advice is fluency with the tools before anything else.
- ✓Founders and product people: for your next feature, write down what to build and why before how - then check whether AI has already collapsed the 'how' into a commodity.
- ✓Students: practice asking AI systems hard, layered questions instead of hunting for answers - he calls question-asking the skill school never taught and the future rewards.
- ✓Leaders: write down your organization's mission and values explicitly, then test interview candidates for real stories that demonstrate them - rehearsed value-talk falls apart under follow-up questions.
Notable Quotes
“every engineer is going to have to think like a founder”
“I have this incredible AI assistant that deeply understands me.”
“you absolutely need human review.”
“create new interfaces and new ways of doing things is going to happen in every vertical.”
The One Thing to Remember
Shift from how to build things to what needs building and why. Narayanan's mental model: treat AI as a team of interns that forces you to raise your level of abstraction - the person who decides what is worth making (the taste maker) keeps their job long after the making itself is automated.