Aakash GuptaIf This 81 Minute Video Doesn't Make You an AI PM, I'll Delete My Channel
CHAPTERS
- 0:00 – 1:29
Why AI PM is the new baseline (and why most roles won’t be labeled ‘AI PM’)
Aakash and Ankit set the stakes: PMs who leverage AI will outpace those who don’t. They explain why AI PM roles often hide inside regular PM job descriptions—and why "every PM job is becoming an AI PM job" through tool usage and AI feature delivery.
- •AI creates a widening performance/comp gap between AI-enabled PMs and everyone else
- •Many companies won’t post “AI PM” titles, but responsibilities include LLM/RAG/evals
- •AI PM work exists globally via applied AI in existing products
- •PMs must use AI tools to stay competitive, even if not building AI models
- 1:29 – 5:21
Compensation reality check: US vs. India AI PM pay premiums
They share concrete salary deltas showing AI PM compensation materially higher than traditional PM compensation. The discussion highlights how the premium scales with location, seniority, and the leverage companies get from AI-focused talent.
- •US: AI PM base comp outpaces comparable PM roles, especially in HCOL hubs
- •India: AI PM ranges can jump from ~25–30 LPA to ~45–65+ LPA
- •Top-tier AI PMs at frontier companies can reach extremely high total comp
- •Premium extends beyond PM into adjacent AI-leverage roles (growth, operators, engineers)
- 5:21 – 8:59
Defining AI PM paths: AI-enabled PM vs. applied AI PM vs. core AI PM
Ankit lays out a taxonomy of AI PM roles and clarifies a common misconception: most opportunity sits in application-layer AI rather than infra/model-layer roles. He frames applied AI PM as the accessible, high-demand path for most PMs.
- •AI-enabled PM: everyone must use AI tools for productivity
- •Applied AI PM: builds AI-powered features/products on top of existing models/infra
- •Core AI PM: works on infra/cloud/models; often requires strong ML/DS background
- •Most value accrues in applications (analogy: UPI infra vs. fintech apps)
- 8:59 – 12:14
What doesn’t change: PM fundamentals that keep the role ‘sticky’
They argue that AI won’t replace product management’s core human and business functions. The enduring edge is user empathy, problem-solving, and stakeholder/customer/partner management—especially as teams get smaller and execution speeds up.
- •Mental model: focus first on what won’t change, then layer new skills
- •Core PM strengths: user empathy, problem solving, stakeholder management
- •AI shrinks teams and time-to-value, shifting PM focus to external stakeholders
- •Distribution/partnerships become more important as execution accelerates
- 12:14 – 16:17
AI-powered product development lifecycle: from inputs → roadmap → execution → GTM
Ankit walks through a standard product development lifecycle and shows where AI compresses cycle time—research, analysis, communication, prototyping, and execution. He emphasizes disciplined validation and prioritization to avoid “shiny object syndrome.”
- •Idea sources: business goals/OKRs, market trends, competition, partners, stakeholders, data
- •PM must translate “feature requests” into validated problems
- •Roadmap thinking prevents thrash and reactive prioritization
- •AI accelerates research, analytics, stakeholder comms, and rapid prototyping
- 16:17 – 22:50
Tooling leverage (and the ‘don’t outsource your brain’ warning)
They highlight practical AI tooling wins (analytics, Jira querying, rewriting communications, prototyping) while warning against letting AI replace critical thinking. PRDs and planning are partly valuable because they force structured reasoning and context-building.
- •Examples: Mixpanel AI, Atlassian Intelligence (natural language queries), doc summarization
- •AI prototyping (Lovable/V0/Cursor/Windsurf) speeds validation with real users
- •Company enablement matters (security/policy + leadership support)
- •AI drafts are fine; PM must add context, judgment, and craft to reach high-quality outcomes
- 22:50 – 27:01
Choosing the right AI approach: predictive AI vs. generative AI
Ankit breaks AI solutions into predictive vs. generative, stressing that generative AI isn’t always the right tool. They outline classic predictive use cases and explain why they remain powerful and often cheaper and more reliable than genAI.
- •Predictive AI: ranking, recommendations, anomaly detection, categorization
- •These use cases remain highly valuable and are not “obsolete” in the genAI era
- •GenAI has higher cost and different failure modes; avoid retrofitting it everywhere
- •Transformer advances improved predictive understanding and enabled today’s genAI
- 27:01 – 36:17
Generative AI use cases + how to build your personal ‘AI use case database’
They define generative AI as contextual content generation across text, code, images, audio, and video. Ankit gives a concrete learning workflow: study customer stories from model providers and scan Product Hunt to compile reusable use-case patterns.
- •GenAI outputs: summaries, writing, code, images, audio/video
- •Learning method: review Anthropic/OpenAI/Gemini customer stories for real implementations
- •Create a personal repository of AI use cases to map problems → solutions faster
- •Use Product Hunt’s launch archives to see what’s working in the market
- 36:17 – 40:52
The AI PM checklist: from problem hypothesis to model choice, data, cost, UX, and evals
Ankit presents a practical checklist for what AI PMs actually do day-to-day. It bridges classic PM work (hypothesis, UX, launch) with AI-specific concerns (model choice, contextualization, pipelines, token costs, evaluation/iteration).
- •Non-negotiable: deep problem understanding + clear product hypothesis
- •AI-specific: model selection (fine-tune vs. “thinking” models), data sourcing, pipelines
- •Business constraints: token/infrastructure costs and scalability tradeoffs
- •Critical capability: evaluation and iteration to manage non-deterministic behavior
- 40:52 – 46:26
Don’t start with ML math: use Marty Cagan’s 4 risks (value, usability, feasibility, viability)
He reframes learning priorities using the value/usability/feasibility/viability model and warns against getting stuck in premature technical depth. PMs should prioritize value and viability first while learning just enough AI fundamentals to collaborate effectively.
- •Four angles: valuable, usable, feasible, viable
- •Common mistake: over-index on ML/statistics early and burn motivation
- •PM advantage: define value/viability and guide the product toward success
- •Zero-to-one: prioritize validation over premature optimization (including token costs)
- 46:26 – 55:52
Contextualization 101: prompt engineering vs. RAG vs. fine-tuning
Ankit explains how AI products become useful by adding context and differentiates the three main methods. He outlines when each approach fits, tradeoffs (cost, real-time updates), and why most products start with prompting and progress to RAG.
- •Contextualization unlocks product value (e.g., Notion + LLM, Stripe docs assistant)
- •Prompting: simplest; works when context fits in the prompt/template
- •RAG: retrieval + embeddings to fetch only relevant chunks; scalable and more real-time
- •Fine-tuning: best with lots of examples; higher cost and not inherently real-time
- 55:52 – 58:25
LLM limitations & why AI evals become a core PM skill
They cover hallucinations, bias, and non-determinism—and why evaluation is essential to ship reliable AI. Ankit distinguishes offline vs. online evals and positions online evals as ongoing product analytics for model quality in production.
- •Key limitations: hallucination, bias, indeterminism, overconfident wrong answers
- •Evals are the new ‘tests’ for AI systems that aren’t deterministic
- •Offline evals: pre-launch quality gates; online evals: production monitoring + iteration
- •Eval criteria examples: factuality, structure/format (e.g., JSON), latency, safety/bias
- 58:25 – 1:03:11
Worked example: building an AI job site + using an eval model-in-the-loop
Using an AI-first job board concept, Ankit shows where failures occur (bad summaries, nonsense interview questions, broken JSON) and how evals catch issues. He demonstrates an evaluator prompt, dashboard approach, and iteration loop across prompts/RAG/data.
- •Pipeline idea: crawl jobs → enrich with summaries, interview questions, learning guides, quizzes
- •Typical failure modes: hallucinations, poor relevance, invalid structured outputs
- •Approach: evaluator prompt + stronger model to score against a checklist
- •Iterate by diagnosing whether the issue is output, eval, prompt, RAG retrieval, or data
- 1:03:11 – 1:11:15
From copilots to agents: action-taking systems, automation stacks, and MCP integration
They move from generative outputs to agentic workflows: LLM + tools + autonomy. Ankit gives a concrete automation example (screening podcast invites) and then introduces Model Context Protocol (MCP) as the emerging standard to connect LLMs to external systems safely.
- •AI agent = LLM reasoning + tool access + autonomy to take actions
- •Example: auto-triage Gmail invites, fetch YouTube stats, send correct Calendly link
- •Tools/platforms: agents.ai (beginner), Zapier/n8n; deeper: LangChain/LangGraph
- •MCP standardizes connecting LLMs to APIs; example: Razorpay MCP for payments with guardrails
- 1:11:15 – 1:20:42
How to actually land AI PM roles: job description mining → portfolio ladder → targeted outreach
Ankit closes with a tactical roadmap to become employable: analyze 20+ AI-tinged PM JDs, derive a skill checklist, and produce proof-of-work. He proposes a portfolio ladder from writing to teardowns to side projects, and a ‘small-company’ outreach hack to bypass heavy gatekeeping.
- •Start by mining real JDs to build your learning roadmap and interview readiness
- •Portfolio ladder: writing → product teardowns → improvement proposals → side projects → past work
- •Side projects must show synthesis + user feedback iterations, not just a generated demo link
- •Outbound hack: pick 15–20 smaller AI companies, propose what you’d do in 6–8 months, follow up 3x
