Aakash GuptaHow To ACE AI Product Sense Interviews (OpenAI PM Mock Interview)
CHAPTERS
- 0:00 – 1:16
Why AI Product Sense interviews are suddenly essential (and what this video will deliver)
Aakash explains the gap in existing PM interview prep: lots of generic product-sense content, but almost nothing specific to AI Product Sense interviews now used by top AI labs. He previews the structure of the episode: a live mock interview, a teardown, and actionable tactics to replicate the approach.
- •AI Product Sense is a 45-minute “speed run” of the PM process tailored to AI products
- •Top AI companies (OpenAI, Anthropic, Google, Meta) use this interview format
- •Video promise: live mock + analysis + how-to guidance
- •Quick plug for cohort-based coaching and expert involvement
- •Sets expectations for what a strong AI PS answer looks like
- 1:16 – 3:57
The case prompt: double ChatGPT image-generation WAUs in 3 months with 3 engineers
The interviewer (Dr. Bart) reveals the core challenge: grow weekly active users of ChatGPT image creation from 175M to 350M in three months with only three engineers. Constraints and the specific product surface (still images, not video) are clarified before Aakash begins structuring.
- •Goal: 175M → 350M weekly active image-generation users
- •Timeline: 3 months; Resourcing: 3 engineers
- •Clarifies the feature includes prompting for images and the “Create Image” entry point
- •Notes limited onboarding/discoverability today
- •Confirms scope excludes Sora/video—still images only
- 3:57 – 5:53
Aligning on scope, mission, and what ‘3 engineers’ really means
Aakash grounds the work in OpenAI’s mission and clarifies technical/organizational boundaries: app/product engineering versus improving the underlying image model. The conversation sets expectations on what can realistically move in three months while still influencing the research org via feedback loops.
- •Connects image creation to multimodal AGI strategy
- •Separates app-layer work from model/research improvements
- •Confirms ‘3 engineers’ are product/application engineers, not model researchers
- •Plans to “shape the mountain” while research can “move mountains”
- •Keeps feasibility and timing constraints front-and-center
- 5:53 – 10:17
Live framework build: six workstreams to drive growth under constraints
Aakash lays out a structured approach on a Miro board to guide the interview from strategy to execution. The framework explicitly includes users, focusing, problems, solutions, metrics, and guardrails—designed to show clear PM thinking rather than jumping to features immediately.
- •Framework: mission/strategy → key users → focus segment → user problems → solutions → metrics → safety/guardrails
- •Uses a collaborative “co-thinking” whiteboard style
- •Frames the interview as a navigation tool to avoid rambling
- •Signals intent to tie everything back to the WAU goal
- •Prepares to iterate the framework based on interviewer feedback
- 10:17 – 15:52
User segmentation and growth math: where the next 175M users must come from
Using guesstimates and known platform scale, Aakash segments current image users and reasons about where incremental growth is realistically available. He argues the biggest growth lever is converting lower-tech-literacy ChatGPT users, not squeezing more from early adopters.
- •Segments current 175M into early adopters, tech-forward users, and low-tech-literacy users
- •Learns ChatGPT overall WAUs are ~800M; ~20–25% generate images weekly
- •Uses population math to argue tech-forward ceiling is limited
- •Infers biggest opportunity is low-tech-literacy users and better conversion/onboarding
- •Chooses a focus segment explicitly to match the growth target
- 15:52 – 19:27
Diagnosing friction: UI/UX issues uncovered by an end-to-end demo
Aakash runs a quick, real usage flow to surface practical friction points that could depress activation and repeat usage. He highlights discoverability problems, confusing “thinking” behavior, and perceived latency—especially important for less technical users.
- •Entry points are unclear: accidental discovery or buried behind a plus button
- •No dedicated, obvious image-first experience (unlike Sora)
- •Perceived speed issues and confusing time counter behavior
- •“Afraid to click thinking” and unclear progress feedback
- •Emphasizes observing real flows as a fast stand-in for research
- 19:27 – 23:13
Competitor scan: what Midjourney, NanoBanana, and Synthesia reveal about gaps
Prompted by the interviewer, Aakash maps competitive strengths to product gaps and candidate improvement areas. He contrasts aesthetic quality/style leadership (Midjourney), editing workflows (NanoBanana), and realistic avatar/likeness capabilities (Synthesia) to inform where ChatGPT image creation underdelivers.
- •Midjourney: superior quality/styles; implies model/style improvements needed
- •NanoBanana: strong selection and editing UX; highlights workflow/productivity gap
- •Synthesia/avatars: realistic likeness and “looks like me” expectations
- •Uses competitors to accelerate problem discovery rather than feature copying
- •Captures insights as inputs to both product UX and research feedback
- 23:13 – 24:55
How to learn what users are really doing: prompt taxonomy + feedback signals
Aakash explains how he would identify the top user jobs-to-be-done using behavioral data: clustering prompts into categories and pairing them with satisfaction signals. This creates a practical prioritization input that can guide both UX work and model-improvement requests.
- •Analyze prompt logs to identify the top ~100 image-generation use cases
- •Build a taxonomy (e.g., anime, photorealism, infographics, portraits, thumbnails)
- •Overlay thumbs up/down by category to find biggest pain points
- •Turn findings into actionable feedback for the research/model team
- •Uses data to avoid opinion-led prioritization
- 24:55 – 30:13
Solution brainstorming: discoverability, loading delight, editing, and use-case tooling
Aakash ideates solutions tied to the surfaced problems, spanning UI discoverability, loading/“thinking” improvements, and functional editing capabilities. He also proposes building lightweight, use-case-specific creators (infographics, thumbnails, memes) to unlock broader adoption.
- •Increase image feature prominence on the ChatGPT homepage
- •Improve loading/thinking UX using game-design principles for wait states
- •Fix misleading time estimates and progress feedback
- •Add selection + image/text editing (Canva-like, not Photoshop)
- •Create templates/tools for infographics, thumbnails, profiles, and memes
- 30:13 – 32:10
Prioritization under tight constraints: sizing impact vs. effort to reach 350M WAUs
With limited engineers and time, Aakash shifts into an impact/effort discussion and sketches a simple “above the line/below the line” cut. He estimates which initiatives could plausibly drive tens of millions of users fast, and which are too costly for the 3-month window.
- •Re-centers on constraints: +175M WAUs, 3 engineers, 3 months
- •Assumes fast iteration on front-end/UI changes; prioritizes quick wins
- •Deprioritizes higher-effort items (e.g., likeness training) given limited impact/time
- •Highlights editing as high impact but non-trivial effort
- •Uses rough sizing to justify a feasible bundle of initiatives
- 32:10 – 34:40
Curveball: ‘Build an Instagram killer in 3 months’—MVP framing and viral loop
Dr. Bart challenges Aakash to imagine shipping a dedicated image app within the same timeframe. Aakash reframes it as an authenticity-first AI photo editor, narrows to a core user group, and identifies viral sharing dynamics—then returns to objective prioritization.
- •Reframes from ‘ImageGPT’ to an authenticity-focused photo editing product
- •Prioritizes AI photo editing features: zoom-out/background fill, resolution upscaling, filters
- •Suggests a narrow MVP target segment to ship quickly
- •Leans on social sharing/word-of-mouth as the growth engine
- •Explicitly returns to the main prioritization stream afterward
- 34:40 – 38:00
Final bundle + metrics linkage: a feasible roadmap to bridge the WAU gap
Aakash consolidates the selected initiatives into a coherent package and sanity-checks whether they sum to the growth target within the resourcing constraint. He connects the roadmap to measurable outcomes and acknowledges uncertainty while keeping momentum to the finish line.
- •Combines UI/discoverability + editing + use-case editors to approximate the +175M goal
- •Treats some initiatives as quick (weeks) and others as longer (1–1.5 months)
- •Uses a back-of-the-envelope tally to justify feasibility with 3 engineers
- •Introduces prioritization+metrics as a combined section for speed
- •Aligns with interviewer preference: the journey and reasoning matter most
- 38:00 – 42:09
Solution specification details: what ‘editing’ and ‘infographic/meme editors’ mean in practice
Before closing, Aakash adds concrete definitions so the roadmap is actionable rather than abstract. He clarifies lightweight editing capabilities, how infographic generation would work on a canvas, and what a meme editor would include—while staying mindful of time.
- •Homepage changes to improve discovery and activation
- •Fix time counter/progress experience and improve thinking UX
- •Editing scope: selection/lasso + basic adjustments (Canva-like)
- •Infographic editor: multi-image generation + editable text boxes on a canvas
- •Meme editor: templated generation and text overlays inspired by existing tools
- 42:09 – 44:16
Guardrails, safety, and copyright: red-teaming, likeness controls, and ‘Ghibli’ lessons
Aakash explicitly covers trust-and-safety concerns that become critical when scaling image creation quickly. He argues for proactive copyright and abuse prevention, citing how viral growth can backfire through user backlash and low ratings if restrictions arrive late.
- •Need verification/controls to prevent misuse of third-party likeness
- •Prevent pornographic, abusive, or dangerous content; emphasize red-teaming
- •Addresses copyright sensitivity after ‘Ghibli’ controversy
- •Recommends being more copyright-safe from the start (vs. ‘bold now, sorry later’)
- •Uses Sora review/rating outcomes as a cautionary signal
- 44:16 – 48:33
Interview feedback and breakdown: what worked and what to improve
Dr. Bart evaluates Aakash’s performance, praising the clarity of the narrative, structured discovery, and ability to handle curveballs without losing the thread. He also notes what could be improved in a real interview: more explicit collaboration assumptions, internal research, and leveraging existing backlog work.
- •Strengths: end-to-end product discovery journey, strong guesstimation, clear narrative
- •Handled curveballs and returned to main line of reasoning
- •Visual structure (Miro) improved followability and communication
- •Improvement ideas: be explicit about validating estimates with engineering/team
- •Recommend internal research: what’s already known, shelved work, or backlog “bangers”
- 48:33 – 52:12
Three takeaways to apply in your own AI Product Sense interviews
Aakash closes by extracting practical lessons viewers can copy: build a custom framework, continuously check in and adapt to interviewer signals, and strategically sell your unique fit while solving the case. Bart reinforces that asking for feedback is rarely a losing move.
- •Create a unique framework tailored to the prompt (don’t reuse a generic template)
- •Check in mid-interview and adapt your structure to interviewer feedback
- •Add missing sections (e.g., ‘what users are solving,’ metrics) as the conversation evolves
- •Sell your candidate-market fit by tying answers to your experience and strengths
- •Ask for feedback even if interviewers may not always offer it proactively
