Aakash GuptaThe OpenAI PM Who Helped Grow ChatGPT to 900M Users Demos His Actual Setup
CHAPTERS
- 0:00 – 5:34
Why Codex changes the PM job: from collaborator to agent
Abhi explains how ChatGPT evolved from a chatbot to a collaborator and now—via Codex—an agent that can complete end-to-end tasks. He frames the impact in two ways: automating repetitive PM work and enabling PMs to build meaningful prototypes/features without deep engineering fluency.
- •Codex as an “agent” that can take on parts of a PM’s job end-to-end
- •Time savings on recurring work (updates, reviews, follow-ups)
- •New capability: PMs building features to 70–80% before handing to engineering
- •Shift from “writing docs” to shipping functional prototypes
- •Empowerment and team inspiration from PM-as-builder
- 5:34 – 8:31
International growth dashboard: unifying scattered metrics into one daily view
Abhi describes an internal pain point: international growth insights were fragmented across many dashboards and tools, making synthesis slow and noisy. He introduces a single web app dashboard that centralizes country performance, highlights strengths/risks, and updates automatically each morning.
- •International growth requires synthesizing across funnel layers and competitors
- •Data lived across 7–8 sources (Databricks, Tableau, etc.) with different cadences
- •Dashboard provides country switching, headline metrics, and peer comparisons
- •LLM-generated strengths/risks and prioritized takeaways reduce cognitive load
- •Automated daily refresh makes the dashboard continuously useful for the team
- 8:31 – 9:32
General principle: synthesis + TL;DR is the real unlock (not just a prettier dashboard)
Aakash challenges why this can’t stay in Databricks alone. Abhi explains the key value: bringing multiple sources into one place and layering an interpretable summary of what matters—something traditional dashboards don’t provide.
- •Decision-making often needs multiple sources, not a single dashboard
- •Centralization reduces context switching and cognitive overload
- •LLM layer adds interpretation: key takeaways and what to focus on
- •Peer set selection and benchmarking add actionable context
- •Dashboards show data; agents can synthesize and recommend
- 9:32 – 14:52
How he built it in Codex: define inputs/outputs, iterate with previews and Playwright
Abhi walks through his build process: he tells Codex the desired user experience (outputs) and the data connectors (inputs), then lets it generate the app. He iterates rapidly using local previews and Playwright-based UI smoke tests to fix styling and layout issues.
- •Prompt structure: clearly specify outputs (tabs, stats, strengths/risks) and inputs (connectors)
- •Codex scaffolds the app, generates demo data when needed, and runs smoke tests
- •Local run + web preview for rapid iteration
- •Playwright screenshots help Codex self-diagnose UI problems
- •In-app preview reduces tool switching for end-to-end building
- 14:52 – 16:38
From PRDs to prototypes (with a companion FAQ spec)
Abhi explains he increasingly skips traditional PRDs in favor of building prototypes that stakeholders can react to visually. He still writes a lightweight companion document—more like an FAQ—to cover hypotheses, metrics, guardrails, and compliance considerations.
- •Prototype becomes the primary artifact; document becomes supporting material
- •Companion doc answers the top questions a PRD would cover
- •Faster stakeholder alignment through something tangible
- •Reduces wasted cycles and improves discussion quality
- •Maintains rigor (hypothesis/metrics/guardrails) without heavy documentation
- 16:38 – 18:17
PM workflow pre-AI vs now: faster alignment without burning design/engineering time
Abhi compares the old workflow (hypothesis → PRD/spec → design iterations → engineering) to the new one where PMs can generate an initial end-state mock/prototype themselves. This accelerates feedback loops and makes conversations more concrete earlier.
- •Classic lifecycle still applies, but prototyping compresses the early phases
- •PM can produce a first-pass experience for others to react to
- •Saves scarce designer/engineer bandwidth for higher-leverage work
- •Engineering discussions shift toward feasibility and iteration vs imagining from text
- •Prototypes function as alignment accelerators
- 18:17 – 20:04
Mechanics of building “80% to a PR”: working locally vs in the main repo
Abhi explains when he prototypes locally (internal tooling) versus when he pulls the main ChatGPT repo and targets real components. He shares a key technique: ask engineers for the closest existing reference so Codex can build on a known pattern and avoid codebase wandering.
- •Local prototyping for internal-only tools that don’t need core infra integration
- •For product features, pull the main repo and build on existing components
- •Critical tactic: identify the most similar existing implementation as a reference
- •Codex helps get to a pull request; engineers then review/finish the last mile
- •Overall effect: shorter path from idea to reviewable code
- 20:04 – 21:09
Codex use cases + honest limitations: context, ambiguity, and data-table confusion
Abhi buckets his Codex usage into repetitive automation and net-new building. He also calls out failure modes: separating signal from noise in Slack/updates and ambiguity across similar metrics/data tables (e.g., consumer vs enterprise WAU).
- •Two work buckets: automation of repetitive tasks vs building net-new tools
- •Slack/data synthesis still needs human review due to signal-to-noise issues
- •Ambiguous prompts can yield “correct but wrong-for-your-intent” answers
- •Need precision: specify which dashboards/tables/definitions to use
- •Best practice: generate drafts for PM review rather than auto-sending externally
- 21:09 – 25:59
Daily PM automations: Slack inbox triage, dashboard refresh, and weekly updates
Abhi demonstrates automations that run before his day starts, centered on Slack. These include a daily triage summary of unread/unreplied messages, an automated morning dashboard refresh, and an auto-drafted weekly stakeholder update assembled from multiple sources.
- •Slack is the operating system; volume makes misses inevitable
- •Daily triage automation highlights important people, threads, and needed responses
- •Automated dashboard run every morning keeps metrics current without manual work
- •Weekly update automation pulls from Slack, Drive/Notion, and dashboards
- •Human-in-the-loop review remains important for quality and nuance
- 25:59 – 28:27
Harness and skills: scaling Codex via reusable workflows (Statsig experiment reviews)
Abhi highlights the “Codex harness” as the major unlock beyond the raw model. He shares a team-created skill that automates experiment reviews end-to-end—from hypothesis through monitoring and recommendations—by connecting to Statsig and generating summaries for review meetings.
- •Harness (connectors + workflows) is a key differentiator for real productivity gains
- •Reusable skills can encode team processes (not just one-off chats)
- •Experiment review skill: writes hypothesis, monitors results, drafts postmortem, suggests decision
- •Evals/metrics rigor becomes easier to operationalize across a growth org
- •Best builder is “whoever cares most,” not a specific job function
- 28:27 – 33:07
Personal ‘computer use’ demo: catching up on WhatsApp and extracting action items
Abhi shifts to personal usage, showing Codex’s computer-use capability with WhatsApp Desktop. He asks it to scan message backlog and summarize what’s actionable, mirroring the same synthesis workflow he uses at work.
- •WhatsApp is central for communication in many international markets
- •Codex computer-use reads the desktop app to identify actionable threads
- •Reduces overwhelm from high-volume group chats and missed commitments
- •Applies the same triage/synthesis approach from work to personal life
- •Demonstrates practical agent behavior beyond coding and dashboards
- 33:07 – 37:00
Codex takes action in ~68 seconds: drafting a reply using Google Calendar + WhatsApp
Abhi demonstrates a full action loop: read a WhatsApp message, check availability across calendars, draft a professional response, and place it into the WhatsApp composer. The system intentionally stops short of sending, preserving user control for final approval.
- •Multi-step agent flow: ingest message → consult calendar → compose reply → open app and draft
- •Human-in-the-loop safeguard: drafts message but asks user to press send
- •Speed improvements make computer use feel viable for real workflows
- •Permissioning and action review mitigate risk of unintended actions
- •Illustrates how agents can reduce “coordination tax” in daily life
- 37:00 – 40:20
Building a 1040 tax-filing app + safety: data handling and control permissions
Abhi recounts building a web app that ingests tax documents and outputs a completed 1040, then cross-checking it against an accountant—catching a missed income source. He then discusses safe usage: data protections (training off/enterprise) and control via permission levels and action review.
- •Agent-built app can produce structured, submission-ready outputs (not just advice)
- •A/B comparison against accountant surfaced a real mistake
- •Safety has two dimensions: data privacy and action/control permissions
- •Use settings like training-off/deletion/enterprise protections as appropriate
- •Set explicit checkpoints in prompts (e.g., draft first, request approval)
- 40:20 – 47:18
International growth and the 900M WAU story: expanding beyond knowledge workers
Abhi explains his role: ensuring OpenAI’s products benefit users outside the US and beyond the Silicon Valley knowledge-worker bubble. He connects global growth to product shifts like Search and multimodal breakthroughs like image generation that better match how many users actually engage with technology.
- •Mission framing: “benefit all of humanity” implies a strong global focus
- •Knowledge-worker share varies dramatically (high in Germany/US, far lower in India/Brazil)
- •Growth requires serving non-knowledge-worker use cases (small businesses, WhatsApp-driven work)
- •Search removed knowledge-cutoff friction and broadened everyday utility
- •Image generation unlocked multimodal engagement and new adoption vectors
- 47:18 – 57:46
ImageGen 2 demo: multilingual text rendering, consistency, and real-world outputs
Abhi shows why ImageGen 2 represents a major leap: better control, editing, multi-image storytelling, and improved performance across languages and scripts. He demos realistic outputs (e.g., Indian-language book titles) and discusses pro workflows like using “thinking” mode and targeted edits.
- •Big improvements: precision edits, batching/storytelling, and multilingual rendering
- •Better character consistency across panels (e.g., manga example)
- •Demo: realistic bookstore/books with multiple Indian languages—previously a weak spot
- •Pro tip: switch to “thinking” for higher-fidelity, more realistic generations
- •Remaining limitation: steerability—changing one detail while preserving everything else
- 57:46 – 1:07:06
Being a PM at a frontier lab + breaking into OpenAI: live AI, evals, and builder mindset
Abhi contrasts frontier-lab PM work with traditional PM roles: the model is itself a product, capabilities are uncertain until release, and roadmaps must stay fluid. For aspiring candidates, he emphasizes using AI deeply, understanding evals as the ‘currency of progress,’ and demonstrating builder behavior—illustrated by his own path through international work and shipping prototypes.
- •Frontier PM reality: ambiguity, emergent behaviors, and dynamic roadmaps
- •“Model is the product” + product layers built on top
- •Evals as the core mechanism for measuring and driving model progress
- •PM depth in evals varies by role; fluency matters more than mastery for entry
- •Career story: international focus + building a translation prototype helped him break in
