Aakash Gupta on AI PM system design

Aakash Gupta’s latest AI product manager advice says the interview has moved from classic product design to AI system design. Candidates are now being tested on how they structure an AI system, choose signals, and make technical trade-offs. (aakashg.com) (news.aakashg.com) In Gupta’s April 17 mock interview, the case is a churn-reduction agent. He frames the goal around revenue, asks for platform-specific signals like mobile versus desktop behavior, and pushes the candidate to define churn before proposing a build. (aakashg.com) That format is different from older PM prompts about redesigning a consumer app or inventing a novelty feature. Gupta says companies are testing product sense and system design together, especially for AI PM roles at OpenAI, Google, and Meta. (aakashg.com) (aakashgupta.medium.com) His broader playbook is simple: an AI PM has to show three things. Mahesh Yadav, in Gupta’s September 2025 interview transcript, lists them as building in AI, handling AI PM decisions around data, models, evaluations, and iteration, and operating under scale and ambiguity. (aakashg.com) The “builder PM” label is central to that shift. Gupta’s site says PMs are increasingly being asked to push pull requests and code, while Yadav’s April 2026 episode says he left a $1.3 million total compensation package after product roles at Microsoft, Amazon, Meta, and Google to build independently. (aakashg.com) (youtube.com) Gupta’s recent Claude Code essay turns that into a practice regimen. He writes that the PMs getting hired in 2026 are not the ones who “tried it once,” but the ones who have built up persistent context files, workflow skills, and connectors to tools like Slack, Drive, Notion, and Figma. (medium.com) He describes the early problem in plain terms: the model starts out generic because it has not been onboarded. His answer is to give the model memory in practice — project context, conventions, evaluation criteria, and recurring task files — so output quality improves over repeated sessions. (medium.com) (news.aakashg.com) That same logic shows up in his product-building advice. In a 2025 guide, Gupta says many teams jump to fine-tuning too early, while durable AI products depend on prompt structure, pipeline design, and reliable system behavior before heavier model customization. (news.aakashg.com) He has made evaluation a separate pillar too. In his March 2026 podcast post with Braintrust chief executive Ankur Goyal, Gupta says the companies running 12.8 evaluation experiments per day are the ones shipping AI products that hold up in use. (news.aakashg.com) Put together, Gupta’s message is less about memorizing an interview framework than building taste through reps. The candidate who can define the user problem, pick the right signals, set up evals, and prototype a working agent is the one his material is training for. (aakashg.com) (medium.com)