We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Big Ideas
Agent products need an architecture audit, not just a compelling demo. Aakash Gupta’s checklist spans eight layers— infrastructure, agent networks, protocols, tooling, cognition, memory, application, and governance—and names three failures: building everything instead of buying selectively, jumping straight to the app, or forgetting a layer. For PMs, his test is to identify the neglected layer and focus the team there; at the application layer, perfect one workflow rather than a feature monster. Apply it as a launch review: assign an owner and evidence to each layer, especially tool execution, memory, monitoring, and compliance. Surface-level usefulness can conceal a system that cannot reliably act or ship.
Consumer AI has an authenticity problem. Andrew Chen contrasts work—patterned drudgery that AI compresses—with consumer experiences, where people seek novelty and parasocial connection and reject “slop”; he says the same adversarial-creativity challenge appears in sales and marketing. Treat novelty, distinctiveness, and human judgment as product requirements where sameness destroys value; do not measure a consumer AI experience only by automation or output volume.
Tactical Playbook
Make discovery a three-way pressure test. A PM/design thread proposed lightweight designer-led research for A/B hypotheses and pre-build discovery. The useful operating model was to bring a designer and lead engineer in as early as possible, jointly validate business goals, and welcome pushback—while keeping scope and time constraints explicit. For each candidate problem, have the trio state the hypothesis, run a small customer check, define what evidence would change direction, and agree the time-to-market boundary before requirements. This preserves collaboration without recreating a waterfall or feature factory.
Case Studies & Lessons
OpenAI uses two design tempos—and treats failure as input. Ian Silber, OpenAI’s head of product design, says some ChatGPT features try 100 ideas, discard 99, and use research and experiments, while Codex and the “super app” build in public, take big swings, and learn quickly; some ideas go from idea to ship in four hours. At Instagram, he says IGTV was a huge flop because the team baked in wrong assumptions and constraints; changing those and iterating produced Reels. Use deep validation for durable, core surfaces and fast experiments for reversible ones; the important output of a launch is what to fix next.
A vertical AI service makes trust part of the product. A Canadian lawyer reports an AI-native firm using its own playbooks, AI first-pass drafting and redlining, versioned audit trails, commercial data terms, and lawyer review and signoff. It offers flat fees and 48-hour turnaround, and says its first month, not yet closed, is tracking to low-to-mid five figures. The firm sells legal work directly to startup and SMB clients rather than software to law firms. The founder’s warning is broader than legal: a polished agreement can still be wrong for the company or jurisdiction. In high-trust workflows, provenance, human verification, and domain checks are features—not post-launch cleanup.
Career Corner
Build the “shaping the build” muscle. Andrew Ng’s map, based on more than 10,000 job postings, expert interviews, and surveys, names four AI-engineering skills, including “shaping the build.” He says that means product sense, business context, customer goals, and judgment about when to ship an MVP versus build carefully. Shreyas Doshi highlighted that product-sense line. For PMs, make this visible in your work and portfolio: show the customer problem, the scope and tempo choice, and how you steered the build. Silber’s hiring rubric points in the same direction—curiosity, prototyping, a point of view, strategic thinking, and systems thinking.