We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Big Ideas
AI-speed delivery is exposing what PMs actually own. One PM reports that a Palantir-inspired Forward Deployed Engineer went from doing little to building and deploying without consulting the PM or development team; the department now routes everything through him. The FDE reportedly learns the department’s needs, proposes high-ROI solutions, builds and tests with Claude, runs UAT, and deploys—effectively combining PM, engineering, and QA. The durable PM response is to own value propositions, customer needs, and priority—not developer coordination—then turn validated bespoke work into a scalable or self-service roadmap. The governance test is whether a feature serves the install base or only one customer; otherwise a software team can quietly become a consulting operation. The counter-risk is technical debt: commenters expect the shortcut to look good for a year and become unmaintainable when the engineer leaves.
Specificity is becoming an AI-product advantage. Hiten Shah describes working almost entirely asynchronously in Slack with agents alongside him; narrowly targeted marketing tools make an expert’s “private operating system” executable, and their specificity is the advantage. His team’s second product is nearing shipment after more than a year, several beta-driven pivots, and iteration alongside local-model progress; it will enter a crowded market with a permanent free version. For PMs, the implication is to encode unusually deep workflow knowledge rather than chase generic AI feature parity.
Tactical Playbook
Run EOL as a go-to-market program, not an engineering ticket. Start with Legal to surface contractual, regulatory, and data obligations; then Finance, Sales, Marketing, Customer Success, and Support, before discussing the technical shutdown. Speak with anchor and difficult customers before a mass announcement to uncover hidden integrations and dependencies. Publish the full sequence—end of sales, upgrades/development, support, and life—and explain migration, carryover, cost, or data export. Assume communications are ignored until acknowledged; one practitioner estimates 90% of customers will not migrate until forced. For online products, trigger notices when users log in, offer parallel or reversible migration where possible, define recovery steps, and put a public deadline months ahead of the technical cutoff.
Normalize launch metrics. When a launch creates a spike, ask for the pre-launch baseline: Paul Graham cites a startup growing 40% per week before its spike, an impossible compounding rate but a useful signal that users liked the product.
Case Studies & Lessons
Robotics founders are selling the job before training autonomy. YC’s Robotics Club discussion says 2026 is still “the year of the demos”: work-cell successes exist, but general-purpose robots remain unavailable and teleoperation data collection is remarkably hard. Rerun’s Niko recommends starting with one paying customer problem, off-the-shelf hardware, and teleop; deploy quickly because the physical world exposes failure modes a lab cannot anticipate. A paper-plane example produced concrete learning: 1,000 perfect planes per day for viability, a tray that cut failures by 50%, 20 hours of operator practice, and pick-and-pack worth 10 times more than folding alone. Build a replica of the customer environment, define a business-specific evaluation manually before automating it, track failure metadata, and train continuously rather than collecting all data upfront. The PM lesson is to treat autonomy as a scaling factor, not a launch prerequisite. The infrastructure side is equally product-led: General Instinct frames inference economics as the constraint, describing roughly $70,000 GPUs and a path toward 500 ms per 16-action chunk on Jetson Orin.
Career Corner
Protect yourself against offer rescission. Aakash Gupta describes a written offer worth $25,000 more annually being pulled after the candidate resigned. His checklist: ask whether headcount is fully approved, wait for the signed offer and cleared background check before giving notice, and keep interviewing. If rescinded, try to restore the current job; otherwise get the withdrawal in writing and ask for one to two months of lost pay.
Tools & Resources
Test the conversational control-plane hypothesis. Saturnial argues that dashboards are no longer the control plane and that “the harness is the interface”; Hiten Shah echoes that “chat killed the dashboard.” For agentic workflows, prototype a surface where users state intent, inspect what happened, and correct it—then measure whether it beats passive reporting rather than assuming chat is always better.