ZeroNoise Logo zeronoise
Post
AI-Speed Delivery Raises the Bar for PM Judgment
1 day ago
4 min read
148 docs
The sharpest signal is AI-enabled delivery bypassing traditional product boundaries. This brief translates that shift into PM operating guidance, alongside a teleop-first robotics playbook, EOL execution tactics, and career risk protection.

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.

AI-Speed Delivery Raises the Bar for PM Judgment
Research extraction
Y Combinator

Direct answer. As of mid-2026 the field is still "the year of the demos": Neo 1X can be pre-ordered but a Figure robot still cannot be bought, and there is work-cell success but no Rosie the Robot . The two startup plays pitched are (1) teleop-first "robotics application companies" that own end-to-end business problems , and (2) infrastructure to make world-action models fast enough and cheap enough to run on robots . The deepest bottlenecks cited are teleop data collection, memory/adaptation, reasoning data, sim-to-real dexterity, embodiment drift, and inference cost/latency .

What founders are building

  • Rerun (Niko, co-founder/CEO, ~4-year-old company) is a "unified data layer for physical AI": open-source SDK for logging, visualizing, querying, and loading physical data, plus Rerun Hub as a cloud data catalog/backend . He argues physical data (multimodal, multi-rate, episodic, 3D-nested) is very different from web data and breaks normal table-based databases, so storage-layer choice drives a lot of operational friction .
  • General Instinct (YC company) is building "infrastructure for you to run physical AI models really fast" — analogous to vLLM/SGLang for LLMs . Their optimizations: distill the VAE encoder-decoder, shrink/split the diffusion transformer, use cross-attention from the video transformer to the action transformer instead of decoding future frames, cut flow-matching sampling from ~50–100 steps to ~1–3 (≈50x speedup), and use KV-cache energy for asymmetric denoising . Target: run a 16-action chunk in ~500 ms on a Jetson Orin . Motivation: a world-action model otherwise needs two "GB 200" GPUs at ~$70k each, which the speaker calls economically unscalable for robotics . They chose infra over an application company because "eventually every model ... needs to go on the edge and needs to go in real time" .
  • Robotics application companies / "neo-integrators": take ownership of a full business problem end-to-end, use as little custom hardware as possible, and often start with pure teleop plus fine-tuned models rather than foundation-model generality . Early traction areas: data centers, warehouse robotics, small-scale manufacturing, food; reliable-robot companies are described as "100% supply constrained" — demand is easy to fill, while labor is scarce and hard to manage .

Startup playbook and PM lessons (Niko)

  • Start with a single customer problem someone will pay for; solve it teleop-first with off-the-shelf hardware, sell/deploy fast, because "the physical world is brutal" and lab work cannot anticipate all failure modes .
  • If it works with teleop, generally a model can later learn it . Teleop-first also means many businesses (even robo-taxi) don't need full autonomy at launch; autonomy becomes a scaling factor, and data-scale needs are learned by plotting your own scaling curves rather than guessed .
  • Fictional paper-plane factory illustrates the learning pattern: ~1,000 perfect planes/day viability; customers care most about speed of onboarding new designs; a simple paper tray cuts picking failures ~50%; a teleoperator needs ~20h of practice to meet demand (changes hiring/ops); adding pick-and-pack is ~10x more valuable; cheap research arms break and force supplier changes .
  • Set up "basics for learning": fine-tune an open model (e.g., Pi) on a few hours of teleop demonstration; keep a replica of the customer environment in your office; build a repeatable, business-specific evaluation (sharp edges, symmetry, weight distribution); do eval manually first, then automate; track metadata and failure classifications — "good data is data that makes the model better" .
  • Iterate constantly: do not collect all data upfront then train; even big-lab debugging often ends in "send a different instruction to your data collectors" .
  • Moats come from non-model operations: sales, assembly, shipping speed, unboxing, support — things pure model companies won't do . Application startups with relatively little capital are already "making a lot of money" and growing fast .
  • Why not billions of revenue yet: robotics is "death by a thousand cuts"; cheap ways to try applications are new, and only recently have arms and base models gotten cheap/good enough; lack of those basics forced general raises .

Technical constraints and metrics for PMs

  • Teleop and data reality: collecting teleop demonstration data is "remarkably hard" and finicky; if the approach depends on scaling this data on gripper/wrist-camera setups, "we're kind of doomed" .
  • Four unsolved walls: world models don't respect real physics (sim-to-real gap, deformable objects, action conditioning), action-space representation is unsolved, robots lack human-like sensing (one F/T sensor per fingertip vs. human nerve endings detecting force, moisture, temperature, vibration, friction), and embodiment drift (dust/corrosion/battery aging) makes teleop data stale and forces recollect/retrain .
  • Long-horizon memory: state-of-the-art policies like PI0.5/Groot have no memory; useful tasks (full-bed cleaning, full meal) run tens of minutes while internship demos top out around 2 minutes; adding memory fixes endless dishwashing/burnt-grilled-cheese loops and enables in-context mistake correction, but naive memory hurts performance and training/inference efficiency .
  • Reasoning: robotics is data-scarce, so annotated chain-of-thought is a data augmenter, but selective reasoning beats exhaustive reasoning; in manipulation, move/gripper-position reasoning helps while perceptual reasoning gets pruned; in self-driving, pruning hallucinated annotations lowers L2 path error/collision rates; "action forcing" lets you drop reasoning at inference to dodge latency .
  • Sim-to-real dexterity: a single goal-conditioned RL policy trained on primitive objects/random goals runs at 60 Hz controlling a 22-DoF hand + 7-DoF arm zero-shot, generalizes to 12 unseen tools, and turns a new task into a new goal-pose sequence ("task prompt") — a new object/task in minutes instead of hours/days . Roughly 60% of failures come from pose tracking (not the policy) ; recovery behaviors emerged because random forces in simulation forced re-grasping .
  • Evaluation is a product problem: real-world eval is expensive and rarely statistically meaningful — "you can almost get a vibe check unless you run it truly 100 trials" (can't tell 90% vs 87%); teams build automated sim evals and still need human insight .
  • WAM inference economics: world action models learn future kinematics from future video, but inference is heavy — a 16-frame chunk needs ~32 diffusion-transformer runs due to CFG, plus 50–100 autoregressive steps; distillation and latent-state approaches with no frame decoder are the fixes .

Tensions and gaps to flag

  • Tension: Niko's thesis is "if you can solve something with teleop, generally you can train a model to do it" , while Tyler's sim-to-real work exists because "highly dextrous actions remain difficult to demonstrate reliably and at scale" via teleop . The transcript does not resolve where the boundary lies.
  • Numeric inconsistency: flow-matching distillation is described as 1–2 steps in one place and 2–3 steps in another , both claimed without performance drop.
  • Gaps: no revenue/ARPU data for application companies is provided beyond qualitative "making a lot of money" ; hardware costs and deployment economics for teleop-first models are not quantified; most metrics are demo/lab-scale rather than production deployments.
Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club
Product Management

Reddit PMs report a growing threat to PM roles from Forward Deployed Engineers (FDEs) using AI coding. An OP describes their CTO, inspired by Palantir, rolling out FDEs as a test; an FDE then built and deployed features without consulting the PM or dev team, and the department now routes everything through the FDE, while leadership reassures 'FDEs will not replace the PM role' . The same PM describes the FDE as a one-person PM/dev/QA using Claude to build, test, and deploy at very high speed, causing code conflicts with the core dev team .

Critics warn FDEs will be 'great for a year until the tech debt is due' , leaving unmaintainable solutions when the engineer leaves , and skipping evals, latency/performance testing, and security checks . Advice for PMs: focus on value propositions and customer needs, not on managing developers ; embrace the FDE, bring them into the core team, and align their output with business priorities ; and productize FDE work into a scalable platform roadmap . Some concede that if an FDE closes the idea-to-build loop faster, PMs could be skipped for minor deviations on mature products . A blunt take: if one FDE with Claude delivered more value in four months, the PM wasn't shipping what users needed .

Veterans see the model as a rebrand of Professional Services/solutions engineering: 'the technology changed. The organizational physics didn't' , though FDEs differ by being embedded in engineering and able to change the product and deploy to prod . The thread also debates internal vs external PM roles: internal PMs are at risk because 'there is no business to run' , while DevEx PMs argue internal products need managing even more due to organizational noise . Custom one-off builds lower gross margins from software's 70-80% to consulting's 20-30% . One PM reports being laid off possibly due to an FDE hired a month earlier .

My PM role seems to be quite... threatened. My company has been exploring Forward Deployed Engineers (FDE's). The FDE assigned to the department I interact with most is completely bypassing me and even the dev team... From what I can tell, FDE sits with the department to understand their processes/needs, proposes solutions for the highest impact/ROI, bu… FDE’s are going to be great for a year until the tech debt is due. He won’t burn out but his solutions will turn in to an unmaintainable mess, probably right around when he leaves to a new company 100% agree. OP, is the FDE also doing the evals, harness, testing for latency and performance, and other scale-out issues? Start intervie… Product Managers continue to be the people who should be focused on value propositions, customer needs and then focusing the lense of pri… If I was you I would embrace the FDE and make him part of your core product team. Get him involved in all the things you are trying to do… Just treat a subset of the FDE work as your platform roadmap, and make it scalable or real or self service. The PMF is done for you. Unsh… Worst case scenario is that the FDE is able to close the feedback loop speed from problem/idea to getting something built. Especially if … This is a problem, but not the way you’re thinking. If a single FDE with Claude can take user confidence and deliver this much value to t… Excuse me, perhaps I'm reading this wrong, but I'm so old I remember when we just called these people Professional Services. Back in the … PS still exists, FDEs are usually inside the engineering teams. Depends on the org, but PS implements, executing an SOW. An FDE can actua… Sounds like this is an internal product? If so, your job as a PM is at risk. Internal projects don’t really need PM. There is no business… Developer platform / DevEx PM here: You're out of your damn mind if you think internal products don't need managing. If anything they ben… The question I would ask is "what business are we in here?" If your business is to produce one off custom solutions, then all is good. If… I literally got laid off this week I think a large factor was the new FDE they brought in a month or two ago
Product Management

Treat EOL/deprecation as a real go-to-market effort, not an engineering task: over-communicate via multiple channels (email, docs, login screen, in-app banner), set expectations on what will and won't work, and assume customers haven't read anything until they explicitly acknowledge it .

  • Customers won't migrate voluntarily: notify paying/licensed customers in advance and honor existing contracts; expect ~90% to take no action until forced, even when migration is trivial; if you give an out, nearly everyone takes it, and customers will drag their feet and ask for extensions — churn risk is real .
  • Timelines will slip: planned EOL dates rarely stick (one PM's Y2K fallback software was promised gone by 2003 but ran until ~2014); build buffer into the schedule and set a public deadline several months before the technical cutoff, expecting stragglers to complain for months after .
  • Migration mechanics: for online products, trigger direct notifications right after active users first connect and repeat at intervals; if old and new products share a data source, let customers toggle between them during migration (Marty Cagan's 'Gentle Deployment'); if data sources differ, build a data-transfer interface so customers can run both in parallel on their own schedule; also define recovery steps ahead of time since undocumented dependencies on legacy systems are common .
  • Sequence stakeholders before announcing: Legal first (contractual promises, regulatory/data obligations that survive EOL), then Finance (revenue already baked into projections, budgets, renewals; cost of keeping the product alive), Sales (bundles, renewals, multi-year contracts, pipeline deals, undocumented field promises), Marketing (campaigns and spend already planned), Customer Success (they absorb the backlash) and Support (what stays supported, escalation) — Engineering is the last conversation . Have one-on-one talks with key/anchor customers before the mass email to surface hidden workflows and dependencies; present the full chain of dates (end of sales, end of upgrades/development, end of support, end of life) and keep communicating as each approaches .
  • Case study (B2B enterprise EOL, 2024): process took ~3 years, driven by a hard end date from third-party data contracts; communication ran through email campaigns, documentation, login screen and in-app banner; tracked customer database, feature usage, integrations and partners; Legal helped deprecate unused features early; after end-of-sales only security/data-export features shipped. The team shrank from ~60 devs to 3 devs plus a DB team, causing institutional-knowledge loss and stability strain; ~4,000 customers were moved off. The PM's biggest regret was not sharing more context with the dev team, who read the nontraditional PM work as laziness. No replacement product was offered .
Over communicate. Then over communicate some more. No one reads anything. Ensure your time lines make sense if there’s work involved on t… Treat deprecation like a real GTM effort - you need to set user expectations on what will and won’t work. Last thing you need is some con… - Assume people have not and will not read any comms you send them, regardless of channel. Do not assume they know or understand anything… I end of lifed a B2B enterprise product in 2024. Communication really is the key to success. We had a hard end date, due to some third pa… You need to notify your customers in advance of the deprecation if they are paying for licenses or support. You probably want to honor yo… Your planned EOL day will not be. Neither will the second one. Or the third. The super duper pretty please with sugar on top fourth date … What kind of a product is it? Is it online so you can engage people directly, or something where you do not really have visibility over w… The key to successful EOL is figuring out how to lose the legacy product without losing the legacy customer. Here's what I learned the ha… Find your most difficult customers. Walk them through the entire EOL and transition and get them to weigh in.
Hiten Shah
  • Product thesis & first product: Typeahead, a Mac AI autocomplete, is the first product from Hiten Shah and Sam Asante's thesis that the next wave of AI software will move closer to the user; local AI is faster, more private, and makes offline work possible by default .
  • Use-case selection: They started with writing because it happens everywhere on a Mac in small moments where a blank AI workspace or long prompt is too heavy; Typeahead's product decision was to offer inline suggestions inside the apps where writing already happens .
  • Lessons from v1 → 2.0: Inline help must feel fast, native, and controllable; voice matters most — "generating more text is easy. Helping someone finish the sentence they were already writing is the harder and more useful job" . The filter for 2.0 became "Does this help the writer stay in flow?" and the strongest product decisions were those that kept the writer in flow .
  • Local as a product decision: Local wasn't added as a privacy feature; it was the product decision first, because writing contains sensitive context and if the Mac can run the model, writing should stay there by default — enabling faster suggestions, offline work, and a real boundary for the user . Hybrid systems will still matter; the product question is where the default should live .
  • Pricing: Typeahead is a one-time purchase to feel like software you own; subscriptions make sense for hosted services, collaboration, storage, and infrastructure, but a local utility has a different shape .
  • User feedback & roadmap: Feedback centers on speed, suggestion placement, context understanding, and native feel; users also value local-only processing, offline use, and one-time purchase. The roadmap is to get faster, understand context better, and adapt to more writing places while preserving the trust boundary .
  • Second product: Their second product is almost ready after a year+ of work; it enters a crowded space, beta feedback drove several pivots, and it will be free to download with a free version expected to remain . The product is being iterated in tandem with local AI model development — the idea has been evolving alongside the technology .
Why Sam and I started a company around local AI software Our second product is almost ready. The space we’re entering is crowded. That makes me even more excited about the bet we’re making on lo… It almost feels silly to say we've been working on a product for a year when AI makes building products so much faster. In the case of th…
The community for ventures designed to scale rapidly | Read our rules before posting ❤️

Freelance or consulting experience builds product management skills — collecting product requirements and building iteratively — which can make former freelancers fit more easily into organizational structures than founders of established product companies, whose opinions may conflict with org priorities . When transitioning back to employment, a reputation for delivering (clients who can vouch for you) carries more weight than a title; such credibility comes from finishing things, not from having had a role .

It depends on the nature of the company and the business. If you are more like a freelancer you can fit nicely into a lot of org structur… The people on those calls are buying, which isn't the same as being willing to hire you. What actually travels is a client who'll pick up…
Product Management

A 24-year-old with a Product Design & Technology degree and about a year running his own startup joined a software company as an SDR and, within two months, exceeded KPIs and was given work beyond the role: building prospecting tools, helping with marketing campaigns, contributing to GTM strategy, and helping plan launch strategy for an acquired product . The company countered a competing offer with £40k base + £15k OTE + a larger-than-normal equity grant, and promised fast progression into a strategic closing role with someone reporting to him . The alternative offer is a Product Manager role in motorsport at £40k base with standard benefits — aligned with his degree and long-standing engineering/product/motorsport interests — but he is unsure whether the day-to-day is hands-on product work or admin/coordination-heavy, and less certain about progression . He frames his decision as fast-tracked commercial/sales leadership vs. a PM title in a passion industry, and is seeking input, especially from those who moved between sales/commercial and product/engineering careers .

24M, stay in a job where I'm progressing quickly, or take a Product Manager role in motorsport?
The community for ventures designed to scale rapidly | Read our rules before posting ❤️

A founder whose prototype is in regulatory review and who listed themselves as a senior employee on their resume is interviewing at a big company in the same space, hoping to gain income and learn how a top player builds and scales products . Advice from the thread:

  • If the company is a direct competitor, the employment agreement may contain a moonlighting clause and other restrictive covenants; have counsel review before signing .
  • Some companies publicly advertise a culture that encourages employees' side startups — mention the founder role there; otherwise keep it vague and frame yourself as a “lead” for the relevant portion .
  • Keeping it vague is the smart call; if asked directly, don't lie, but you're under no obligation to volunteer founder status unless the title genuinely matters to the role .
  • If the startup is in the same or a vaguely competitive space, the IP assignment agreement (EPIA) will result in either the company owning it or disclosure leading to a rescinded offer .
Mentioning I own the startup on my resume or framing it as an employee? I will not promote If it’s a direct competitor I’d be very surprised if there isn’t some sort of moonlighting clause in the employment agreement. There may … Some companies encourage employees to have external activities, such as their own startups. They’ll most certainly advertise that as thei… yeah keep it vague is the smart call. if they ask directly dont lie, but you're under zero obligation to volunteer that you're a founder … If it’s in the same space and vaguely competitive your EPIA is going to result in either them owning it or you disclosing it and them res…
Product Management
  • In product domains where industry knowledge matters (e.g., finance, healthcare), PMs debating product network vs. domain network leaned toward the domain: knowing people in target companies aids job referrals better than a generic PM network, and vertical connections matter more than fellow PMs ; several said build both if possible .
  • Industry connections double as low-effort market research: access to new information through established relationships can lead to new product features and a competitive edge .
  • When competitors and vendors are based in other countries, maintaining genuine industry connections is hard: LinkedIn feels superficial, and PMs seek other platforms for real relationships .
Product network or domain network? Industry. If I want to switch jobs, knowing people who work at other companies in the same industry and can refer me is more important th… Vertical is more important than fellow product folk. I mean if a person can successfully pull off both, then why not go for both? Por qué no los dos? Additionally, if you want your product to have a competitive edge, having more connections could provide you with access to new informati… Can you leave some suggestions on how one can maintain connections within their industry? I work in a company where most competitors and …
Hiten Shah

@hnshah (Hiten Shah) posted that chat has replaced dashboards as the product control plane , endorsing @saturnial's view that 'dashboards are dead as a control plane' — users should 'just talk to the computer about what you want,' and 'the harness is the interface' .

Chat killed the dashboard. [https://x.com/saturnial/status/2086131203636384025](https://x.com/saturnial/status/2086131203636384025) Dashboards are dead as a control plane. Just talk to the computer about what you want. The harness is the interface.
Paul Graham

Paul Graham observes that much of his office-hours work with founders involves giving them "permission to do something simpler than they were planning to," often saying "You don't need to x. Just y." This exemplifies a useful product-management practice: explicitly authorizing simpler, scaled-down approaches to reduce scope and waste.

A surprising amount of what I do in office hours is give founders permission to do something simpler than they were planning to. I'm alwa…
Hiten Shah

Hiten Shah: "You can vibe code the product. The user should never feel like you did." — building with vibe coding is acceptable, but the shipped product must not let users perceive that shortcut .

You can vibe code the product. The user should never feel like you did.
Paul Graham

Paul Graham asked a startup how fast they were growing before a launch-induced spike; they said 40% per week. He notes that 1.4^52 = 39.7 million, so no one can sustain that rate, but it shows users like the product.

I asked a startup today how fast they were growing. They said the current number was meaningless because it was just a spike from launchi… 1.4^52 = 39.7 million. So obviously no one can keep this up for long. But it shows users like the product.
Aakash Gupta

Job offers can be rescinded even after you've given notice; one job seeker lost a $25K raise this way, and in coaching 160+ people over 12 months this scenario came up more than admitted . Prevention tactics: ask the recruiter if headcount is fully approved before signing, wait until the offer is signed AND background checks cleared before giving notice, and keep interviewing to have other irons in the fire . If your offer is pulled and you've already resigned, try to keep your current job by having an honest conversation with your manager ("I'm committed to this team"), and if that fails, get the rescission in writing and ask for 1-2 months pay for what you're giving up — big companies often agree .

Employers will pull your offer, even after you’ve given notice to your current employer. This Redditor got a written job offer for $25K m…