Junya Li

Waymo’s recent robotaxi incidents surface one of the most important and least discussed questions in AI product design: What should a system do when it doesn’t know what to do? Waymo’s answer is refreshingly honest, i.e. stop. Its vehicles are explicitly trained to halt when uncertainty rises (power outages, broken traffic lights, unexpected human behavior). This isn’t a model failure. It’s a product decision. When the safest action is inaction, the system chooses restraint over guesswork. At the level of one car, this is rational. At the level of ten cars, still sensible. At the level of hundreds, it becomes a system-wide failure. Waymo optimizes for local correctness, i.e. when unsure, do no harm. But without orchestration above the vehicle, local safety becomes global brittleness. Correct behavior in isolation can cascade into city-scale paralysis. The real lesson: building AI products isn’t about maximizing intelligence. It’s about designing systems that behave predictably when intelligence runs out and understanding how those behaviors compound at scale. Contrast this with China’s Apollo Go (Baidu) and Tesla and you will see three different strategies with different trade-offs. Baidu reduces ambiguity upstream with an “infrastructure-first” strategy. They’re not just building autonomous cars, they’re building V2X cities where roads, signals, and networks talk to vehicles. When things break, the city steps in. Fleets are centrally directed into predefined “safe harbor” states. The result is system-level resilience albeit at the cost of heavy coordination, regulation, and long-term capital alignment. Tesla takes the opposite bet. It assumes ambiguity will reach the vehicle and designs for it. No smart roads. Vision-only autonomy with end-to-end neural networks learning from millions of edge cases. Tesla isn’t optimizing for certainty in the moment, but for learning at scale over time. The product takeaway is clear: robust AI systems can’t rely on a single philosophy. The most resilient products layer their defenses: + Local safety behaviors that fail predictably + System-level coordination that prevents cascading failure + Continuous learning loops that improve with scale In AI, uncertainty isn’t the edge case, it’s the design constraint.

15

3 Comments

I took my second Waymo across Los Angeles last night, and the experience was largely excellent. From the smooth pickup to navigating the freeway (which is now supported here in LA!), the technology is undeniably impressive. However, the "last mile" wasn't quite perfect. I was dropped off on the left side of a two lane, one way street, directly in traffic rather than on the shoulder. Also, my destination was on the right. As I've written previously (https://lnkd.in/gHw7drrc), there is great promise in using technologies like AI in countless ways, from autonomous driving to talent measurement. However, my recent Waymo experience highlights two critical lessons: 1. Humans Must Own the Outcome: Whether you are a passenger in an autonomous vehicle or managing talent, you cannot delegate accountability to the machine. Technology is a powerful tool, but humans must own the final result. In hiring, as in transport, we must remain in the driver's seat to ensure the "drop off" (i.e. the final decision) is the right one. 2. The Feedback Loop is Non Negotiable: Without human feedback to these systems, technology will never improve. I made sure to provide feedback on my ride, and we must do the same in our talent systems. If we don’t provide constant, iterative feedback, we lose the opportunity to refine them into something better. Progress happens when human intuition meets technological scale. We provide the feedback; the tech provides the lift.

33

4 Comments

I was CPO at Blend before their $4 billion IPO. We rejected the “forward deployed engineer” Palantir model that everyone is copying now. Here’s why: Blend could have followed the forward-deployed engineer approach. Both our CEO and CTO were early Palantir. They'd seen it work. Instead, we built a SaaS product with core components. Move fast, iterate, use data to learn. Not custom code for every client. That Palantir model is having a resurgence. Now Sierra, Decagon, and others are hiring forward-deployed engineers. But it’s not the right approach for every company. It’s fundamentally the wrong answer for Assembled’s customers. Here's what we understand about support teams and why we’re not going the forward deployed route: Customer Support teams are used to getting no resources and no support. On the surface, you'd think they'd be excited about engineers building custom solutions. "Finally, someone's offering us developers!" The reality is support teams are terrified of getting help they can't afford to keep. Support teams are wary because they know the math. They don't have budget for expensive, scarce resources. At some point the shoe drops. The engineers disappear when priorities shift. And custom solutions break. Support is left holding the bag with something they can't maintain. They're resource-starved by design. The idea of a dependency on resources they can't control is terrifying. So instead, the goal should be to give support teams the ability to manage, change, and solve their own problems. Not dependencies on engineers they can't keep. Smart, competent people who teach them to do it themselves. Palantir wasn't even a product company for years. They were a custom development shop with reusable components. Show up, build something bespoke, maybe extract learnings. That's not scalable for any but the largest support teams who need sustainability. Forward-deployed engineers are the new shiny object. I wasn’t a fan of it at Blend. And I’m not a fan today either. What I see in the customer support space is that teams need empowerment, not dependency. I want the teams we work with to be able to say: "I love that you’re partnering with me. And I love that I can do this myself too." That's the difference between building a product and building a consulting practice.

80

6 Comments

Could non-technical staff build 9x AI workflows in one day? A week or so ago, I ran a one-day hackathon inviting engineers and colleagues from onboarding, RevOps, support and operations. The winning team's fraud detector identified a bizarre pattern; a customer buying a chicken carcass every few minutes for two days straight... 🐓 Over 20 participants formed 9 cross-functional teams and, by day’s end, each pitched an AI workflow: - CS and knowledge-base chatbot, - underperforming merchant alerts, - renewal-email sentiment analysis, - fraud detection 🥇(winning team!), - sales-opportunity optimiser, - loyalty-socials content generator, - optimised outbound workflow, - CI deprecation solver, - lead qualification and more... 𝗛𝗲𝗿𝗲'𝘀 𝗺𝘆 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝗮𝗱𝘃𝗶𝗰𝗲.. Within the next week, invite one non-technical colleague to co-create a simple Agentic workflow. Dedicate four hours to tackle a routine process, such as handling support tickets or generating renewal reminders, using a no-code interface that connects to your data (via an MCP server) such as N8N or Zapier. Focus on a clear pain point, prototype rapidly and iterate based on immediate feedback. What cross-functional AI workflow will your team prototype next? Let me know in the comments if you get stuck! #AIWorkflows #CrossFunctionalAI #Hackathon #AgenticAI #DataInnovation #DiverseTeams

23

1 Comment

Kudos to Moein Hasani, Trace Levinson, and the team for sharing their work on evolving recommendations at Instacart. They’re building an AI-native, LLM-powered discovery stack that generates more cohesive, personalized pages — grounded with RAG and supported by strong evaluation and relevance models. Great to see thoughtful innovation in how we approach discovery.

22

I wrote about how product managers should deliberately decide which features should be built to an A grade and which ones should be built to a B or even only a C grade. Making the resource tradeoffs to optimize business outcomes is a core job of product management, and building every capability to an A grade is a poor use of scarce resources. https://lnkd.in/gTXBfNYE

31

4 Comments

Excited to share our team's work on developing a Ranking Engineer Agent (REA), an autonomous AI agent, to fundamentally transform the end-to-end ML lifecycle for ads ranking, one of the most sophisticated and complex machine learning systems. Improving ads ranking has been a deeply complex, manual, and multi-week-long process, heavily reliant on manual effort across hypothesis formation, configuration, execution, and failure recovery. REA is changing that paradigm. It now autonomously manages the entire workflow from generating high-impact hypotheses and designing experiments to executing training jobs, diagnosing system failures, and optimizing iteratively. This critical shift allows our engineers to pivot from operating the experiment loop to directing it: setting the strategic vision, evaluating macro-level results, and making crucial architectural decisions that propel our business forward. Congratulations to the entire team for pushing the boundaries of what autonomous ML engineering means at Meta scale. Please see the full technical deep dive on the Meta Engineering Blog here. https://lnkd.in/gtjGyrBv

265

1 Comment

My Startup Was Working… Until It Wasn’t We had paying customers. We had local press. We were gaining traction. It was 2016. Then UberEats, DoorDash, and Postmates entered Portland with billions in funding. At the same time, Intel—where I worked full-time—announced massive layoffs. Overnight, my 40-hour week almost doubled. Midnight fab shifts, weekend work, endless overtime. My startup didn’t die because the idea was bad. It died because I didn’t have the financial runway to compete. Here’s what that painful chapter taught me: 1. Passion + grit aren’t enough. Startups need time, capital, and breathing room to survive. 2. Corporate is just built for stability. It can’t nurture the creative chaos that startups need. 3. Adaptation beats stubbornness. Sometimes the smartest move is to regroup, save, and come back stronger. So I made a choice: doubled down on my W-2 career, moved to Seattle, and joined Amazon to maximize income. Every dollar earned—and every loan I could pull—went straight into real estate. One property a year turned into two… then into double digits in a single year. Deals kept coming, I kept closing, and I built a local team to keep momentum alive. By the end of 2022, my portfolio was generating more passive income than my paycheck. That’s when WoolooAI.com was born—not as a flashy startup, but as a necessary system to simplify and automate the chaos: acquisitions, CRM, portfolio management, project management, all streamlined. The truth is: - Without the grind of my early startup days, I wouldn’t have landed the Amazon role. - Without Intel’s layoff threats, I wouldn’t have learned the importance of building a financial runway first. - And without the expertise I developed at Amazon, I wouldn’t have gained the skills that helped me build WOOLOO AI. The irony isn’t lost: the same challenges that once crushed me ended up shaping the freedom I enjoy today. 💡 Takeaway: Sometimes what feels like an ending is really just the foundation for the next chapter. 👉 If you want the full story, I wrote it here: https://lnkd.in/geEVye36 💬 I’d love to hear your story too—what twists or setbacks ended up shaping where you are today? #Entrepreneurship #Startups #RealEstateInvesting #AmazonAlumni #FinancialFreedom #LessonsLearned #HumbleGrowth

226

8 Comments