Ryan Drapeau

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

Waymo is steadily capturing market share from ride-hailing giants Uber and Lyft. According to a recent Sensor Tower report, the rise of self-driving technology is no longer a distant future—it’s reshaping the mobility landscape right now. What does this mean for the traditional ride-hailing business model? As autonomous vehicles become more reliable and accessible, companies like Uber and Lyft face a fundamental challenge: adapt or risk losing relevance. The convenience and potential cost savings of self-driving fleets could quickly shift consumer preferences. Are we witnessing the beginning of the end for human-driven ride-hailing, or will established players find new ways to compete? The next few years will be critical for the entire industry.

1

Two companies are building self-driving cars, using two very different philosophies. 🚗 Waymo High quality inputs. Curated data. Maps, constraints, labeled edge cases. Teach the model the world before letting it drive in it. 🚙 Tesla Scale over curation. Massive real-world video. Let the model learn patterns by watching humans. Both approaches work. Both fail in different ways. And it maps strangely well to AI agents. Do we design with structure, domain rules, and controlled context? Or do we let systems learn behavior from scale and refine over time? Or does the future belong somewhere in between? Curious where people land. I know where I lean, but the debate is the interesting part.

17

11 Comments

I've been discussing this recently with Frances Buontempo in the context of coding assistants. We hear a lot from developers saying that LLM assistants in their fave IDEs are great for generating a lot of the (boring) boilerplate code needed to stand up a non-trivial system. However, most popular IDEs have been able to do exactly this for decades--whether with "wizards", code-block and project templates, or other various forms of meta-programming. The difference is that the "old" way is much less hungry of resources, while simultaneously being much more predictable and reliable at automatically generating the "boring" stuff. Coding assistants are also good at other things, of course, but many of the popular (in my conversations with developers) tasks are those I would much prefer to do myself, such as writing tests.

15

6 Comments

Uber is investing over $100M to build fast-charging hubs for autonomous vehicles in SF Bay Area, LA, and Dallas, prime spots for its upcoming public robotaxi launches. Uber said owning these stations boosts efficiency, cuts costs, and keeps vehicles on the road longer, maximizing utilization and uptime. What’s next for ride-hailing?

5

Last Tuesday, Deliveroo's rider-switch Rails endpoint spiked to 4s latency and intermittent 503s. The flame graphs told the real story and became our speed playbook. Technical context: Flame graphs visualize time across call stacks using sampling profilers. In Rails, latency often stems from DB calls, serialization, or middleware. This makes root causes visible to newcomers. Recent tech context includes AI agents and RAG patterns (LangGraph, Claude 3.5, GPT-4o, Gemini 2.0) guiding performance analyses. KEY INSIGHTS: 🔍 Flame graphs reveal hot stacks and where wall time sits. ⚡ Latency on the hot path fell from 4s to ~800ms after fixes. ───────────────────────── 🔗 Read the full article: https://lnkd.in/ghcREzpZ 🎯 Practice interview questions: https://lnkd.in/gmy5drNw #performancetesting #cpuprofiling #memoryprofiling #flamegraphs #latency

1

1 Comment