Open Letter to Elon Musk
Subject: Urgent Warning on Your Elevator-FSD Analogy – A Fatal Design Blind Spot Threatens Optimus and Tesla's Entire Safety-First Mission
Sir,
I have followed your work long enough to take it seriously. The drive to make electric vehicles safer than human-driven ones, and now to birth general-purpose humanoid robots through Optimus—you have a real chance to do something life-changing for so many. But after your recent X post comparing modern elevators to Tesla's Full Self-Driving (FSD) system, I must warn you—in the strongest possible terms—that your frame of reference contains a dangerous distortion. One that could cripple Optimus before it ever reaches its potential. Safety first is not a slogan. It must be the non-negotiable foundation of every design decision. Right now, it is not.
You are correct that elevators are extraordinarily safe. Their fatality rate is microscopic—roughly 0.00000015% per trip across billions of rides—because they operate in a sealed, predictable shaft with redundant mechanical fail-safes. But your analogy collapses the moment we examine what actually defines "quality" and "safety" in vertical mobility systems. Death is a blunt, lagging, and frankly inadequate diagnostic. It only appears after the damage is done. Real engineering excellence is judged by leading indicators that prevent harm long before anyone dies.
Look at stairs—the simplest vertical system we all use daily. Their objective safety standard is not body counts. It is uniform spacing: riser heights capped at 7–7.75 inches, tread depths at 10–11 inches, with no more than ⅜-inch variation between steps in a single flight. These are hard, measurable biomechanical rules derived from human gait. Deviate even slightly and misstep rates explode. Injuries—over a million ER visits annually in the U.S.—become inevitable. We enforce these standards upfront in building codes because we understand that human factors plus poor physical design equals predictable harm. Deaths are rare; the spacing diagnostic catches the problem first.
Now apply the same lens to elevators themselves. Not all are equal. Otis and Schindler both clear the same minimum safety codes, yet facility managers, owners, and riders know there is a palpable quality gap. Otis often excels in predictive maintenance and energy-efficient flat-belt systems; Schindler frequently leads in ride smoothness, door precision, and low-vibration profiles. One elevator might deliver a silk-smooth, perfectly leveled stop with whisper-quiet doors. Another—same code compliance—feels jerky, mislevels by fractions of an inch, or slams doors. These differences manifest as daily annoyances, minor trip hazards, higher maintenance calls, and lost productivity. No one dies in either, yet one is objectively superior. Quality beyond death is what separates good from great.
Tesla cars already follow this manufacturing logic quite well. Gigafactory precision, tight tolerances, and data-driven iteration produce vehicles that outperform the industry on ride quality and reliability. But when you pivot to FSD and extrapolate to roads, your frame of reference warps. Roads are not shafts. They are adversarial, multi-agent chaos where vehicles can and do kill each other. Your neural nets are forged in that crucible—predicting, buffering, dodging. That adversarial tuning is why FSD has driven billions of miles with collision rates improving: one major incident every approximately 5.3 million miles in recent supervised data. It is impressive. But it is also the exact handicap that worries me for Optimus.
Optimus is not a car. It is not an elevator. It will operate in the intimate, unpredictable spaces of human life—factories, homes, offices—where it must hand tools, fold laundry, walk crowded floors, and interact gently with people and objects.
Here is where the engineering record demands scrutiny. Optimus Gen 2's specifications are worth examining on paper: the platform stands 1.73 meters tall, weighs approximately 57 kilograms—10 kilograms lighter than Gen 1—and achieves a 30% improvement in walking speed over its predecessor, now approaching 0.68 m/s on flat terrain. Its hands alone represent a serious leap: 22 degrees of freedom across 11 actuated joints per hand, driven by 6 independently actuated fingers with tactile sensors embedded in the fingertips, capable of a 5 kg payload per hand. Whole-body articulation reaches 28+ degrees of freedom with 50 actuators total. The AI5 compute chip delivers 4× the inference throughput of AI4, running the same vision stack—8 cameras—that powers FSD. During December 2023 demonstrations, Gen 2 sorted objects by color, handled eggs without breakage, and performed controlled yoga sequences, all of which signal impressive closed-loop dexterity.
— Exact force-compliance curves and maximum contact pressure limits under dynamic load
— Gait uniformity tolerances—the robotic equivalent of stair riser variation, quantifying step-to-step consistency across surfaces and fatigue states
— Whole-body jerk and acceleration profiles during close-quarters maneuvering
— Real-time tactile feedback thresholds that prevent crushing an egg one moment and bruising a child's arm the next
— Reaction latency under unexpected contact from a human body at close range
The 22-DoF hands can sense contact. The wrists carry force-torque sensors. The vision stack can detect obstacles at distance. These are the right ingredients. But without published limits, without a human interaction diagnostic framework, the world cannot verify that the system errs on the side of gentleness rather than efficiency.
Without those leading indicators, the inherited FSD mindset risks producing a robot that is brilliantly defensive in traffic-like scenarios but tentative, jerky, or dangerously inconsistent in close-quarters collaboration. A robot that "avoids death" perfectly may still cause thousands of minor injuries, dropped objects, or ergonomic strain on human coworkers—exactly the way a poorly spaced staircase or a bouncy elevator degrades quality without ever killing anyone. Public trust will evaporate. Regulators will pounce. And the very mission of safe, abundant humanoid labor could be set back years.
I am not being theoretical. You have repeatedly said Optimus will be "the biggest product ever made." You plan to scale it into factories, then homes. The same data flywheel that improved FSD can improve Optimus—but only if you deliberately expand the safety diagnostic beyond collision rates and fatalities. The 50-actuator platform, the tactile fingertips, the force-torque wrists—these are the hardware levers for a gentleness standard. Use them. Demand measurable uniformity in motion, ride-quality metrics stricter than any elevator, force-limiting envelopes that are hardware-enforced, not just softcoded. Publish the full Optimus safety spec sheet immediately—force curves, gait precision tolerances, contact pressure limits, injury-prevention validation data—so the world can see you are not gambling with human bodies.
I urge you, in the strongest terms: recalibrate now. Do not let the road-war mindset distort the design of a machine that must coexist with us. Safety first is not optional. It is the only way Optimus earns the right to walk among humanity. Ignore the subtler diagnostics—spacing, smoothness, consistent quality—and you risk turning a revolutionary technology into a cautionary tale of injuries, lawsuits, and lost trust. Something along the lines of a Human Interaction Diagnostic standard is needed. Feedback alone marginally improves what will require tailored, use-case-specific calibration for every deployment context.
Tesla and xAI exist to benefit humanity for all intents and purposes. Prove it by making safety the uncompromising North Star, measured by the right metrics from day one. I await the spec sheet. The world is watching.
~A
Concerned Engineer and Neoplaneteer
