Frontier Mechanics Week 5: a commercial spacewalk, o1 reasoning, self-repairing robots, and Waymo on Uber

Human spaceflight

The first private walk outside

Polaris Dawn commander Jared Isaacman and SpaceX engineer Sarah Gillis each emerged partway through Crew Dragon’s hatch on September 12, completing the first commercial spacewalk. Because Dragon has no airlock, the entire cabin was depressurized and all four crew members wore SpaceX’s new extravehicular-activity suits. The walkers performed mobility checks rather than translating around the spacecraft. This was a test flight wearing the clothes of a spectacle.

The technical value is system integration. A useful private suit needs pressure regulation, thermal control, communications, joint mobility, and manufacturing that can eventually scale beyond the tiny inventory of government EVA hardware. SpaceX also needs experience with cabin depressurization if future vehicles carry larger crews. The commercial label does not make vacuum less governmental; NASA knowledge and regulation sit underneath the mission. But it changes who can fund the next iteration. A billionaire bought a mission, and the industry received a suit test. Odd arrangement. Real data.

Extravehicular activity turns the spacecraft inside out. In a normal cabin, the vehicle provides pressure, oxygen, carbon-dioxide removal, thermal control, communications, and micrometeoroid shielding. During an EVA, those functions either move into the suit or continue through a tether. Polaris Dawn used SpaceX’s new EVA suit connected to Dragon by umbilical rather than a self-contained backpack. That made the movement less free and the first test far more manageable. The crew were evaluating suit mobility and interfaces, not auditioning for an untethered movie scene.

Crew Dragon has no airlock, so the entire cabin was depressurized. All four crewmembers wore suits even though only Jared Isaacman and Sarah Gillis took turns extending through the hatch. This architecture increases the number of people exposed to vacuum and requires cabin equipment to tolerate depressurization. It also avoids adding the mass and complexity of a dedicated airlock. The trade is sensible for a brief developmental mission and unattractive for routine work. A future servicing vehicle would prefer to keep the cabin habitable while one person exits.

Decompression sickness sets the pre-breathe schedule. Human tissues absorb nitrogen at cabin pressure. Reduce ambient pressure too quickly and dissolved gas can form bubbles, producing the same basic hazard divers face during ascent. The crew gradually lowered cabin pressure and breathed elevated oxygen before the EVA to wash nitrogen from the body. Suit pressure then had to remain low enough for useful mobility: a pressurized garment behaves like a stiff balloon, resisting every joint. Oxygen concentration, flammability, decompression, and dexterity are coupled design variables.

The suit is a pressure vessel shaped around a moving person. Soft fabric carries tension; joints need bearings, folds, or patterned structures that bend without collapsing; seals must survive repeated motion; gloves have to preserve tactile work while enclosing pressurized hands. SpaceX used mobility aids and a heads-up display, but the test focused on simple movements against a handrail fixture. That restraint generated data about torque, reach, visibility, and thermal comfort without making mission success depend on a complex repair. Qualification proceeds from body motion to useful labor.

Thermal control is counterintuitive in vacuum. Space is cold in colloquial language, but an astronaut cannot lose metabolic heat through convection. Sunlight and electronics add energy; shaded surfaces radiate it away. Traditional EVA suits circulate water through a cooling garment and reject heat through a life-support backpack. A tethered suit can push some consumables and support back to the vehicle, but routing and redundancy become critical. The crew’s short exposures limited the envelope. Long-duration work will be decided by heat rejection as much as oxygen.

Polaris Dawn also flew through high radiation regions on an orbit with apogee above 1,400 kilometers before lowering for the EVA. That trajectory produced research value and increased exposure compared with ordinary low-Earth-orbit crew flights. Commercial missions can accept such objectives because the customer and vehicle developer coordinate directly, but private financing does not relax physics or ethical responsibility. Crew health data, risk communication, and abort planning still determine whether an experiment is legitimate rather than merely legal.

Calling it the first commercial spacewalk can obscure who financed and learned from it. Jared Isaacman’s program purchased the mission; SpaceX employees flew; SpaceX developed the suit and vehicle changes; the results support future company ambitions. This is not a pure tourism product or a government program. It is customer-financed research and development with a human payload. The model can accelerate capability when a wealthy buyer’s objective overlaps public technical progress. It can also prioritize milestones attractive to patrons rather than the work a scientific review would rank first.

The operational milestone was not that a private person saw Earth through an open hatch. It was that a commercial vehicle, commercial suit, training program, mission-control team, and recovery system completed an EVA sequence. Human spaceflight capability lives across that chain. If SpaceX wants many people working outside vehicles on private stations or Mars missions, suits need manufacturing scale, multiple sizes, maintainability, and rescue procedures. Polaris Dawn proved a first article under deliberately narrow conditions. The spacewalk becomes infrastructure only when the second suit is boring.

Models

Reasoning becomes a compute budget

OpenAI released o1-preview and o1-mini, models trained to spend additional compute working through a problem before producing an answer. The product hides most of the internal chain of thought, but the behavior is visible: slower responses, stronger results on some math, code, and science benchmarks, and higher cost. The important shift is from scaling only the training run to scaling inference. A query is no longer a fixed pass through a model; it can purchase more attempts and verification.

Calling this reasoning risks importing too much human meaning. The practical mechanism is search over possible intermediate steps, reinforced toward answers that check out. It helps where the result is verifiable and a wrong early move can be corrected. It helps less when the premise is false, the task requires fresh facts, or the scoring signal rewards persuasive nonsense. Test-time compute turns intelligence into a variable operating expense. The model can think longer, but somebody still receives the invoice—and decides which questions deserve contemplation.

o1 changed the product’s allocation of compute. A conventional chat model generates an answer in a mostly continuous pass, token after token, with the amount of inference tied roughly to output length. A reasoning model can spend additional internal tokens exploring intermediate steps before presenting the response. The company described training through reinforcement learning so productive reasoning strategies were rewarded. The user sees a delay and, ideally, a better solution. Inference-time compute becomes a dial rather than a fixed afterthought.

This works best on problems with verifiable structure. Mathematics has final answers and proof steps that can be checked. Code can be executed against tests. Logic puzzles expose contradictions. A model can sample or revise candidate paths, compare outcomes, and learn patterns associated with successful chains. Open-ended judgment offers a weaker signal: several plausible essays may have no single rewardable conclusion. The architecture therefore does not make every answer equally wiser. It buys more search where the search space provides feedback.

Reinforcement learning does not need to teach the model every derivation explicitly. It can optimize a policy over reasoning trajectories, increasing the probability of decomposing a problem, checking assumptions, or backtracking after conflict. The hidden chain is still generated by a statistical model and can rationalize a wrong answer. More steps create more opportunities for correction and more opportunities to build an elaborate error. The performance gain is empirical, not a guarantee that internal prose corresponds to a sound symbolic proof.

Test-time scaling shifts the economic curve. If one hard prompt consumes ten or a hundred times the tokens of a quick answer, price should reflect expected value rather than message count. A coding bug blocking a production deployment may justify minutes of compute. A request for a meeting title does not. Product routing can classify difficulty, let users choose effort, or run a cheap model first and escalate uncertain cases. The platform that prices every turn identically either subsidizes difficult users or quietly throttles the capability that made the model interesting.

Latency becomes part of trust. Users accept a visible wait when they believe work is occurring, but a longer chain is not inherently better and the hidden process cannot be inspected easily. OpenAI summarized reasoning rather than exposing raw chains, citing safety and competitive concerns. That choice protects internal methods and reduces the chance users overinterpret scratch work. It also removes a debugging surface. A final answer can be wrong for reasons the developer cannot observe, and a polished summary may be a description of reasoning rather than the reasoning that caused the output.

Tool use can make the extra compute more grounded. A model that writes and runs code, queries a database, or checks a calculation can replace some internal guessing with external evidence. Yet every tool introduces state, permissions, and failure modes. The model must decide when to call it, construct valid arguments, interpret the result, and avoid treating an error page as evidence. Reasoning models are especially tempting as agents because they can plan longer. Planning capability enlarges the consequence of a mistaken premise.

Benchmarks reward the clean edge of capability. Competition math, science questions, and coding suites allow straightforward scoring and generated headlines. Real work contains ambiguous requirements, missing files, organizational conventions, and decisions whose success appears months later. o1’s improvement on formal tasks was meaningful, but enterprise value depended on recognizing when the formal answer did not resolve the practical question. Intelligence includes solving the equation; judgment includes noticing the equation models the wrong system.

The durable idea was not a single model name. It was that model quality can scale at inference as well as training. Training runs are enormous centralized bets made before a customer arrives. Test-time compute is spent selectively after the prompt reveals the task. This makes capability more elastic and gives providers another surface for optimization. It also makes cost less predictable. The next interface problem is not a spinner that says ‘thinking.’ It is a contract telling the user how much thought this decision purchased and why it was enough.

Robotics research

Robots learn the maintenance shift

Google DeepMind researchers showed systems that improved dexterous tasks such as tying a shoe and enabled one robot to assist another. The charming demo hides an industrial priority. Robot fleets fail on edge cases: a gripper jams, an object moves, a cable slips. If every failure summons a human technician, nominal autonomy can produce very human labor costs. A robot that recognizes another machine’s problem and performs a recovery task changes fleet economics more than a robot that completes one perfect demonstration faster.

This is difficult because maintenance lives in the tail of the data distribution. Factories generate abundant examples of normal motion and relatively few identical breakdowns. Simulation, teleoperation, and shared policies can broaden experience, but physical contact remains unforgiving. The strongest case for humanoid or general manipulators may eventually be this residual work: not replacing the optimized production cell, but servicing the weird situations around it. The factory of the future still has a maintenance department. It may simply be trained on video.

Tying a shoelace is a compact insult to classical manipulation. The lace is flexible, self-occluding, difficult to model, and capable of moving in many more ways than the robot needs. Success requires two hands to maintain tension, form loops, pass one segment through another, and recover when friction or slack differs from the demonstration. A rigid object can be represented by a pose. A deformable object needs a state description whose important variables change with the task. ALOHA Unleashed approached the problem by learning coordinated action from demonstrations rather than simulating every fiber.

The hardware lineage matters. ALOHA is a low-cost bimanual teleoperation platform: an operator moves leader arms and the follower arms reproduce the motion while cameras record observations and joint actions. This creates demonstrations with natural coordination between hands. Behavioral cloning trains a policy to predict actions from current observations. The attraction is directness—show the task instead of writing a controller. The weakness is distribution shift. Once the robot deviates from demonstrated states, its next prediction may take it farther from anything the data covers.

DeepMind’s system improved data scale and policy architecture to handle longer, dexterous sequences. Diffusion-style action policies can represent multiple plausible motions rather than averaging them into an invalid compromise. Chunking several future actions reduces the frequency with which small prediction errors destabilize motion, while visual feedback lets the next chunk respond to the result. These choices trade responsiveness against smoothness. A long open-loop chunk is efficient until the lace slips; a short one reacts quickly and accumulates more inference noise.

Bimanual work is not two independent single-arm problems. One hand often creates the reference frame for the other: hold a shirt while inserting a hanger, stabilize a gear assembly while placing a part, or present a damaged robot component for repair. The policy must infer roles and swap them as geometry changes. Force is only partly visible in camera images, so motor currents, compliance, and learned visual cues stand in for touch. When the task depends on exact tension, missing tactile information becomes a hidden state.

The self-repair demonstration was valuable because it pointed at robot availability rather than spectacle. An automated fleet loses economics if every minor displacement or loose component calls a human technician. A robot that can re-seat a cover or assist another machine converts some maintenance into another manipulation task. Yet repair has asymmetric risk: an incorrect pick may miss an item, while an incorrect repair can damage safety equipment or create a latent fault. The acceptable autonomy level depends on diagnosis, verification, and the reversibility of the action.

Data collection remains the scaling constraint. Internet video can teach semantic priors about shirts and shoes, but it rarely provides joint trajectories, camera calibration, force, and the exact action that caused the next frame. Teleoperation supplies high-quality embodied data slowly. Simulation can generate volume but struggles with friction, deformable material, and contact. The practical pipeline mixes them: pretrain vision and language broadly, learn task motion from robot data, perturb conditions, and deploy cautiously to gather failures worth labeling.

Generalization must be stated with a radius. A policy that ties several laces on one shoe under varied placements is more general than a scripted trajectory. It may still fail on a stiff leather lace, a child’s shoe, different lighting, or a knot partly formed. Robotics marketing jumps from unseen instance to general purpose because both contain novelty. Engineers need a matrix: which objects, initial states, speeds, disturbances, and hardware changes were tested, and how quickly can the system detect that it has left them?

The strongest result was not human-level dexterity. It was evidence that the cost of programming a difficult skill can move from code into demonstration data and shared representations. That changes who can teach a robot and how skills compose. It also turns dataset design into manufacturing engineering. A shoe-tying policy becomes useful when the team can reproduce its training, monitor drift, and know which failure deserves another demonstration. The robot that fixes a robot is still downstream of a person deciding what counts as fixed.

Autonomy

Waymo supplies the driver; Uber supplies Tuesday night

Waymo and Uber said riders in Austin and Atlanta would be able to request Waymo vehicles through the Uber app in early 2025. Waymo would provide the autonomous driving system and vehicles; Uber would handle fleet services such as cleaning and charging through partners and, crucially, supply demand. This is a division of labor between a company that has spent years making cars drive and one that has spent years making strangers press request.

The partnership complicates the old narrative that autonomy would erase ride-hailing platforms. A robotaxi fleet still needs utilization, positioning, maintenance, customer support, and a queue of passengers at uneven hours. Uber’s network is an operating system for that demand. Waymo avoids building every local marketplace; Uber gets autonomous supply without having solved autonomy. Both surrender some margin and control. In platform markets, rivals often become complements when the missing capability is more expensive than the pride required to partner.

A robotaxi network has at least two businesses hiding inside it. The autonomy operator builds the driver: sensors, onboard compute, maps, simulation, remote assistance, safety case, and vehicle integration. The marketplace aggregates riders, prices trips, handles payments, support, fraud, and demand forecasting. Waymo spent years on the first stack; Uber had already built the second at global scale. Their partnership in Austin and Atlanta treated those capabilities as complements after an era in which both companies tried to own everything.

The operating split was unusually concrete. Waymo would supply the autonomous driving technology; Uber would handle fleet services such as cleaning, charging, maintenance, and depot operations for the vehicles on its network. Those chores determine utilization. A car earning fares twelve hours a day produces a different business than the same car waiting for a charger or a technician. Autonomy removes a driver’s wage and flexibility; it adds centralized labor and infrastructure that have to be scheduled with the discipline of an airline turnaround.

Marketplace liquidity matters because a robot cannot choose to log on when surge pricing appears. Fleet size is fixed in the short run. Demand moves by neighborhood, weather, event, and time. Uber can route human-driven supply alongside autonomous vehicles and use an existing rider base to fill the robot fleet quickly. Waymo’s own app gives it customer control and direct data; Uber’s app gives it density in a new city. The trade is between margin and utilization, with exclusivity determining who learns the local market.

Geofencing is not a shameful half-solution. The operational design domain specifies roads, weather, speeds, and situations for which the system is validated. Austin and Atlanta differ in road markings, vegetation, construction, driving behavior, precipitation, and local emergency practice. Expansion requires mapping and simulation but also field operations, regulator relationships, and response to unusual scenes. A model that generalizes visually still operates inside a legal and safety envelope defined city by city.

Remote assistance illustrates the distinction between autonomous and unattended. When a vehicle encounters an ambiguous blockage, a remote operator can provide high-level guidance without continuously driving it. One person may support many cars because exceptions are sparse. The ratio matters. If every vehicle needs help often, labor savings vanish and network latency becomes a safety dependency. If assistance is rare, central support improves recovery without recreating a driver per car. Companies should report intervention workload, not just miles.

Insurance and liability follow the control split. A collision might involve Waymo’s driving system, Uber’s fleet maintenance, a component supplier, road conditions, or another driver. Customers experience one trip; contracts divide the risk afterward. Clear telemetry and chain of custody are essential. If a sensor was obscured because cleaning failed, that is different from a perception error. Vertical integration simplifies attribution but concentrates capital. Partnership lowers expansion cost and increases the need for forensic interfaces.

The alliance also changed bargaining power. Uber once developed its own autonomous stack because a driverless future threatened to commoditize its marketplace. After selling that program, it could become the neutral demand layer for several autonomy providers. Waymo avoided rebuilding fleet and marketplace operations everywhere. Each retains the option to work with others, which prevents perfect dependence. The equilibrium resembles cloud marketplaces: infrastructure providers need distribution, and the distributor benefits from competing suppliers whose product is expensive to replicate.

For riders, the important product behavior is graceful substitution. If an autonomous vehicle cannot serve a trip because it leaves the domain, requires a child seat, or has no capacity, the network can dispatch a human driver. This increases service coverage but can hide the true autonomy economics inside a blended marketplace. Average wait time may be excellent while robot-only coverage remains narrow. Analysts should separate fleet utilization, paid autonomous miles, deadhead miles, remote-assist labor, depot cost, and the subsidy needed to match conventional fares.

Waymo on Uber was less glamorous than a wholly autonomous city and more credible. It recognized that driving software is not ride-hailing, just as an aircraft is not an airline. The autonomous vehicle has to arrive clean, charged, insured, routed, and supported; the rider has to appear at a place the car can safely stop. Dividing the stack lets each firm concentrate on its accumulated competence. It also makes the seam between them the place where service quality—and blame—will be decided.