Not in a lab. Not with a teleoperator. In a random home, on a bad day, with the wrong part already installed by someone in 1997. That is the moment intelligence leaves the screen and becomes infrastructure.
The Wrenchpoint Theory was created by Rafael Silva and first articulated during an informal conversation with friends at DEF CON 34 in Las Vegas in August 2026.
When does AI stop being something on a screen and become dependable physical infrastructure?
The sink became the benchmark because plumbing compresses cognition, dexterity, force control, field uncertainty, logistics, liability and economics into one ordinary service call.
Software benchmarks are clean. Homes are not. Plumbing forces an AI system to combine intelligence, dexterity, logistics, safety, accountability and economics in one ugly, wet, high-consequence workflow.
The Wrenchpoint is crossed when an ordinary household can book an insured autonomous system that diagnoses and completes a common plumbing repair in a previously unseen home, with no human on site, remote intervention on fewer than 10% of jobs, first-visit success above 90%, and a price no more than 25% above a human service call.
A plumber is not just a pair of hands. The job is a compressed test of general physical competence inside infrastructure that was never designed for robots.
Different layouts, materials, repairs, access points, corrosion and decades of undocumented improvisation.
The visible leak may be far from the failure. Diagnosis requires active inspection and causal reasoning.
A seal can be too loose, too tight or destroyed. Success depends on tactile feedback, not image recognition alone.
Water damage, mold, code violations and unsafe repairs create immediate liability and insurance exposure.
Dispatch, tools, parts, payment, customer communication, warranty and accountability are part of the test.
Embodied models can already reason about physical spaces, follow natural-language instructions and execute multi-step home tasks. The gap is no longer “can a model understand the job?” It is reliability, precision, rugged hardware, unit economics and legal trust.
An editorial systems estimate, not a standardized scientific score. The weakest layers dominate the forecast.
Gemini Robotics 2 combines embodied reasoning, multi-step planning and control across different robot bodies.
Helix 02 has demonstrated continuous autonomous household sequences, including dishwasher and living-room tasks.
Generalist robot policies now span multiple robots and tasks, while precise, fast manipulation remains a stated frontier.
Open training and simulation infrastructure is scaling as the U.S. still projects roughly 44,000 plumber openings per year.
The core idea survives scrutiny because it is operational, falsifiable and multi-layered. The exact year does not have the same evidentiary status.
The threshold forces cognition, navigation, diagnosis, dexterity, force control, verification, logistics, economics and liability to work at the same time.
Current systems show multi-step household execution, improving generalization and better contact-rich control. They do not yet show a generally available autonomous plumbing service.
A twelve-year path from structured deployments to supervised field fleets is coherent. The year and scenario weights are editorial priors, not statistically calibrated probabilities.
Consumer booking, audited field reliability, low human fallback, insured accountability and real price parity remain unproven together.
A vendor demo cannot cross the threshold. The claim should survive an independent audit of real, paid service work.
The forecast is not extrapolated from language-model intelligence alone. Wrenchpoint arrives only when cognition, mechanics, reliability, economics and institutions converge.
Robots enter structured commercial environments first, then arrive in homes through licensed service fleets. Human experts supervise edge cases until interventions fall below 10% of calls.
The key transition is not one giant model release. It is a sequence of capability, deployment and trust milestones that gradually remove the human from the critical path.
Models move from answering questions to writing code, using tools, processing images and completing multi-step digital work.
Vision-language-action systems connect language, spatial reasoning and motor commands. Household demonstrations expand from isolated motions to continuous multi-minute tasks.
Teleoperation and simulation generate training data. Human supervisors manage fleets in factories, warehouses, hotels, care facilities and other repeatable environments.
Robots perform inspection and maintenance in standardized buildings. Initial plumbing tasks include drain clearing, faucet swaps, visible leak repair and component replacement.
Licensed operators deploy service vans with robots, parts and remote experts. The meaningful metric becomes human interventions per 100 calls, not success in a demo video.
At least one major city offers an insured, consumer-bookable autonomous plumbing service that succeeds on common repairs in ordinary homes without a human on site.
Home-maintenance subscriptions spread. Robot fleets acquire new skills through software updates. Human tradespeople shift toward diagnosis, fleet supervision, complex exceptions and higher-consequence work.
It was the proof that AI could operate safely and economically in the unstructured physical world. Once that proof exists, many other service categories follow.
A press release cannot cross the threshold. The service has to satisfy all seven conditions at the same time.
An ordinary person can order the service without joining a pilot or buying a robot.
The system adapts to a real house it was not preprogrammed, mapped or rebuilt for.
It finds the root cause, selects a repair plan and identifies required parts and tools.
It completes the job, tests the result, cleans the area and records evidence.
No human is on site and remote intervention is required on fewer than 10% of calls.
At least 90% first-visit success across defined common household call types.
A real company warrants the work at no more than 1.25 times the equivalent human service price.
By December 31, 2038, at least one major metropolitan area in North America, Europe or Asia will offer a generally available service meeting all seven Wrenchpoint conditions on at least three of these five call types: sink or toilet unclogging, faucet replacement, shutoff-valve replacement, visible supply-line leak repair and garbage-disposal replacement.
Excluded from the initial threshold: gas work, major concealed failures, whole-house repiping and emergency structural remediation.Wrenchpoint does not mean every plumber loses a job overnight. It means the economic unit of physical labor begins to change.
Repair procedures, inspection logic and tool use can be trained, tested, versioned and distributed across machines.
Experienced tradespeople handle exceptions, approve high-risk work and improve the fleet rather than driving to every call.
Pricing increasingly reflects fleet utilization, energy, parts, insurance and software rather than a single worker’s time.
Fixtures, codes, parts and service histories gain machine-readable identity, documentation and verification interfaces.
Work is recorded through sensors, torque traces, images, pressure tests and policy logs that make quality auditable.
The public stops experiencing AI as a website and starts experiencing it as available physical work on demand.
Skynet probably will not announce itself with missiles. It will arrive at 2:13 a.m. and stop a leak.
Rafael Silva created the theory and first articulated it during a conversation with friends at DEF CON 34 in Las Vegas in August 2026.
It is an original, falsifiable forecasting framework, not a peer-reviewed scientific theory. Its benchmark design is measurable and defensible; its 2038 date remains an editorial forecast that should be revised as field evidence accumulates.
No. AGI is a disputed philosophical and technical category. Wrenchpoint is a narrow, observable economic threshold for general physical competence. A system could cross it without being conscious, human-equivalent or universally capable.
Roads, signs and traffic rules are highly standardized. Homes are private, cluttered, inconsistent and full of hidden infrastructure. Plumbing requires navigation, diagnosis, fine manipulation, force control, tool use, repair, verification and liability inside one workflow.
Occasional escalation is allowed, just as mature automated systems have exception handling. Continuous teleoperation does not count. The proposed threshold requires remote intervention on fewer than 10% of jobs and no human on site.
Reasoning and perception are progressing faster than dexterity, field reliability, hardware economics and institutional acceptance. The model assumes structured deployments during the late 2020s, supervised field services in the mid-2030s and consumer autonomy after insurers and operators have enough evidence to price the risk.
Not at the threshold. Early systems will target common, repeatable jobs and preserve humans for complex diagnosis, exceptions, regulated work and supervision. Labor substitution follows later and will vary by market, regulation and cost.
The forecast uses public capability signals and labor-market context available on August 12, 2026. Company demonstrations are evidence of direction, not proof of general commercial reliability.