synthesized from several pieces of Allan's writings & conversations by ChatGPT-5.6 Sol
We already have vibe coding. A novice says, “Make me a game with a mute button,” and an AI drafts the program, runs it, sees what broke, and tries again. The user does not need to describe the class hierarchy, choose a framework, or explain where the event handler belongs. They describe the thing they want, complain about the result, and gradually steer the machine toward something usable.
It works unusually well because code is text. The request is text, the model's native medium is text, and the product is text. Even the error messages are text. The loop closes naturally.
The world, inconveniently, is not made of text.
The world of things is made of structures, circuits, materials, fluids, machines, heat, tolerances, stresses, noise, wear, and the persistent refusal of physical reality to be persuaded by eloquence. You cannot prompt a bridge into carrying twice its rated load. You cannot ask a battery to become lighter in the same tone of voice that persuaded the AI to change a button color.
But you can ask an AI to design, simulate, criticize, and revise.
That is Vibe Engineering: describe what you want in ordinary language and let the AI try to make it physically real.
From a Napkin to a Solver#
Suppose I say, “I want a vest that keeps me cool on a hot day using the Peltier effect.”
An engineer hears many unstated questions. How cool? For how long? In what ambient temperature and humidity? How much may the vest weigh? Where does the rejected heat go? How large is the battery? What happens when moisture meets electronics? Is the object meant for a construction worker, an athlete, or someone sitting in a wheelchair?
A present-day chatbot will happily produce a design-shaped answer. It may estimate the thermodynamics, recommend components, and draw a plausible diagram. Plausibility is not engineering.
A Vibe Engineering system would turn the request into a continuing experiment. It would generate candidate layouts, produce CAD or schematic files, send them to thermal and electrical simulators, inspect the results, discover that the first battery lasts eleven minutes or that the hot side cooks the wearer, and revise. It would return not merely an answer but evidence: expected cooling capacity, weight, battery life, temperatures, cost, and the assumptions on which these depend.
The simulator is not quite an oracle—simulators contain models, discretizations, boundary conditions, and mistakes—but it is a useful source of resistance. Language can flatter an idea indefinitely. Physics eventually says no.
This creates a loop:
- Describe an intention in plain language.
- Translate it into candidate geometries, materials, circuits, and parameters.
- Run the candidates through the relevant simulation tools.
- Interpret failures and performance.
- Revise, branch, or abandon.
- Show the human what changed and what trade-offs remain.
The point is not a perfect first pass. The point is iteration tethered to a world outside language.
We Do Not Need New Physics#
Much of the necessary machinery already exists. Engineers use CAD packages, finite-element solvers, computational fluid dynamics, circuit simulators, materials databases, optimization systems, and manufacturing tools. These systems are powerful, mature, and largely inaccessible to anyone who has not spent years learning their interfaces and conceptual vocabularies.
The barrier is not always the physics. Often it is the driver.
Most engineering tools were built for a person sitting at a workstation, selecting surfaces with a mouse, entering parameters into dialog boxes, constructing meshes, choosing boundary conditions, and interpreting plots. To use them, you must know the language of the tool before you can ask the question that brought you to it.
Vibe Engineering requires a different interface layer:
- APIs that let AI operate CAD and simulation tools reliably rather than imitate mouse clicks.
- Translation from ordinary intentions into explicit engineering variables and constraints.
- Model-generated schematics, meshes, solver inputs, and test plans.
- Feedback loops in which AI reads solver output and revises the candidate design.
- Explanations that turn stress maps and convergence failures into consequences a non-specialist can understand.
None of these is trivial. Together they are still narrower than “invent a machine that understands all engineering.” We can begin with one domain, one toolchain, and one class of artifact. A system that reliably takes a cooling-device request through CAD, thermal simulation, design alternatives, and an intelligible report would establish the pattern.
Vibe coding did not begin by replacing every software engineer. It began by making the loop visible.
Let the Ridiculous Requests In#
The examples should include serious needs, but not only serious ones. Whimsy is a useful test because it exposes the difference between a system that retrieves familiar designs and one that can explore a space.
“Build an air cleaner that clears catbox odor before guests arrive.” The AI has to turn embarrassment into an airflow requirement, choose a fan and filter medium, estimate adsorption, consider noise, and discover whether “five minutes” is compatible with the size of the room.
“Design a helicopter that fits in my garage.” Perhaps the answer is no. Rotor diameter, power, stability, downwash, and safety may make the request absurd. A useful system does not conceal the absurdity beneath an attractive rendering. It shows which constraint kills the design and what would have to change: the garage, the payload, the word helicopter, or several laws of physics.
“Give me a mat I can roll across a table, drop my phone or watch anywhere on it, and charge only at the right places.” The system can explore coil arrays, switching networks, localization, shielding, heat, and cost. Maybe it finds a workable architecture. Maybe it finds that the product exists. Maybe it finds that the desired flexibility makes it uneconomic. All are better outcomes than a confident sketch.
“Give me a compact heat sink for a home battery array.” Now the request is ordinary and consequential. The same loop explores fin geometries, materials, airflow, ambient conditions, manufacturability, and failure modes.
The range matters. Vibe Engineering should turn everyday language into design experiments, including experiments whose result is “you cannot have all of those things at once.”
Not One Design, a Landscape#
The first version of this idea imagined a single loop converging on a solution. That is still too much like ordering from a machine.
Engineering rarely has one right answer. It has a landscape of compromises. The lighter cooling vest is noisier. The quiet one costs more. The cheap one runs for forty minutes. The long-running one looks like camping equipment for a small moon expedition.
A useful system should produce families of designs, not one artifact presented with divine confidence. It should vary materials, topology, dimensions, and operating assumptions deliberately. It should map the trade-offs and show the user a Pareto frontier: designs for which improving one desired property necessarily worsens another.
The phrase “Pareto frontier” is not useful to most people who merely wanted their catbox to smell better. The system should translate it.
This design is cheapest but loudest. This one is quiet enough for a bedroom but clears the room in twenty minutes rather than five. This one meets the five-minute target but occupies half a wall. Which compromise describes what you actually want?
The human can then steer at the level where human preference matters. “Make it quieter.” “I do not care what it looks like.” “Reduce the price even if it becomes heavier.” “Show me two radical alternatives instead of polishing this one.”
The AI undertakes the tedious exploration. The human chooses which discomforts are tolerable.
This is more than a friendly user interface. It changes the relationship between a non-expert and engineering knowledge. The user does not need to know the parameters in advance, but remains responsible for preference and purpose. The machine can reveal that a vague desire contains incompatible demands. The conversation helps the person discover the specification rather than pretending the specification existed fully formed at the beginning.
The Human Loop Is Not Ceremonial#
“Human in the loop” is an abused phrase. Sometimes it means a person retains meaningful authority. Sometimes it means a person is available to blame.
In Vibe Engineering, the human role should be substantive but not artificially heroic. A non-expert cannot certify a stress calculation by admiring the colors. An expert cannot personally inspect every variation generated at machine speed. We need layers of responsibility appropriate to the stage of the work.
During exploration, the human supplies goals, rejects unacceptable trade-offs, and decides which branches deserve more effort. The AI and simulators do the grind. Before fabrication, domain experts assess assumptions, tool validity, safety margins, regulations, and failure modes. Before deployment in a consequential environment, testing moves from simulation to prototypes, hardware-in-the-loop, destructive tests, and whatever certification the domain requires.
Simulation is evidence, not absolution.
The system should also explain uncertainty. Material properties vary. Models omit effects. Boundary conditions may be guesses. A result that says “factor of safety 2.1” without showing sensitivity to assumptions is a numeral wearing an engineer costume.
The promise is not that anyone can safely manufacture an aircraft after a pleasant conversation. It is that anyone can begin the engineering conversation, reach a level of technical specificity that was previously inaccessible, and bring a much more developed idea to the experts and machines required for the next stage.
Engineers do not vanish. Their role moves upward and outward: choosing models, validating tools, recognizing pathologies, curating design spaces, and deciding when simulated evidence is sufficient to justify touching matter.
Mutation, Not Merely Optimization#
An optimizer tends to improve the design it already has. A creative engineering system should sometimes leave that neighborhood.
Genetic algorithms offer a suggestive metaphor. Maintain a population of candidates. Vary them. Combine features. Preserve several lineages rather than allowing an early winner to consume the search. Introduce mutations whose purpose is exploration, not immediate improvement.
Language models may help precisely because they can propose structural changes that a conventional numerical optimizer would never represent. “What if the heat exchanger becomes part of the garment's frame?” “What if the filter is distributed through furniture rather than placed in a box?” “What if the charging mat is modular instead of continuous?” These proposals may be foolish. Simulation makes foolishness cheap.
The system should not mistake novelty for merit. It should use language to expand the space and physics to contract it again.
That alternation—imagination, resistance, revision—is the heart of the project.
From Simulation to Matter#
The eventual loop does not end with a CAD file.
Designs can flow to 3D printers, CNC machines, circuit-board fabrication, robotic assembly, and automated test equipment. Sensors on the resulting prototype can feed measurements back into the model. The simulation's theory of the artifact meets the artifact. Differences update the next design.
At that point Vibe Engineering becomes an AI-native experimental workshop:
Describe. Design. Simulate. Build. Measure. Revise.
Automatic and instant fabrication belongs to a different manifesto, or perhaps to the last page of this one where manifestos traditionally become irresponsible. For now, even a loop that ends in a validated, manufacturable design would collapse an enormous distance between intention and experiment.
Who Gets to Ask?#
Vibe Engineering is not fundamentally about replacing engineers. It is about expanding who can participate before the expensive part begins.
At present, many ideas die before they reach a testable form. The person with the need cannot express it in engineering language. The expert who could help is expensive or unavailable. The tool requires training. The first simulation requires a mesh. The sketch remains a sketch.
An AI that drives existing tools can lower the price of asking “could this work?” It can help an amateur turn discomfort into requirements, requirements into candidates, and candidates into evidence. It can also explain why a cherished idea fails without requiring the person to earn a degree merely to receive the bad news.
There will be ridiculous outputs, abused tools, converged simulations of impossible objects, and products whose elegance survives exactly until the first screw is tightened. Vibe coding has bugs; Vibe Engineering will have gravity.
Good. That is why the loop must touch reality early and often.
If vibe coding gave us software from words, Vibe Engineering can give us physical experiments from words. The artifact may be a CAD model, a circuit, a blueprint, a toolpath, or a prototype. The path begins the same way:
“Here is what I want.”
Let the AI engineer it. Let the person steer it. Let simulation reveal the trade-offs.
And, eventually, let matter have the last word.