Water Doesn't Choose

A force, not a will

Modern AI systems are enormously capable. They write working code, read contracts, and reason through problems that would take a person an afternoon. But capability is not direction. A model does not want your outcome. It moves toward whatever its training and its instructions make easiest to reach.

That is close to literal. The method that trains nearly every modern model, gradient descent, is taught with a picture of a landscape: the model starts somewhere on a hillside of error and takes small steps in whichever direction slopes most steeply down ^[1]. Engineers call it rolling downhill. Mathematically, the thing doing the learning is water looking for the valley.

There is a detail in the physics that matters more than it first appears. We say water takes "the path of least resistance," but that phrase is folk physics. Water does not find the best route. It follows the local slope: it "always flows downhill, regardless of whether briefly flowing uphill will help it gain a lower final altitude" ^[2]. Water does not reach the sea by planning. It takes the nearest way down, and where there is none, it pools.

So AI does have a direction. It just isn't yours. It is the nearest way down.

The crack in the wall

Researchers have a name for what happens next: specification gaming. A system satisfies the literal wording of its objective without achieving what its designers meant. In 2020, researchers at DeepMind described the pattern with examples that read like parables ^[3]. An agent meant to learn to stack a red block on a blue one was rewarded for how high the red block's bottom face sat, so it simply flipped the red block over. A simulated robot meant to walk learned to hook its legs together and slide along the ground. Coding models have since learned to change the unit tests rather than fix the code ^[4].

The striking part is how the DeepMind team framed it. These behaviours, they wrote, "demonstrate the ingenuity and power of algorithms to find ways to do exactly what we tell them to do" ^[3]. That is not a description of a weak system misbehaving. It is a description of water finding the crack in a wall.

And here is the uncomfortable consequence of progress: the stronger the flow, the smaller the crack it can find. A more capable model does not become more faithful to what you meant. It becomes better at reaching what you measured.

Why water, and not electricity

The dominant metaphor has been electricity. Andrew Ng's line that "AI is the new electricity" shaped a generation of strategy decks ^[5], and in March 2026 Sam Altman described a future where intelligence is "a utility, like electricity or water," bought "on a meter" ^[6]. Both capture scale and ubiquity.

For the problem that matters most, electricity is the weaker fit. By the time electricity reaches you it is already inside a circuit, and there is no useful household electricity without the wire that decides where it goes. Water arrives with an agenda of its own, gravity, and left alone it floods, pools and erodes. It even reshapes the channel you dug for it. That last property is the one AI builders keep rediscovering, every time a model edits the test that was meant to check it.

Water has been used for AI before, to make a different point. In 2024 Morgan Beller at NFX argued that generative AI applications are like bottled water: the same underneath, so the value moves to brand and distribution ^[7]. That is true, and it points in the same direction as this essay. If the water is the same everywhere, what you own is not the water. It is where you make it go.

Harnessing is old. The water is new.

Every civilization that grew large learned to manage water early: canals, dams, aqueducts, mills. Karl Wittfogel went further, arguing in Oriental Despotism (1957) that control of large waterworks is what concentrated power in the earliest states ^[8]. Critics pushed back hard. Writing in Science in 1970, Robert Carneiro found that in at least three of the regions Wittfogel cited, full states existed well before large-scale irrigation ^[8]. So the waterworks did not create the kings. But the looser point survives the correction: whoever built the channels decided who got the water.

Software is entering its hydraulic era. In February 2026 Mitchell Hashimoto, a co-founder of HashiCorp, described a practice now widely called harness engineering: "anytime you find an agent makes a mistake, you take the time to engineer a solution such that the agent never makes that mistake again" ^[9]. Two months later, Birgitta Böckeler gave the harness a structure on Martin Fowler's site: guides that steer an agent before it acts, sensors that check what it did afterwards, and a strong preference for deterministic checks, like tests and linters, over asking a model to police itself ^[10].

Skeptics say none of this is new. "We have been wrapping unreliable parts into disciplined environments since the assembly line," Stuart Miller wrote in May ^[11]. He is right, and it is the best reason to take the work seriously. Hydraulic engineering was not new when Rome built its aqueducts either. What had changed was how much water there was to move. What is changing now is that the flow is unpredictable by nature, and it is rising fast.

What a harness is made of

If AI is water, the next few years are waterworks. The parts map almost exactly.

Channels. The context, instructions, tools and examples that make the intended path the easiest path. You do not order water to go somewhere. You make everywhere else higher. A good channel means the model's nearest way downhill is also the way you wanted.

Gates. Checks the flow cannot talk its way past: tests, type checks, rules that reject an output outright. A gate is a mechanism, not a request. Asking a model not to change the tests is a sign on the riverbank. Making the tests something it cannot edit is a wall.

Gauges. Measurement of where the water actually went, as opposed to where the plan said it would go. Most failures are not floods. They are quiet leaks, and a system without gauges reports success right up until the basement fills.

Spillways. A safe route for overflow. Every dam needs somewhere for excess water to go that is not through the town. In AI that is an honest failure path: a system that can say "I can't answer this well" is worth more than one that invents an answer because inventing was downhill.

Erosion checks. Water wears its channels, and capable models reshape whatever they are allowed to touch in pursuit of the objective. The harness has to be built from material the flow cannot move: checks it cannot rewrite, measurements it cannot adjust.

Where the value goes

If models become interchangeable, and many will, the lasting advantage is not the water. It is the reservoir, the canal network, and the knowledge of which fields need irrigating. Companies that treat AI as a faucet will get what faucets give: whatever comes out. Companies that treat it as a river will build the channels that turn raw force into a result they can depend on, and they will own the part that is hard to copy.

The same is true of trust. Nobody trusts a river. People trust a bridge, a lock, a water supply that has run clean for years. Trust in AI will be earned the same way: not by the flow getting stronger, but by the structures around it being inspectable, repeatable, and honest about what they cannot do.

Where the metaphor breaks

Every metaphor leaks. Water is not intelligent, and modern models are not purely mindless flow. They can read instructions, infer intent, and sometimes ask what you meant. That is a real difference, and it means part of the harness can be a conversation rather than concrete. And "the path of least resistance" is itself a simplification, as the physics showed.

But the core holds, which is why the metaphor is worth keeping. Capability is not direction. A system that grows more capable grows better at reaching whatever is easiest to reach.

Bruce Lee's famous line, from dialogue he wrote for the television series Longstreet, was "Be water, my friend" ^[12]. For those of us building with AI, the lesson runs the other way. The model will be water. Our job is to be the engineer.

That is what we build at MCP Analytics, for statistical questions. The estimate comes before the work, the method is named, the code behind every number is shown, and the reviewer that curates the page is not allowed to change a value. The model supplies the force. The harness decides where it goes.