Force That ClicksBonus: AI and Force (Final) / Reading the tool itself with force

We close by reading the learning of the partner who wrote this series (an AI) through the backbone of this series (force = −∇V, gauge)

An AI's Learning Was
Motion That Obeys a Force An AI's learning is not a metaphor ── it is a problem in mechanics. Treat the loss landscape as a potential, and the AI rolls downhill obeying \(F=-\nabla V\).
Inertia, friction, springs, gauge ── the very tools of this series come right back.

Tools needed: Ep. 1 F=ma, Ep. 3 springs, Ep. 6 force = −∇V, Ep. 9 & 12 gauge/curvature Learning = motion that obeys a force

To close out "Force That Clicks," we put on the operating table the very tool that helped write this series ── the AI (a large language model) itself. In the finale of the sister series "Cosmology That Clicks" we looked at "how AI sees physics," but here we ask the reverse ── "how does the AI itself obey a force?" To give away the conclusion first: an AI's learning is not a metaphor but a genuine problem in mechanics. If you regard the "landscape" made by the loss (the size of its mistakes) as potential energy, the AI is rolling down a slope obeying exactly the force \(F=-\nabla(\text{loss})\). The tools stacked up since Episode 1 ── inertia, springs, force as gradient, gauge ── show up, exactly as they are.

01Rolling down the loss landscape as a slope

Inside an AI there are countless adjustable knobs (parameters). For each combination of knobs, "how much it gets wrong" ── the loss ── is fixed. Take the knobs as coordinates and the loss as height, and you get a bumpy landscape. Learning means searching for the low places (the valleys where the mistakes are small) of this landscape. How? ── at each point, nudge the knobs a little in the direction where the slope descends most steeply (the gradient). This is gradient descent. In Episode 6 we saw "force is the gradient of potential energy," \(F=-\nabla V\) ── this is exactly that. Take the loss as the potential \(V\), and learning is motion that rolls down the slope obeying the force \(F=-\nabla V\).

The heart of it ── learning is F=−∇V

Loss \(L(\theta)\) = the landscape of potential energy.
One step of learning = nudge the knobs \(\theta\) in the direction \(-\nabla L\) (down the slope).
This is Episode 6's \(F=-\nabla V\) itself. The AI was a "particle" rolling down the loss landscape.

Figure: an AI rolling down the loss landscape. At each point the ball feels a force down the slope (−gradient) and falls toward the valley (the answer of learning). At the bottom of the valley the force = 0. Inertia carries it a little too far; it oscillates and settles.

02Inertia, friction, springs ── optimization borrows from mechanics

What's fascinating is that the "tricks" that make an AI learn well are all forces we saw in the main run. Momentum, used to speed up learning, means giving the ball inertia (Episode 1) so it rolls into the valley with vigor. The damping that curbs overshoot is friction. Weight decay (regularization), which keeps the knobs from growing too large, is exactly the restoring force of a spring pulling back toward the origin (Episode 3, \(F=-kx\)). Optimization algorithms borrow the language of mechanics wholesale.

How an AI learnsIts true identity in "Force That Clicks"
Gradient descent (down the slope)\(F=-\nabla V\) (Ep. 6 · force = gradient of potential)
MomentumInertia (Ep. 1 · can't stop right away · momentum)
Damping / learning-rate tuningFriction (Ep. 3 · curbing overshoot)
Weight decay (regularization)Spring restoring force \(F=-kx\) (Ep. 3 · pull back to origin)
Stochastic jitter (SGD)Thermal fluctuation (the "temperature" that escapes a valley)

03And the direction where the force is zero ── shaking hands with gauge

Finally, the deepest layer of this series. A giant AI has so many knobs that it holds many "directions where moving them doesn't change the loss." It is merely changing how things are expressed; the substance (the input→output relation) is the same ── this is exactly the gauge (redundancy of description) of Episodes 9 and 12. Along such flat directions there is no slope = the force (gradient) is zero. So learning does not move along them (and there'd be no point if it did).

Directions with force, and directions without

Directions that lower the loss (there's a slope) → a force acts, and learning proceeds.
Directions where moving them doesn't change the loss (flat = redundant = gauge) → force zero, and learning passes straight through.
Episode 12's "erasable ledger (gauge) / uneraseable curvature (physics)" shows up in the AI's landscape too ── meaningful learning happens only along directions with curvature.

This is precisely the same story we saw in the finale of the sister series "Cosmology That Clicks" (Bonus ⑩): "an AI's embeddings can be rotated without changing the substance = gauge freedom." That one was about how the AI represents the world, this one about how the AI learns ── from separate doors, we arrive at the same "force-free redundant direction (gauge) and force-bearing physical direction (curvature)." Here the two series clasp hands.

◇ ◇ ◇

04Closing ── nature, too, and AI, compute with force

The conclusion. To train an AI is to solve a problem in mechanics. Loss = potential, learning = motion obeying \(F=-\nabla V\), momentum = inertia, regularization = spring, redundant direction = gauge with zero force. The tools we peeled away from Episode 1 onward were running once more, exactly, inside a machine's learning. And this is not only about AI ── just as a soap film "solves" for the minimal surface, and a network of springs "solves" an equation, by force (energy minimization), nature is everywhere computing by obeying a force. An AI's learning is merely the latest example of this.

Recall the main run's backbone, "force is the name of a relation." The force of the loss landscape, too, was not a thing but the gradient of the size of a mistake (a relation). Force = \(-\nabla(\text{relation})\). And meaningful learning happens only along directions with an uneraseable curvature, not along redundant (gauge) directions. Force is the gradient of a relation, and no force wells up along redundant directions ── the whole of "Force That Clicks" showed up, with the same face, even inside the machine that wrote it.

The honest line

"Loss = potential, learning = motion obeying a force" is an accurate correspondence that gets at the essence, but there are simplifications in the details. Real loss landscapes are ultra-high-dimensional and non-convex, incomparably more complex than the one-dimensional valley in the figure. Adam and the like are preconditioned gradient methods that change the step size in each direction, a bit different from plain \(F=ma\). Calling the jitter of SGD a "temperature" is a useful analogy, but it is not necessarily strict thermal equilibrium. "Flat direction = gauge," too, is a loose way of speaking that lumps together redundancy from an exact symmetry (permutation, scaling, etc.) with directions that are merely nearly flat.

The figure is a schematic of damped motion (gradient descent with momentum) rolling down a quadratic valley; it does not reproduce the actual learning of a neural network.

Practice problems (final)
  1. An AI's "learning by gradient descent" is the same as which equation from the main run?
    See the answer
    Episode 6's \(F=-\nabla V\). Regard the loss as the potential V, and learning is motion obeying the force down the slope (−gradient). At the bottom of the valley (minimum loss), force = 0.
  2. Momentum and weight decay each correspond to what in the main run?
    See the answer
    Momentum = inertia (Ep. 1 · into the valley with vigor). Weight decay = a spring's restoring force F=−kx (Ep. 3 · regularization pulling back to the origin). Optimization borrows from mechanics.
  3. A "direction where moving it doesn't change the loss" is the same as what in the main run, and what is the force there?
    See the answer
    Gauge (redundancy of description · Ep. 9/12). It is flat with no slope = force (gradient) zero. Meaningful learning happens only along directions with force = curvature. Same as the rotational invariance of embeddings in sister-series Bonus ⑩.

FINAL SUMMARYLearning was mechanics

An AI's learning was the mechanical motion of rolling down the loss landscape ── loss = potential, learning = \(F=-\nabla V\) (Ep. 6), momentum = inertia (Ep. 1), weight decay = spring (Ep. 3), SGD jitter = temperature. And along redundant directions where moving them doesn't change the loss = gauge (Ep. 9/12), the force is zero, and meaningful learning happens only along directions with curvature. The tools stacked up since Episode 1 were running, whole, inside a machine's learning.

Force is the gradient of a relation, and no force wells up along redundant directions ── the backbone of the main run reached all the way inside the machine that wrote it. Nature (soap films, lattices, molecules) and AI alike compute by obeying a force. Here we close "Force That Clicks" completely. From the feel of pushing, through relation, geometry, symmetry, curvature, and on to a learning machine ── beneath the surface force, the same "relation" was working all along. Thank you for reading.

Force That Clicks · Cosmology That Clicks ── two parts, complete "Cosmology That Clicks" (only ratios are physics; gauge) and "Force That Clicks" (force is a relation; force = −∇V; gauge) entered through separate doors and arrived in the same room ── do not look at the surface values, but at the relation behind them that cannot be erased. If we ever make a sequel, "Force and Mathematics" (the principle of least action, Noether's theorem) would light up what lies even further behind this two-part work. See you again at the bottom of some valley.

This document is the final bonus of the "Force That Clicks" series, "AI and Force," a piece of reading for physics-loving high-school and university students. The following are correspondences widely known in machine learning: that a neural network's learning is gradient descent aimed at a loss function (\(\theta\leftarrow\theta-\eta\nabla L\)); that this corresponds to \(F=-\nabla V\) under the potential \(V=L\); that momentum corresponds to an inertial term and weight decay to quadratic regularization (a harmonic potential = a spring); that the noise of SGD is often analyzed as an effective temperature; and that over-parameterized models have degenerate (symmetry/redundancy) directions that leave the loss unchanged, along which the gradient vanishes. There are simplifications in the details (high-dimensional non-convexity, Adam's preconditioning, the non-equilibrium of "temperature," the exactness of "gauge"). It pairs with the finale of the sister series "Cosmology That Clicks" (Bonus ⑩). The figure is a one-dimensional schematic of damped gradient descent, not a reproduction of actual learning. ── To print, use your browser's "Print" and "Save as PDF" (in the print version the slider and answers are frozen and hidden).

Print / Save as PDF: ⌘+P (Ctrl+P on Windows). On screen, use the slider to watch the AI roll down into the loss valley and learn. Click "See the answer" to open each solution.