The chapter shows three facts that pull in different directions. Loss falls along a power law as parameters, data and compute grow. Hidden vectors carry structure that a probe or a sparse autoencoder can read. An explanation can be plausible to a person and still fail when the model is intervened on. The field has no agreed answer to whether larger models are easier or harder to explain and align.
Take a position on this question: as models grow, does interpretability and alignment get easier, harder, or neither? Design a measurement that would support your position, say what result would change your mind, and name one thing your measurement cannot establish.
Rubric. Tick each criterion your answer meets, then save the self-assessment. The platform does not grade it.
Worked solution
There is no single correct answer, so this is a model of the shape a complete answer takes, not a key. A level-3 answer begins by fixing the comparison: two model sizes trained on the same data with the same tokeniser, evaluated on the same held-out tasks, with the same budget for the explanation method. It then names a quantity, for example the fraction of a refusal behaviour that a fixed set of directions can turn off and turn back on, or the fraction of a probe's accuracy that survives a change of random seed. It states a prediction: if scale makes alignment easier, the intervention success rate should rise with model size at fixed explanation budget; if scale makes it harder, the rate should fall or the number of features needed should grow faster than the model. It says what would change its mind: a reversal at a larger size, or a result that disappears when the tokeniser or the training mix changes. It names a limit: a direction that controls refusal on the prompts tested may do nothing on prompts nobody wrote, and a feature that a sparse autoencoder reads may not be the unit the model uses. The answer also separates the two questions in the chapter. Interpretability asks whether a person can read the computation; alignment asks whether the model pursues the intended outcome. A model can become more readable and less aligned, or the reverse. A level-3 answer does not claim that one curve settles both.