Kirsh argued that we are not just passively reacting to computer-generated options. If we look at interaction at a higher level, beyond a single action, we see that users are also actively influencing their environments. Users are "architects" of their environments, as Kirsh put it. For example, users may change the settings to turn on or off a function or change the way it behaves. They also choose the applications they use. Such tailoring behaviors are not explained by Norman's intention–action–response–interpretation–evaluation cycle.

Kirsh points out that Norman's model makes an unrealistic assumption: The user is assumed to know the environment and its options and is merely picking an option. In practice, we do not always know what the options mean or even what options are available. Kirsh argued that users need to actively explore interfaces to become aware of the available functions and how they work. Via exploration, they also learn about their own abilities in using them. Consider the first time you launch an application; you probably try out various actions to see what happens. Kirsh argued that the discoverability of such options is as important as their visibility; however, discoverability is not covered well by Norman's theory.

The modelling subscribes to a linear account of the cognitive mechanism, going from goals to actions and back. However, according to current understanding in cognitive sciences, the picture is more complicated. One thing that is missing is an account of how beliefs about the computer are formed and updated and how they drive action specification. The current understanding is that users form internal models that predict how their actions produce perceived outputs, and they learn to minimize prediction errors. This explains why people explore interfaces (to develop better internal models) and why, eventually, they no longer need to compare outcomes against goals. Moreover, the model was initially used in a weak, heuristic sense and did not converge with efforts to implement interactive systems.

However, self-attention alone is permutation-invariant, i.e., if we reorder the rows of X, then the mechanism has no built-in sense of which token came first. Since word order matters, we must inject positional information. We often add a position vector pt to the token embedding: h(0)_t = e(xt) + pt One classical choice for the positional encoding is called the sinusoidal positional encoding. pt[2k] = sin(t / 10000^{2k/d}), pt[2k+1] = cos(t / 10000^{2k/d}) The sinusoidal features give each position a distinct geometric signature across many frequencies. Nearby positions have related encodings while distant positions remain distinguishable. This lets the network reason about relative offsets.