Old assumptions about learning relied on a simple signal. If the learner could produce the answer on demand, the understanding was in there somewhere, and the path from “produced an answer” to “understood the material” was short enough to be implicit in most teaching.
Since LLMs have entered the learning loop, the learner can now produce the answer by reaching for the model. From the outside, the output looks the same, which means that the bottleneck moves from what the learner produces to whether the capability actually transferred.
This shift does more than break the signal: it changes the weights between two underlying skills that have always mattered for lasting capability.
Compounding and agency
The first skill is that what a learner already knows shapes how fast the next thing sticks, and over years that compounding is what builds real capability. The second skill is agency, the choice between reaching for a model and doing the work yourself - a choice that now arrives every time the learner has a gap.
With a model in the learning loop, a learner who reaches for the model whenever a gap appears never has to find out whether their understanding holds.
They can move through the material producing answers without ever testing whether the knowledge is theirs or whether the capability has transferred.
What I’ve found is that real capability is the residue of effort that the learner did themselves. Anything that bypasses the effort also bypasses the capability. However, this is not a new observation - it is roughly what spaced repetition and active recall have been about for decades. What is new is that the easiest way to bypass effort is now sitting in a tab, ready to produce whatever the learner asks for.
Active recall, partial answer
Active recall is one design response to this. It forces retrieval in a context where reaching for the model defeats the purpose. The classic flashcard format is the simplest version: prompt on one side, answer on the other, retrieval as the act of remembering rather than looking up. When the alternative to retrieving from memory is retrieving from a model, the cost of the easier path is no longer obvious to the learner. They are not aware they are bypassing the work that makes the work stick.
Active recall is also a first approximation, not the full answer. Recall at the flashcard level is not the same as reasoning through a new situation. Most real-world capability revolves around the ability to take what you know and use it in a context that does not look like the practice context. Flashcards train the retrieval muscle, but do not train the transfer muscle directly.
The harder problem
I have built Brainbank around this distinction. It handles retrieval reasonably well: spaced repetition forces internalisation, and the capability is meant to live in the user, not in the tool. What it does not solve is the harder question of whether that capability actually transferred.
This might be where agents change something - probing what a learner actually understands, not just what they produce, might get closer to measuring transfer than anything we have now. That shifts the measurement problem from “what did the learner produce” to “what does the learner understand”, and lets learning systems build agency alongside compounding.
Not sure yet what this looks like in practice, but when a model can produce an answer to almost anything a learner might ask, the lasting capability comes from agency not first reach. The tools that will actually push education forward are the ones that build both.