Education Technology - Condensed milk version

Education raises the floor for all civilization (but also enables those who would damage it in search of treasure). The most efficient way to help everyone is by systemic improvement, such as researching optimal protocols and building technology for education.

As informational beings we use understanding to make the things that happen in the universe around us increasingly preferential. As a tutor I motivate students to take interest in their own learning. I also support them to learn how to learn.

1. Identify concept in question, relevant knowledge components (KCs)

2. Infer/Assess interest and familiarity with proximal KCs

3. Identify missing/incorrect concepts. Shine (gradually bright) light on it if unnoticed. Disentangle the faulty connections with existing knowledge.

4. Provide just enough scaffolding as needed, titrating information until progress is made.

5. Re-inforcement / competence building tasks (hypothesised Learning Events, in spaced repetition protocols)

6. Assess competence and goto 3 or 4 as needed (infer LE efficacy, update ZPD and learner KC tree)

Applying knowledge of knowledge to its acquisition for learning

As other thought leaders in ed-tech have posited, a future where you’re holed in to some device to learn seems dystopian. Also, to be in alignment with the definition above (regarding utility of learning in the world) we implicitly require measurements and demonstrations of competence to be ‘in-situ’ as much as possible, using toy models if not the real thing. It may be useful to consider examples of subject matter here - mathematics on one hand which is perhaps the subject with the most objectivity and abstraction, amenable to interactive visualizations we can code and customize, such that the education can live almost entirely within technology. On the other hand, we may have something highly hands-on like carpentry or driving, where most of the education happens ‘out there’ in the real world (until we can simulate it well enough, like the virtual labs for medicine etc that the pandemic made us make).

The causal chain that we can use to improve is: Changes in instruction -> changes in learning -> changes in knowledge -> changes in robust learning measures. For the academic robustness, we can look to learning science stuff, like the KLI framework, which is built on that same causal chain.

The biological technology those steps are dancing with

Every one of those steps has to work with hardware we did not design and cannot patch: human memory. Its constraints are predictable enough to design around, and four of them matter most here.

Memory forgets on a readable schedule. The gap that best fixes something in place is not fixed; it scales with how long you want to keep the thing, shrinking as a fraction of that horizon as the horizon grows (Cepeda and colleagues: nearer a fifth to a third of the way out for a week, a twentieth to a tenth for a year). Cramming loses the memory; reviewing it after roughly the right gap helps it set. Step 5 applies that schedule by feel.

We think through a small, brief window while long-term memory stays effectively bottomless, and everything has to pass through the window first (Sweller’s cognitive load theory). The material’s own difficulty you can only sequence around; the difficulty you add by explaining badly is pure tax, and the more expensive of the two (watching people work with GPT-4o, Lepine and colleagues found it dragged on the output about three times as hard as the intrinsic kind). The scaffolding that saves a novice becomes clutter to an expert. That is why step 4 has to move with the learner: add support when working memory is overloaded, then fade it as the material becomes familiar.

Blocked practice looks productive because it keeps a learner succeeding, but it predicts poor retention. Mixing the problem types forces the learner to keep deciding which method applies; it feels worse, lasts longer, and carries over to cases they have not seen (Schorn and Knowlton; Rohrer and Taylor). A dip in today’s accuracy can come from the kind of practice that will last, which is an awkward thing to put on a dashboard.

And knowledge is not one substance. A fact you need fluent, a rule you must induce, and an idea you must understand each want different treatment, which is why the literature seems to argue with itself about testing versus worked examples (the KLI framework’s answer: they were studying different kinds of knowledge component). Decompose the domain into its components, ask what each one needs, then choose the method. Steps 1 to 6 already separate those components before choosing what to do next.

Software can support these habits by noticing which one is in play and responding the way the pencil-and-paper tutor already does.