Understanding Is More Efficient Because It's More Robust – and More Robust Because It's More Efficient

Understanding is more efficient because it's more robust – and more robust because it's more efficient.

When I ran AI trainings, I always placed a lot of value on giving learners not just facts, but understanding.

So at the start of each training, I'd invest about an hour and a half explaining how large language models (LLMs – ChatGPT and the like) actually work: what happens under the hood when I type a question and get an answer. Even though I simply find it fascinating to demystify something so “magical”, that investment was never about finding or creating like-minded people. Instead, understanding the technology was the foundation that made every practical tip easier to grasp.

All of these are insights from trainings in 2023/24 – and even though roughly 43 years have passed in AI time since then, these insights are all still relevant. What was understood was how answers come about – and that process has barely changed despite all the development.

But if you've understood how to use the models' intelligence by proactively enriching your own context to prepare a better answer, you realize: the game plan is really just a bigger chain of thought – a logical consequence of rising autonomy. A continuum at the end of which, in two years, you'll probably be holding a weekly with your agent team.

A wide variety of concrete use cases can be derived from the same conceptual understanding – which makes what was learned more robust.

But is robustness also important for … ? Yes.

Take a company running cybersecurity trainings, for example. You can invest tens of thousands so employees learn what a phishing email looks like right now. Purely didactically, that's not wrong – except the scam signals will keep getting subtler over the coming years, until bad actors build pitch-perfect websites with supposedly genuine team photos, videos of supposedly genuine people, and of course a download of a (trial) version of their supposedly genuine terms and conditions – which then contains a prompt injection for your local agent.

But if the fundamentals of social engineering, attack vectors, and the like were actually understood, that knowledge still holds even as the surface form changes. You just need to tighten a few conceptual screws, rather than throw out everything you learned before.

The same goes for data protection, leadership skills, and so on. A stronger focus on conceptual understanding, through its flexibility, protects not only employees from automation, but also employers from bad investments in perishable e-learning – a risk that has never been greater than now.

Understanding is therefore economically more efficient, because it makes expertise more robustly relevant. You save, because what's understood stays relevant longer.

Once you've understood how multiplication works, that understanding gets activated every time you multiply two numbers – there's an 'if … then …' link. If, instead, you only know the times tables by rote, there are 55 separate 'if … then …' combinations. For each individual one of those number combinations, it's far less likely to be recalled than the general rule is. The algorithm – just multiply – is a single gateway to enormously many solutions. Like a busy highway in your head, instead of 55 footpaths.

The same applies to my AI workshop from earlier: once you've learned how a word gets predicted, that's a gateway to enormously many conclusions about meaningful interaction – whereas individual if-then rules remain isolated footpaths. And with many footpaths, some inevitably grow over (what was 9 times 6 again?).

I like to sum this up as: good learning is good compression. Trying to accumulate a lot of knowledge is often futile if that knowledge can only be reached via badly paved roads. That's exactly why things like the famous 'activating prior knowledge' and analogies are so valuable – they turn isolated forest paths into a connected network: few gateways to a lot of knowledge. And roads that are regularly traveled like that don't grow over either.

Understanding is therefore cognitively more robust, because it maps content more efficiently. And if you additionally manage to make that one gateway to the efficiently mapped content 'catchy', you have a good chance that something will actually stick.

I can't compress it any further than that.

About the author

Dr. Christoph Kenntemich is Co-Founder of TreadEd.

TreadEd is a creator tool that makes it possible to design dialogue-based learning paths in which learners are always optimally supported through one-on-one guidance by an AI tutor. Experienced employees fully shape the tutor didactically, passing on their knowledge and experience at scale.