.png)
This post is a follow-on from “Hello World. We are NGL” and covers a deeper dive into the learning theory and technologies we are using to integrate AI tutors and copilots into formal learning programs.
Our inspiration began by listening to audio summaries from NotebookLM.
It blew me away.
If you haven’t seen it, you have to try it.
After uploading any document, presentation or link to a website, it generates an audio podcast that helps you quickly understand the subject matter through a casual, editorialised, highly engaging 2-person conversation.
It feels like magic. And it sounds like real people. But, of course, it isn’t either.
It is a very clever implementation of generative AI with the latest voice models from Google’s Deep Mind that no one else has access to.
My first thought was that I had seen the future of education.
But after a few days, I realised that wasn’t quite right. I hadn’t seen the future. This is how learning has always been.
We learn best by listening to others. And all the better when those people know enough about us to make the learning personal and we can speak back to them!
This is the domain of great teaching and 1-1 tutoring and it's the gold standard for any learner who can afford it. We’ve known this since the 1980s (see Bloom’s 2 sigma problem) and while we’ve made some progress in using technology to close the gap between the efficacy of group and 1-1 instruction, a meaningful gap still exists today.
Last September, NotebookLM didn’t show me the future of education. It showed me how ancient learning can be scaled in a digital, AI-enabled world.
This combination of generative AI and voice model technologies will, in time, give us the tools we need to lessen the constraints of scaling personalised teaching.
That is what I saw last September.
And wow, am I excited.
Human-like voice matters a lot here.
We have been learning by speaking and listening for 60,000 years. By contrast, we’ve been writing for 3,000. And so when we hear people speaking about something that matters to us, and in a way that is designed to help us understand and gain insight, something ancient inside of us makes it easy to comprehend the subject matter. And quickly too.
Don’t take my word for it. Just ask the tens of thousands of university students who are using NotebookLM to study for their exams. They upload their course work and lecture slides, ask NotebookLM to help them learn ahead of their exam and then listen to an audio summary that halves their time to comprehension. That’s a lot of extra time for other things students want to do; some of whom may choose to use it to deepen their learning.
There are very smart people at Google’s Deep Mind who developed the voice models I’m talking about. And there are other different but equally smart people on the NotebookLM team who deployed these models in a way that makes any uploaded subject material engaging to listen to and helpful for learning.
This post is not meant to be a promotion for NotebookLM. But they have been the leaders in implementing generative AI in education to date. And their work with NotebookLM, Illuminate and similar tools gives us the clearest view of where gen AI is going in education.
I’ve listened to all the public interviews I can find with the creators of NotebookLM (and listened to some engineers and other insiders too) to learn as much as I can. They speak about the “science of interestingness”. They have a formula for it. And they use it in how they construct their audio overviews.
I see a future where teachers and learning theorists develop a science for effective lecturing, effective group work facilitation, effective 1-1 tutoring, and more. And where, increasingly, teachers delegate responsibility to their AI tutors for defined pieces of teaching with their students.
The key takeaway for me is that we can begin using computers to further develop the science of teaching in a way that was previously impossible. And then we can scale those things at a really low cost.
OpenAI’s RealTime API, Google Gemini’s Live API, ElevenLabs, Mistral and many more emerging technologies are making it possible for companies like NGL and educators around the world to use AI tutors and AI copilots in their teaching. And these technologies are just getting better and better every month.
This matters because the major problem with formal education is not access. The problem is engagement.
The internet solved the access problem long ago. Now, the real challenge is keeping learners engaged, especially in an increasingly noisy digital world.
Engagement is a complex issue. And educators can only take so much responsibility for solving it. But it is my view that these new tools (generative AI + voice models) give us new ways to solve these problems. And this new technology is now a major opportunity (obligation?) for educators around the world.
We’re early and there isn’t a lot of good science yet. But the general view so far is that these tools can be used to hyper personalise the learning experience (achieving 1/2 to to 2/3rds the efficacy of a human 1-1 tutor) and that it doubles the learning outcomes in the same period of time (see Harvard report here).
These are seismic gains and they can’t be ignored.
But this promise of progress doesn’t come without risks. Data privacy and intellectual property issues are two in clear sight. And as we progress, further issues will arise. A responsible, multi-disciplinary approach to this innovation is crucial to its successful implementation so that its positive impact is felt by all stakeholders who adopt it.
We are progressing quickly but carefully with our implementation of AI tutors and copilots into education experiences. And we’ll be writing regular articles to share what we’re learning at this exciting new frontier of learning.
