Contents of this post:
My wife has recently started learning Spanish in preparation for a trip to Mexico and Colombia. She uses a mix of Duolingo and Zoom conversations with a native speaker through iTalki. Duolingo is good for gamification, and it helps her practice phrase translation and vocabulary training every day. The conversations via iTalki are excellent for actually talking to other humans. However, between the two, she recently mentioned that she was missing something - an interaction a bit like conversations but with a stronger emphasis on correcting her mistakes. Her conversation partner does provide some corrections, but they can feel unnatural in a conversation and thus don’t occur frequently.
I asked her if she would like to try using ChatGPT as a teacher and she was up for it. I prompted ChatGPT with the following prompt:
„Please assume the role of a native Spanish speaker conversation partner. You will ask me questions about my trip through Colombia. If I make a mistake or if my reply is not worded as a native speaker would say this then please gently tell me what a native speaker would say instead before asking the next follow up question.“
ChatGPT readily complied and my wife was sold after two paragraphs back and forth.
Having an infinitely patient personal tutor that assumes the role you need is a really powerful concept. I have used variations of this for my work. It is fascinating to see how using GPT makes a much wider area of your profession available to you. I am fluent in five programming languages and can somewhat work in maybe ten more. There are many other interesting or useful programming languages that might be a slightly better fit for a problem, but the effort of learning a new language is often not worth the benefit. This calculation changes with GPT. All of a sudden, writing small programs in a language that I am not at all familiar with goes from “very tedious” to “not as fast” as one of the languages I know well. This doesn’t scale to big programs because understanding things at scale usually requires a deep understanding of the language, but quite often being able to write a simple tool in the optimal language is very valuable. Simon Willison has written a great post on a similar observation.
“Girl with a Mandolin” by Pablo Picasso
I also think that having a digital assistant at hand that can write simple pieces of code could unleash a new golden era of automation. It used to be the case that if you wanted something specific to happen automatically in your workflow, you either had to know how to code and be willing to put in the time to learn the API of the relevant tools, or you had to be lucky and the vendor of the tool had to already support this kind of automation in the UI. With GPT, it is now vastly more feasible to write little scripts, etc., without knowing how to program. This kind of use hits a complexity ceiling quickly, of course, but the fact that all of a sudden, the number of people who can reasonably attempt automating stuff goes from “all software developers” to roughly “all humans” is pretty wild. Hillel Wayne wrote about this possibility a while ago in his excellent newsletter.
Coming back to the idea of your own, personal non-human tutor - I wonder if some people might prefer opening up to a machine that they know will not judge them. A lot of people have substantial areas of middle or high school math that they never truly understood but were afraid to admit not truly understanding. If I can just ask my infinitely patient, non-judgemental personal teacher to walk me through trigonometry again, it might just click, finally.
Reflecting on all of this and trying to put it into a bigger context, I wonder about the added benefit that something like a personal tutor could have for pupils, especially in poor countries with poor educational outcomes. Of course, there are many problems in impoverished countries that will not be helped by adding GPT - if kids don’t spend much time in school, a personal tutor will not help. Similarly, GPT is probably not a great solution for teaching kids to read and write in the first place. However, once kids get far enough along in school, I assume that a lot of potential brilliance is wasted due to individual students not receiving enough attention or teachers not being very well-educated themselves.
I can definitely imagine that with a bit of work, a specially crafted solution that uses something like GPT could be very useful in helping kids achieve higher educational outcomes. Such a system would have to be prompted with the goal of the lesson and be able to keep some notes on students’ weaknesses - but then it could explain and question the student at whichever pace the student needs. Such a system could be pretty inexpensive to run, once you manage to have a simple internet-connected device available for each student (and with smart-phones becoming cheaper all the time, this might not be a significant hurdle).
The scale of the challenge of educating kids better in poor countries is certainly big enough to warrant more attention. I just had a look at Nigeria as an example. It is home to 220 million people and projected to grow to 380 million by 2050. The median age is 17 years old. About 200 million children will need to go through school in the next 3 decades in Nigeria alone. Wouldn’t it be great if this new technology could be used to help many of them realize their potential?
“Panel 58 - In the North the Negro had better educational facilities” by Jacob Lawrence
I am pretty excited about what LLMs will do for our ability to learn things more efficiently. If you haven’t used ChatGPT yet to learn something or write a little script for yourself then I suggest giving it a try. You might have to experiment a bit with a good start prompt and you may have to remind it again in a longer chat to remember its role and instructions. If you run into an issue with coding just copy paste the error message into the conversation with GPT. Give it a spin - the outcomes can be quite remarkable.
(Note: this post was published also on my Substack)