A personal assistant for all

Happy Friday, everyone.

This morning, I was helping to deal with an incident of scratching in a Children’s House classroom. I was running short on time, so I asked a colleague for help in drafting a quick email to support the classroom teacher. The email draft is worth analyzing together.

First, here’s what I told my colleague:

I need to send two emails to two sets of parents in a Montessori classroom. One child, Alice, scratched another child, Sally. It left a mark on Sally, which I think her parents may be sensitive to. This was not typical behavior for Alice. Both children are 3. I need to protect the privacy each of the children to the other set of parents. Can you draft both emails please?

Here’s what my colleague sent back for the first email:

Subject: Incident Report: Classroom Update

Dear [Sally's Parents],

I hope this email finds you well. I wanted to inform you about an incident that occurred in our classroom today. During playtime, another child accidentally scratched Sally, which resulted in a visible mark. Please know that we promptly attended to Sally, ensuring her well-being and comfort.

I would like to assure you that this type of behavior is not typical in our classroom, and we take such incidents seriously. We have spoken with the other child and their parents to address the situation and ensure that appropriate steps are taken. Please note that we are committed to maintaining the privacy of all children involved, and we request the same from you.

In our Montessori environment, we strive to teach our students the importance of respect, empathy, and communication. We will use this opportunity to reinforce these values with our class and help the children understand how to interact with each other in a positive manner.

If you have any concerns or questions, please feel free to reach out to me. I am always available to discuss any matter that pertains to your child's well-being and education.

Warm regards,

The second email, to Alice’s parents, was similar overall, but with a key paragraph changed:

We have addressed the situation with Alice, emphasizing the importance of gentle touch and respecting personal boundaries. We are confident that this is an isolated incident, and we will continue to support Alice in understanding and practicing appropriate behavior in the classroom.

Before I point out what was interesting about these emails, it’s worth forming your own opinion: What do you think of them? How good are they? Is there anything that you might add or modify? Is there any special context that you might want to know about so you can highlight or add something in the emails?

Take a minute.

Okay. What I think was interesting—and I’m sure some of you have guessed this by now—is actually that this exercise is a bit of a bait and switch. Sorry. My “colleague” was an artificial intelligence model. The emails were written by a computer. Specifically, by OpenAI’s GPT4 model.

The incident is based roughly on something that my daughter Alice (unfortunately) actually did to one of her classmates last week. (The need for support in this case is fictitious; our guides communicated to my wife and me about the incident wonderfully without any need for support.) But the emails themselves were generated by a robot.

Over the last 10 or so years, the field of machine learning has made startling advances in technology that can replicate flexible, intelligent human behavior. Over the past year or two, these models have been trained on enormous data sets—namely, the entirety of the internet. And over the past several months, interfaces to these models have become eminently usable. Using an AI language model is now precisely as difficult as sending a text message. Seriously—go try it if you haven’t already. And don’t limit your asks, you may very well be surprised.

The way these models work is profoundly alien. They are using complex statistical language patterns (which they've picked up from all of the language content on the internet) to generate a reasonable response, word by word. Despite this inhumanity, interacting with these models feels very human. Indeed, the best way to interact with them is naturally, as you would with a human: have a back-and-forth. Ask for an answer then ask questions about the answer. Or ask for whatever revisions you want to make to the email above. You can ask the robot for more work, clarification, or additional drafts, all in ordinary language.

I could tell it to redo the email to Sally’s parents, keeping in mind the (fictitious) context that she’s had several incident reports already, that her parents are concerned that the campus isn’t safe, and that I’m really worried about losing this family’s trust. I in fact just tried this, and it did a pretty good job. (My wife Gena was using AI to help her fill out an academic Call for Papers, and when she didn’t like the first draft, I told the AI to “try again as though you actually have a personality”. Gena was horrified at my manners but the AI actually did better on the next try.)

The world of technology is full of empty hype. But sometimes a genuinely new frontier opens up, one that will effect massive change, one that benefits virtually everyone to learn and use extensively. The internet itself, especially the modern, smartphone-accessible internet, is one such change. It is extremely likely that AI will be another.

ChatGPT is the fastest-growing consumer app in history, faster than TikTok, and for good reasons (again, go try it). Software developers have been using AI support for programming for some time now, asking in natural language for bits of code and help with debugging. (Presently AI is even better at programming than it is at ordinary conversational language.) Images generated with AI tools like Midjourney, Dall-E, and Stable Diffusion are already ubiquitous. AI video generation is well on its way. There are many, many people who are working very quickly to create AI tools that help people do everything from clean and analyze financial data to assist with medical diagnoses to describe video content. In the very near future, you’ll be able to ask a robot to shop for online groceries for you or book flights for you. There was a 60 Minutes segment on AI earlier this week.

People are already talking. But soon this technology will dominate the cultural discourse. It will be the only thing people are talking about. The capacity to greatly extend the reach of what computers can do brings with it the specter of major social change. There are very real risks but also enormous benefits.

We are, as an organization, exploring ways to integrate AI into our work and programming. Right now, that mostly means using it to support adults in various ways: making it possible, for example, for a parent or educator to talk to our Montessori curriculum and ask questions of it. We’ve been exploring for some time having AI analyze classroom footage and automatically producing reports analogous to Transparent Classroom reports and Montessori’s concentration curves.

In lieu of specific tools, there’s already so, so much that you can do with the general tools. You can produce first drafts or even second and third drafts of emails. You can edit emails—you can just paste an email into it and ask the model to clean up the grammar and typos. (You can even ask it to produce a report for exactly what it changed, so you can review it.) You can have it summarize lengthy material for you. For clerical writing tasks, which are innumerable in our work, AI is a very competent personal assistant who has infinite patience and works for free or, at most, $20/month, if you want to use the very latest language model.

You can use it as a brainstorming partner—describe your problem and situation and ask for suggestions or resources. If you know even a little bit of programming, AI will give you the power of divine fire, generating functions, scripts, or even entire small apps at your request and to your specifications. AI is very good as a learning tool, already better than Google in many cases. Over the last few months I’ve used it to learn how to make pancakes fluffier, the difference between mortgages and construction loans, how to use version control software, and to get leads on finding specific quotes that I’m having trouble with. As a tutor for adults, AI gets a lot right but a lot wrong—beware that it will make very confident mistakes, called “hallucinations”. I’ve asked it many questions about Montessori, and it gets surprisingly specialized questions correct—it knows that the Pink Tower comes before the Brown Stair, and why—but also produces a significant number of errors. Still, despite being prone to hallucination, it has virtues that are hard to come by in human tutors. Two big ones: the aforementioned “infinite patience”, and the content base of “has been trained on the entire internet”.

Getting back to the original email draft, I think what the model generated is quite good but not stellar. It’s a solid B. If I were actually sending it, I’d want to add more personal touches and adjust for individual parent needs. I’d probably cut the stuff about privacy. Some of these shortcomings AI could help with, with another round or two of editing, and others would just be easier to do myself. But I would still absolutely prefer to have this draft as a starting point. It’s much easier, for me at least, to start from a perfectly structured, grammatically correct, professional email, and add a bit of context or remove a bit of unnecessary hedging where I need to.

My primary goal in this note is to encourage you to try to use AI models if you haven’t already. The main thing to use is ChatGPT, which will require you to create an account. Use it instead of a Google search. Use it to try to do work for you. Use it for fun—just mess around. It’s not a perfect tool, but it’s already valuable, and it’s worth building the habit now.

There’s one more thing worth mentioning, which I’m sure will merit its own note in the near future: child-facing AI products are coming. They are coming fast, in weeks, or, at the latest, months. The technological capacity already exists for a child as young as three to converse with an AI model, in ways that make talking to Siri or Alexa or Google Home speaker devices look like a primitive joke. But the default AI models are likely to go off the rails in terms of giving responses that aren’t child-appropriate, usually in a benign way but very occasionally in an alarming way. These issues are being solved, and when they are, there will be a new normal: everything from AI-empowered teddy bears to very young children independently interfacing with the internet.

There’s a lot I haven’t discussed. I haven’t discussed how this technology relates to Montessori in particular (it very much does), nor how it will threaten traditional modes of education (ChatGPT cheating is already rampant; AI can be used to write essays and pass sophisticated, specialized exams). My slant here is deliberate, because my main goal is just to get you intrigued and hopefully enough excited to try it out. A genuinely powerful, flexible new technology has been born, and, in the words of Montessori:

It is not always imperative to see big things, but it is of paramount importance to see the beginnings of things.

Matt Bateman
Executive Director, Montessorium