Roshi - Logo
Roshi

 - Lessons in 1 Click

GPTeach-3 - A Beginner's Guide to Language Models in Education

By Jonny Kalambay

(@jonnykalambay)

Published on 

Robotic Teacher

By the end of this article, you’ll be able to understand the latest wave of AI tools.

You’ll also understand why, as mind-blowing as they are, they're not replacing teachers anytime soon.

Large Language Models 101

Robotic Teacher

Then feel free to skip this section.

For the rest of you, if you understand the autocomplete on your phone, then you understand large language models (“LLM”). An LLM is basically an overpowered autocomplete. It’s a giant machine learning model that’s trained to predict what comes after a given piece of text.

While your autocomplete can guess how you’re gonna finish a sentence…

iOS automcpleted
message

… LLMs can guess how you’ll write an entire essay.

Auto-generating an
essay

You can do countless things with this, but you first have to understand how to speak the model’s language. This is called prompt engineering.

How to Speak LLM: Prompt Engineering

Language models are dumb geniuses. They don’t have brains, they can’t reason, and they can’t even do basic math. However, given a piece of text, it’s extremely good at guessing what text might follow. Prompt engineering is all about crafting that first piece of text to get exactly what you’re looking for in the second.

LLMs can be pretty straightforward, like in the essay example above. It gets more involved when you want to do things that are complex, like run a chatbot or extract and organize information from some text.

In many cases, it involves writing very detailed instructions and perhaps an example or two in the prompt. Here's an an example from

OpenAI's playground

:

Answering
bot

If you want to see some insane examples of prompt engineering, follow

Riley Goodside

and

Simon Willison

on Twitter, they have language models doing things I would have never imagined, like

describing Better Call Saul using a diagram

.

Diagram of Better Call
Saul

LLM-Based Startups

Technology this powerful opens a floodgate for useful new tools.

Here are some of my favourite examples:

Copy.ai logo

Copy.ai

writes marketing content for you automatically.

Latitude logo

Latitude

enables AI-Powered gaming experiences.

Snapwrite
logo

SnapWrite

lets you generate product descriptions from photos.

What About Education?

What really surprised me was the absence of education-focused LLM-based tech amongst all the amazing tools out there. Some education companies like Duolingo use LLMs in the background to make their lesson materials more quickly. However, there are very few solutions that put the tech in the hands of teachers and learners, to speed up their workflow.

So that’s exactly what I set out to do Roshi. When I launched it a couple of months ago, it was just a Chrome extension that can be used to simplify text. It’s since grown into a platform that does a lot more, like creating and sharing learning exercises. It’s now being used by hundreds of educators in universities, language schools, and even by the Canadian government.

Let me give you an example of ways LLMs can be used to make lesson material:

Here’s how I can make a quick explanation of a frog’s life cycle:

An explanation on the life-scycle of
frogs

If I’m teaching French and want to write some example sentences for a particular word, I can do that instantly:

French
sentences

To get a little more complex, I can take myself back to high school bio class and make a quiz on mitochondria.

A quiz on
mitochondria

Just like that, you can create learning exercises in seconds.

No more need for teachers right?

No, not even close.

Human Educators are Here to Stay.

It's important to remember here that this AI isn't doing any reasoning when it outputs this texts. It's simpliy making guesses based on all of the text that it's been trained on.

To show you what I mean, GPT-3 can easily guess that "1 + 1 = 2" because of how often that specific string of characters has been written its training data.

However, when presented with one that's unfamiliar...

GPT-3 Doing math

11.546 * 2.5 = 28.865

...it just guesses based on similar looking equations it's seen before, interpreting it only as a string a tokens, no different from a song lyric or a restaurant review.

Even if we only focus on language-based information, if anyone here has ever been a teacher, you can see a lot of things that are missing in the above:

GPT-3 Can’t:

  • Understand the learner and adjust the material accordingly
  • Explain context and nuances
  • Give personalized feedback and support
Mr. Feeny from Boy Meets
World

The greatest teacher in history.

What AI can do is support educators by significantly reducing the tedium of preparing lesson materials, to make their jobs easier. Making adjustments to AI-generated lesson materials takes a lot less time than writing those materials from scratch. Also, for people like me who like to self-study, it’s game-changing to be able to make study materials automatically.

This is exactly what I’m building

Roshi

for. It started out as a Chrome extension to simplify text and now lets you do a lot more, like create custom learning activities.

Screenshot of
Roshi.ai

Ultimately, we still need teachers to craft lesson materials to serve learners’ specific needs and, most importantly, support learners’ growth. AI won’t be replacing teachers, but it will improve their workflow, and, in turn, improve education overall.

If you’d like to join me on this journey of bringing large language models into education,

please try Roshi

and share with me your feedback.