How the emergence of Large Language Models and eloquent Generative AI interfaces will empower the adoption of the Socratic Method for Learning
AI-generated image created with MidJourney | Owned by The AI Academy as per MidJourney ToS
I have asked around (my kids, my mother, colleagues and friends across multiple continents) and I have not found someone who have not heard about #ChatGPT. Feels like entire generations have been overnight brought up to speed on something that only 18 months ago seemed a very small innovation signal in the eyes of few AI experts (see my October 2021 post on the state of AI).
All it took was to create a friendly User Interface to expose the power of LLMs (Large Language Models) and put it out in the hands of millions of users*.
Open AI’s decision of publishing chatGPT has received a lot of criticism and we would need an entire article to address the consequences of this choice, but I believe most will recognize that the adoption of the underlying technology is a transformational inflection point in human history.
Education is often cited as one of the sectors who will experience a major impact by the widespread adoption of this technology. I’m fortunate enough to have focused on AI and Learning well before chatGPT created this enormous curiosity wave and would like to share where I see the biggest potential of the technology behind it when applied to Learning.
Is GPT good or bad for Education?
We know that binary models of the world like this good/bad, black/white approach are attractive (popular) because are simple but not very good at explaining the complexity of most human systems.
The truth is that this is the wrong question to ask. The truth is that the positive and negative consequences of most technological advance are dependent on how humans choose to adopt it and implement it into our lives: we can choose to deploy nuclear fission to create extremely clean energy sources or use it to create nuclear bombs that will destroy us but there is nothing intrinsically good or bad in the discover of nuclear fission.
As a father of two teenagers, I confess that back in December 2022 I lived through that “holy s**t” moment most parents and educators have gone through when you realize how good the text produced by chatGPT is on almost every topic. And how easily accessible is as a tool.
Some reacted to the release of chatGPT with their control instinct (NYC education department blocks chatGPT in school devices and networks, Italy became the first Western country to ban ChatGPT, ChatGPT Is the Wake-Up Call Schools Need to Limit Tech in Classrooms, When Chat GPT writes the homework: how should the school react?). Some have realized that control is a losing strategy and started looking at ways to embrace the technology advancement to tip the balance to a net-positive situation (Don’t Ban ChatGPT in Schools. Teach With It., ChatGPT is going to change education, not destroy it). Dr Philippa Hardman’s is one of the most eloquent promoters in the net-positive camp and shows with her EDUCATE framework a path to a possible post-AI education scenario.
What I would like to show in the rest of this article is that, beyond the positive and negative hype we are bombarded with on a daily (hourly) basis, there are responsible ways to deploy Generative AI and LLMs (Large Language Models) that will lead to a significant improvement in the way we humans learn.
Beyond chatGPT: “specialized LLMs”
AI-generated image created with MidJourney and edited by The AI Academy | Owned by The AI Academy as per MidJourney ToS
In a way chatGPT has “packaging problem”. The design choices OpenAI has done are phenomenal, but they induce most non-technical people to view the chat interface as an Oracle: you ask questions and magically get eloquent answers on almost any topic. The feeling we get is that you are interacting with an “intelligent creature”: “he said”, “she replied”, are the ways the majority of people choose to describe their experience with the bot. Which by the way is a great testimony of how much we associate language skills with intelligence (“my 2 years old is so clever: she’s already putting together 5 words sentences”).
Indeed, the sheer amount of data these Large Language Models have been trained on give them the ability to provide accurate responses on a very large number of topics. At the same time, we know that even the latest implementations (GPT4) are not always a reliable source of information: the term everybody has adopted for these situations is "hallucination” which is a fancy way of saying “these models sometimes make stuff up”.
I’m sure the attention and funding the teams building these models receives will allow to quickly make improvements but the good news is that we have already today the ability to build “specialized versions” of these LLMs (Large Language Models) that allow us to query domain-specific knowledge bases. These specialized versions bring three main benefits:
reduce the probability of the model to hallucinate by being less general purpose.
extend the knowledge domain beyond the knowledge acquired during the model training.
can be “private” so that specialized knowledge doesn’t have to be shared with platform owners or any 3rd party.
Without going into the technical details of how this is done, the key message here is that today is fairly straightforward to build a private Knowledge Base and integrate it with foundational LLMs (proprietary and/or open source) to leverage their powerful Natural Language skills and create search or chat interfaces to access your private Knowledge Base.
Perhaps is time for some examples to understand the idea better.
Example 1
Say you company is a service provider (for example a banking or internet service provider) and you have a large number of documents describing all the policies about the service is offering. You can easily build a Knowledge Base that contains all these policies, integrate it with an LLM and make this knowledge accessible to your customers via a chat interface: they can just ask questions as they would ask to a person and the chat interface will provide the answer to that question using natural language (in any language).
If you are wondering how this is different than what service providers do today via chat interfaces the quick answer is that this actually works: customer service chat interfaces today provide a notoriously poor customer experience and the power of specialized LLM can transform this.
Example 2
Let’s see an example closer to Learning. Say you are a Neuroscience student and your school has created a digital Knowledge Base from 10,000 Neuroscience books and papers. Your school is integrating it with an LLM and providing a chat interface you can start exploring this vast knowledge territory by asking questions.
Perhaps you may start mapping this territory asking what the main branches of Neuroscience are. From there you could dive into the one that interest you the most asking who the most prominent scientists in that particular branch are or what is the timeline of the major discoveries in the field.
You can go on and on to discover facts and concepts with your curiosity as the only limiting factor. Through this process you can potentially reach the boundaries of the existing knowledge on this topic achieving that exciting wisdom level were the only thing you know is that there is a lot we do not know. And you can start charting your plan to explore unexplored knowledge territories from there.
SIDE NOTE: Building a demo of a specialized LLM is quite easy these days: a quick search on YouTube will give you a ton of videos of how to do that with few lines of code. Creating a robust and scalable enterprise application for this is a different ball game but our team has already gained experience on how to do this. If you are interested in creating a specialized LLM for your company, please contact us.
What has Socrates to do with this?
Socrates believed that knowledge is not something that can be imparted from one person to another, but rather something that must be discovered through one's own efforts.
This is very different from the most popular approaches to learning where the student is a passive actor, an “empty bucket” that need to be “filled” with knowledge by consuming information. This could be read 10,000 Neuroscience books and papers in the example above and memorize most of it or access an online learning platform and watch 1,000 hours of recorded lessons.
There are two sides in this “Learning Architecture”:
The Facilitator
The Learner
AI and LLMs can be powerful tools to for both aspects of this new learning experience.
The Facilitator
Depending on the learning context the Facilitator might be a Teacher, a Tutor or a Mentor. It can also be simply a Peer in your class or your company that is exploring the same knowledge territory and has discovered some key insight that is willing to share. In any case it is a role that interact with the Learner with one main objective:
support the Learner in the discovery of the truth (knowledge)
Socrates saw the role of the Teacher not as someone who imparts knowledge, but rather as someone who helps the student to discover knowledge for themselves. In this sense we can extend this role to the Tutor or the Mentor or a Peer depending on the context.
Through his method of questioning, Socrates believed that he could help others to uncover the truth by encouraging them to question their assumptions and beliefs. He also believed that by examining their beliefs and considering alternative viewpoints, individuals could arrive at a deeper understanding of the truth.
Using AI is possible to create a digital Facilitator that monitors the quality of the learning experience at scale providing nudges and probing questions or recommend personalized content to help Learners acquire new skills.
The Learner
Perhaps the most powerful aspect of the Socratic method is the fact that turns learners into “active” actors promoting critical thinking and intellectual development. This has direct positive impact both in the Learner’s engagement and the Learner’s proficiency level - two very sought-after benefits in both educational and corporate learning contexts.
Learners can apply the Socratic method to learn a new skill by using a structured approach that involves asking and answering questions. Here are some steps that Learners can follow to apply the Socratic method to their learning:
Identify the topic: Choose a specific topic or concept that you want to learn more about. This could be a topic that you are studying in class, a subject that you are interested in exploring on your own or something you need during your work.
Generate questions: Think of questions that you have about the topic. These questions should be open-ended and thought-provoking and should encourage you to think deeply about the subject.
Research and explore: Use your questions as a guide to research and explore the topic. Look for information in books, articles, online resources, and other sources that can help you to answer your questions.
Evaluate the information: As you gather information, evaluate it critically. Consider the sources of the information, the credibility of the authors, and any biases or limitations that may be present.
Ask more questions: As you explore the topic and gather information, continue to ask questions that challenge your assumptions and encourage deeper thinking. Use your previous questions and answers to guide your thinking and generate new questions.
Reflect on your learning: After exploring the topic, take time to reflect on what you have learned. Ask reflective questions that encourage self-assessment and critical thinking, such as "What did I learn from this experience?" or "How can I apply what I learned to other areas of my life?"
By following these steps, Learners can use the Socratic method to learn a new skill in a structured, thoughtful, and reflective way.
Learning through questions and answers, exploring a specific or specialized Knowledge Base is exactly what LLMs (Large Language Models) and GenAI are very good at, and it is now possible to create entirely new learning experiences using these technologies.
Early examples of the use of GenAI and LLMs for Learning
Education
Perhaps the most notable example of how companies are already using Generative AI technology to transform learning in a positive way is Khan Academy. The application of this technology is so impactful that OpenAI has given Sal Khan’s team early access to their GPT 4 models 6 months prior to the chatGPT announcement.
You should definitely watch this 15-minute TED video where Sal Khan explain how they have used GPT4, the latest and more powerful LLM released in March 2023 by Open AI, to create a huge impact in education. The two main use cases presented are:
AI Tutor - a bot that interact with each user at global scale to empower their learning experience.
AI Assistant - a way to make teachers and educators more efficient accelerating several tasks required in their day-to-day activities.
It gives me great satisfaction to hear Sal using the word Socratic several times in this video, further validating the observations I share in this article.
Corporate Learning
We at The AI Academy believe so much in the transformative power of GenAI and LLM that we have been fiercely working on a Workplace Learning Platform that will bring the benefits of these technologies to corporate learning.
We call our platform The Village and we believe it will help make learning in the flow of work more engaging, measurable and effective bringing real benefits both to the company and to the professionals.
The main difference with the Khan Academy is that we have tailored the learning experience to a Corporate Learning environment and instead of using the general knowledge available in the pre-trained OpenAI model, we can use “specialized LLMs” to contextualize the learning to the specific context of each learning community. This allows internal teams to share and access internal knowledge with the appropriate security and privacy levels required in any corporate environments while using the power of GenAI to discover information and insights through the questions and answers paradigm.
Closing note
I hope I managed to present my case on why I believe GPT and LLMs will have a positive impact on Learning and how using these technologies it is possible to create more engaging and effective learning experience following the Socratic method. I would love to hear your comments and feedback on it.
*I’m of course simplifying here, and we should all recognize that a lot of work has gone into the evolution from the 2021’s to the 2023’s state-of-the-art LLMs not just in terms of performance but also in safety.
Comments