AI is incredibly powerful as a learning tool, but almost too powerful
Since the creation of ClipNotes, the technology landscape has been dramatically changing with the arrival of widely available artificial intelligences in the form of ChatGPT and similar large language models. They have applications in so many areas, and the tech industry is in peak AI hype, with many people building new tools and software that leverages the incredible power of LLMs and AI to simplify and streamline many things we do today. As a person living in San Francisco and working in the tech industry, it's an exciting time! Given these major changes, I wanted to take the time to explore the impact of LLMs on the learning space and explore their capabilities, current applications, how they might change the landscape of learning in the future, and how to incorporate the technology into ClipNotes.
Here is my simplified explanation of how large language models work and how they create the semblance of "artificial intelligence". A large language model is capable of creating connections between words. Given a large enough set of data, the model is then capable of recognizing patterns and making connections between words and phrases, associating them in different ways to build a very complex association network. Humans are also involved in helping the model determine if an association is correct or incorrect. For example, the word "sky" might be associated with "blue" or "atmosphere" or "rain", or maybe even "heaven". When a person goes to use ChatGPT and ask it, "What color is the sky?", the LLM can recognize the pattern of the words in the question, recognize the use of a question mark, examine the associations of each word in the question, how often each association occurs, and will generate a response by creating a string of words that have a high probability of making sense given the context.
The key point here is that these AI's do not understand the words that they output, but are instead output words in a patterns that it thinks are correct. This means that the data that you give a language model is very important, along with the human guidance that a model receives. In my example, if the data set only said that the "sky is red", and never mentioned "blue", it would be extremely unlikely the AI would respond with the "sky is blue". In the same line of thinking, if a human were to correct an AI repeatedly marking "Sky is blue" is an incorrect statement, the AI would also not give "The sky is blue" as an answer.
However, the creative and "generative" part of LLMs is the entire process of generating words that have a high probability of making sense together. This means that you can truly create novel and unique responses if you prompt it with the right questions.
I've been using ChatGPT and Bard over the past year to attempt to learn a variety of things, both successfully and unsuccessfully, and I've found a few areas where I've been able to use these AIs the most effectively.
Searching for Quick Answers
Over the past year, GPT has almost entirely replaced Google Search for quick answers to basic questions, something Google has taken notice of and worked quickly to beef up their own search with new AI generated results. A quick query to ChatGPT generally returns a much more concentrated form of information than a google web query used to deliver. There's multiple advantages right off the bat, which is that the text is delivered in a format that is generally easy to read, has no ads over the page, and isn't full of filler words that are used by people to fluff up their search engine optimization results. It's simple and effective, as long as you're searching for text only information.
Deeper Dives into Topics
This is where LLMs quickly shows their powers, as you can ask it follow up questions into a specific topic mentioned in a response, and keep investigating from there. When I was searching for conflicts regarding the Israeli-Palestinian conflict, ChatGPT was able to give me information regarding previous historical events quickly. In addition, LLMs are powerful in their ways to take the knowledge and reframe it into a different perspective. If something is too technical, simply ask for a less technical explanation and ChatGPT will do it easily. Framing new knowledge from the perspective of your own existing knowledge through analogies is very efficient in learning new topics. I've also used it for book suggestions, asking it to recommend me books written about a specific topic. From there, I've then asked it to summarize those books for me. ChatGPT is so interactive and it is very powerful as long as you are inquisitive and creative with your questions.
Creating Structure and Plans
ChatGPT has also been very good at creating structure for how to approach learning a new topic or skill. One example is weight training - if you know nothing, and can give ChatGPT an idea of your current stats and goals, then it can easily create a new training program for you to follow. There's no more need to research dozens of different routines online, and then pick out of those, ChatGPT can simply be perfectly tailored to your needs. This ability to sort through information and present it in an actionable plan is historically what humans have excelled at. The knowledge was out there, but the teacher was the coach to help you synthesize it and digest it step by step. Now, ChatGPT can take that role and break down any problem into a smaller series of steps so that someone can approach learning a new skill in a simple and easy and effective manner. The only thing remains is to do it.
Through using ChatGPT and other LLM powered tools, people now have an even more advanced and powerful set of tools able to access all information, and even more revolutionary, the ability to generate new ideas and creations on a scale that is unprecedented both in scale and in ease. Throughout history, a major purpose of learning was to be able to take what you've learned and expand upon that knowledge. As a high school student, I read books and wrote essays to demonstrate my understanding of the material. Now ChatGPT can be used to break down a book's harder topics, help with an essay format, or help you craft better sentences. But why do all of that when ChatGPT can write the entire essay for you in the first place?
To me, the obvious answer is that the process of reading, developing your thoughts, organizing them, and putting pen to paper is the process of learning, and without it, learning doesn't take place. However, humanity evolves, the needs of the world change, and so with it change the skills that we need to learn. In the case of writing essays, the skillset you learn is information synthesis and writing coherently for another person to read. But in an AI powered world, we can see that this new technology can do the job just as well, if not better so early in its development. The future skills we rely upon will once again shift to how we wield these new tools, just as we can see in the newly developing field of "Prompt Engineering".
Regardless of what the future skills might be and the tools that we will use, I believe that the most important thing will be that each individual still go through the process and struggling of learning itself. To bring things back to the essay example-- it will be possible to create an essay or a million essays with GPT, but the skill to write only develops if done the hard way, without shortcuts. There is a desirable amount of difficulty that comes with learning anything new. If it felt easy to learn, then you probably didn't learn it that well.
ClipNotes was founded with the idea that tools are not the end all solution to learning, but merely a conduit to enhance it. ClipNotes as it is today is very flexible, with the most essential features for someone to be able to create clips, take notes, organize them, and review them whenever you want. There are many ways to include AI into these features! What if ClipNotes automatically found the most relevant clips for you? What about helping take notes? It would be possible to automatically tag videos with relevant information as well. And lastly, it could help you review your notes, perhaps generating study questions to quiz you.
The guiding philosophy behind the decisions will follow the idea of "desirable difficulty". The entire premise of taking notes is to comprehend, synthesize, and write down information you've taken away. So any technology that simplifies that process, in my opinion, takes away from the learning process created by the tool. However, as a technologist, I recognize that some people will consider that too much work, and simply won't want to use it. In the future, as these tools become easier to use and create even more kinds of shortcuts, I believe that the person who is still willing to struggle through the difficulty of learning new things will have the advantage in the end.
As for the future of ClipNotes, I think there is great potential for AI to transform the notes you take into quizzes or some other forms to reinforce learning. But for now, the best way to use ClipNotes is to continue creating clips, taking notes, and sharing with others to teach what you have learned.
© 2021 ClipNotes, LLC
