How to train your dragon .... oops.. AI?
If you cant beat em, join em
Can we reverse the process where students avoid learning by using AI tools, and instead have them train a model on a specific domain or topic? Teaching is the best way to learn and maybe we could have students teach an LLM?
As I researched this idea, I found a similar initiative that worked with custom software built at Vanderbilt, called 'Betty's brain' . Not sure if anyone has tried this with LLM's in class yet, but this is exactly what I am getting at. Another trend that informed this idea is the proliferation of firms providing capabilities to provide human feedback to improve LLM’s. One of the biggest firms in the space is Scale, which provides services to OpenAI to help improve the output from ChatGPT and such. So when you give a ‘thumbs down’ to a chatGPT response, some poor college student contracted by Outlier AI, which gets business from Scale AI, might land up with your conversation a couple days later. How do I know? my kids are training the models on outlierAI :). Perhaps we could leverage the protocols these firms use to train the models in class?
The final score for the class or module could then be based on the performance of their AI model, assessed through automatic benchmarking by a more advanced language model. The instructor would only need to train this smarter model or could use the most advanced general purpose model and create the bench marking questions and answers.
We constrain the capability of the LLM that students have been given, host it in a controlled environment so they cannot just dump an entire PDF textbook into it or feed it unlimited text for context. For starters we can stick to text based models and not include vision enabled models. They are encouraged to use external AI tools to ask questions, get answers, and synthesize the knowledge and then train their individual LLM. Another interesting spin could be if they are expected to create a knowledge graph from the readings they are given or other external sources and then use that along with the LLM.
The instructor would most likely use a flipped classroom model and perhaps could continue as earlier with some in class group problem solving/discussions. As the students learn the foundations of a topic, they could get feedback from other students testing their model or perhaps another advanced LLM, so as to replicate scaffolding. We could have some kind of instructor/TA in the loop as well. Students reflect on their training process and document it.
Each LLM instance for a student could be a docker container that could be used to create multiple instances in a cloud provider. We would use an open source LLM, so there are no licensing costs and no data leakage issues. A cheaper way would be to setup the docker container in each students laptop as long as there is a way to ensure that they dont tamper with the install.
By training their AIs, students gain insights into learning and teaching with AI as a collaborator and gives them skills that they can use to learn any topic and also prepares them for the workplace.
What do you think?