Learning complex topics is not easy, and it is always helpful to have a friend who can explain the direct question. With the rise of the first general language model - GPT-3, new possibilities have emerged for visualizing information with the context and even directly asking questions. LearThis is an experiment with the possibilities of GPT-3 as a tool for learning complex topics such as blockchain, AI or beer fermentation.
Note: This is a prototype. All the content is generated by AI and it is not always 100% accurate. Due to the time constraints, we did not implement a content filter; thus in rare cases some of the content generated by AI can be inappropriate.
What it does
LearnThis visualizes a given topic in the context and provides important additional information such as definition, FAQs, and possibility to ask additional questions.
LearnThis leverages the advantage of GPT-3 of having functionally multiple language models without retraining. We are using it for several language tasks:
1) Finding relevant keywords for the topic
2) Filtering topics based on the schematic score
3) Defining the keywords
4) Generating FAQs about the topic
5) Answering users questions
How we built it
Backend runs on serverless Firebase Cloud Functions, which communicates with GPT-3, after that we use Google Cloud Storage & IBM Watson TTS for audio file generation and storage. For creating the summary document, we are using Google Apps Script.
For the UI we are using Flutter (ready to deploy as mobile app also)
Challenges we ran into
Writing prompts for GPT-3 with respect to minimizing the chance of getting factually wrong answers is challenging.
With graph we realised why on interview are binary tree search that important. We used it for parent regeneration.