We were inspired primarily by the co-founder's (A. Brady-Mine) experience developing educational plans to teach human rights through their non-profit. From this, we already had an idea of the value of well-designed and well-informed lesson plans, especially in increasing access to education. The other co-founder (J. Falagan) is an environmental engineer with specific interests in conservation and climate change mitigation. Combining these two interests led to an obvious conclusion, provide relevant, high-quality information about the global environment that incorporates scientific information alongside studies on global sustainability and action efforts. Due to the desire for the product to be easy to access and highly scalable, it was decided the ideal prototyping platform would be a Squarespace website integrated with OpenAI.
What it does
After entering the website, the teacher is given a description of the product and is directed to fill out the lesson plan request form. They select a range of student ages, a lesson plan length, and submit a comma-separated list of desired environmental topics. Within 5 minutes, the teacher receives an email with the lesson plan.
How we built it
The initial form response was taken from the Squarespace page and stored in Google Sheets. From there, the information was passed through Zapier where it was combined with a preset prompt that showed to be the most effective for generating GPT-3 prompt. Multiple prompts were tested and optimized to produce lesson plans that are informative, concise, and accurate relative to the user-selected options. The prompts are then passed to GPT-3 and the developed lesson plan is passed back to Zapier where it is emailed to the user.
Challenges we ran into
The two major challenges we faced were determining a singe prompt capable of producing consistent results for a wide range of topics related to environmental issues. We went through several iterations of prompts and eliminated those that provided inaccurate or incomplete results. We also had issues with triggering the workflow when the forms are submitted. Due to the method that Squarespace uses to update google sheets, the responses were not being triggered. We were able to work around this issue by using a modified or changed command in Zapier.
Accomplishments that we're proud of
We were able to create a functional app and integrate previous outside knowledge with practical development tools.
What we learned
How to create effective workflows using GPT3. Incorporating language learning models into simple user-facing projects.
What's next for EcoLessons
A crucial next step is working on validating the information provided by the model, in particular identifying and reducing hallucinations. In the future we hope to feed GPT3 a pre-selected set of external sources with known accuracy related to the topic requested. This has been shown to significantly reduce the likelihood of hallucinations. This method would also allow us to provide educators with the resources that were used in crafting the final material.
View and Password: https://turtle-dove-hx63.squarespace.com/ Password: 123abc