How Luna uses AI
3 minute read
Just like a lot of recent products worldwide, Luna incorporates AI. Of course, this opens up the question… how does it use AI?
It’s a fair question! Nowadays, it seems like AI is used everywhere, from massive consultancies to small-scale digital artists. They can’t all be using the same AI model in the same way, as this would open up a can of worms and make it harder for consumers to know what’s going on.
We want to be upfront about the AI technology Luna is utilising and how it is doing this, in order to put minds at ease.
What model is Luna using?
Luna uses ChatGPT 3.5-turbo, which is one of many OpenAI models available. The GPT models are perhaps the best known, and lots of you may have used the free ChatGPT to help with your work, as an alternative to Google, or just to have a bit of fun. If you haven’t played around with it yet, it’s well worth a look; you’ll probably be surprised by just how intelligent artificial intelligence is!
GPT-3.5 models can understand and generate natural language. This was perfect for Luna, as we wanted to create flashcards from natural language (i.e., students’ study materials), and then generate questions and answers in language that is used in typical everyday life.
Other Open AI models include DALL·E, which creates images from a text prompt, and Whisper, a speech recognition software. There are other natural language models too, and Luna was initially built using the
text-davinci-003 model. However, during the development process ChatGPT 3.5 quickly became the preferable option, highlighting just how rapid AI’s development is.
The odds are that Luna will switch models again with the release of ChatGPT 4 so it can benefit from performance improvements. We’ve designed Luna so we can easily switch the AI model it uses because we expect so many developments in a short period of time.
How does that become flashcards?
The text that the user puts into Luna is fed to the AI model, which then identifies key insights. It does this by analysing the text against billions of other data points it’s been trained on: it’s a bit like having an expert in the field skim the text for key points. The AI is leveraging an immense amount of related knowledge that it has access to.
Once it has those key points, Luna allows the user to check and edit any information they feel needs tweaking, then feeds it back through the AI again - this time, the AI is figuring out how to turn each insight into a true/false and a short answer question. That then forms the basis of the flashcard. Et voila, Luna has turned a significant chunk of text into a collection of flashcards. The user can then customise them however they please, but the AI has done the grunt work that used to put students off making flashcards.
How OpenAI uses data from Luna
Data sent by Luna to the OpenAI API will not be used to train or improve OpenAI models. Data from Luna may be retained by OpenAI for up to 30 days (in order to help identify abuse), after which it will be deleted (unless otherwise required by law).
AI can seem monstrous, given its many uses and how it’s discussed in the media at times. But what’s been described here is the extent to which Luna uses AI to help students. It can be likened to having a really, really, really clever person pick out the bits they need to learn from and turn them into questions. It’s a huge timesaver, and it helps with the beginning of the learning process.