Lola (Learn online, like actually)

• 

  LOLA!

Inspiration

We built LOLA to let anyone learn accurate, interesting information from Youtube in a more accessible way. We've built a few subject-specific tutors before for classes like AP World History and skills for financial literacy, and our goal in building LOLA was to let student's have the same kind of personal 1:1 tutoring experience on any subject they'd like to learn about, from coding to cookie-making.

What it does

LOLA starts off by understanding what its student wants to learn about. It asks questions trying to specify what kind of Youtube playlist it should use, finds some that are appropriate for the student's topic of interest and skill level, and shares those with the student, letting them select/confirm the playlist they want to learn about.

Then, LOLA forms a memory of the playlist as a whole and of each video in the playlist. This memory is two-part: its topic/fact-based, so that LOLA can use it to answer questions answerable by the playlist, and its playlist and video organized, which means LOLA can select the playlists and/or videos that are most useful to helping a student and even intelligently query its own memory, using those video or playlist memories as navigation to answer a question requiring multiple different kinds of information. LOLA uses this memory to answer questions for students in a digestable way, and even play relevant video clips for them, as short-form and on topic as a Tik Tok.

So, overall, LOLA is a new form of learner with structured memory and a more intuitive understanding of the Youtube landscape, and its also a new form of AI tutor with smarter memory retrievals and a student-oriented and personalized presentation style. It seeks to make the majority of learning topics dramatically more accessible by creating a knowledgeable friend in that area, trained on entire playlists.

How we built it

LOLA is built on two tracks.

The first track is for understanding Youtube, according to a student's topic of interest and experience level, and the second is for delivering that information efficiently and digestably to its students.

The first track, comprehension, is built of the following chain of actions:

1. LOLA decides what the student wants to learn about, in conversation with them
2. At some point in the conversation, it outputs in a specialized and parsable format, the most relevant Youtube query and the behavior of the student notetaker.
3. The chain picks up on the fact that a query has been made, and finds the top Youtube playlists for that query.
4. It shows the playlists to the students, and lets them choose their preference.

The second track, tutoring, is built of the following chain:

1. LOLA finds the most relevant video in a playlist for a query, using Weaviate natural language vector similarity to search through video search labels (texts containing descriptions of what's covered in a video).

2. LOLA finds the most relevant topic from that video by vector searching that video's topic chunks.

3. It uses the information its retrieved to answer student questions with highly accurate, engaging knowledge.

Challenges we ran into

Prompting chatGPT to build its own knowledge is always super difficult, especially since we set out to build an efficient (non GPT4) system that can interpret ANY kind of information. Writing prompts that successfully build topic-based knowledge, the kind a tutor of that area would need to know, on any dataset is very difficult to do. We also ran into challenges setting up a demo of LOLA, since our backgrounds are more in AI, prompting and chaining than any kind of user-project deployment. This was definitely one of our greater difficulties.

Accomplishments that we're proud of

We're proud that LOLA can teach anyone anything from cookie-making to caching, and we're proud of its new organized memory system, which we think will contribute to the progression of sub-LLM AI systems into the future.

We're most proud of the ease LOLA brings to learners. As students, we feel the biggest impact we've made is saving ourselves and other students hours on searching for information that could be intelligently synthesized almost instantly.

What we learned

Lots. We learned most on the language side of the project. We learned how Youtube queries represent certain meanings and most of all we learned a ton about developing frontends. :P

What's next for Lola (Learn online, like actually)

Next for LOLA, we're going to try to improve efficiency with an LLM-specific cache, and make complex memory retrieval actually work in our project. We'll be releasing it to the public in the coming weeks, once we solve efficiency concerns and add efficient voice.

Built With

• faiss
• fastapi
• openai
• python
• tailwind
• weaviate