Building with Gen AI #1
Challenges, Learnings & a Bit of News from the Builders of Quivr.app
Welcome to the first Building with Gen AI newsletter!
Here we will discuss the new features and learnings from the team at Quivr as well as news from across the exciting world of Generative AI.
Quivr is sponsored by Theodo, Aleios, Padok and BAM. Resources are shared from these companies to work on the project. Outside contributions are always welcome and we have a growing network of active contributors.
If you were forwarded this email you can sign up for the newsletter here.
This Week in Numbers
12 Releases: v0.0.18 to v0.0.29
9 active Contributors this week
1800 new users
6.2k Questions asked
9k Files Uploaded
New Features and Improvements
Multiple brains - Users can create and use multiple brains!
OpenAI Functions implementation
Implemented Tests with Pytest & Vitest
Created a Preview environment
Learnings from the Week
Stan - Sensei: We now have a preview environment which reduces bugs in production. Automated tests are key to ship with confidence! Automate database migrations to avoid pain.
Mamadou - Frontend Wizard: Clean code is more important than comments or docs. Python is not bad and very approachable coming from a Typescript background. Testing and observability are game changers - We know what went wrong without relying on user feedback!
Matt - Backend Dude: Thinking about lowering the barrier to entry for a new contributor. Using Docker rather than native cloud Serverless for example. Also splitting up the project across multiple domains and shipping small features is super important to avoid merge conflicts with many contributors working quickly.
Chloe - Product Maetro: The goodwill of the community is our greatest asset! Contributors share their opinions on the direction of the project and give their time, strength and encouragement. As the core team we must concentrate on communicating and maintaining transparency with regard to goals, decisions and challenges.
Brian - Methodology King: Needed to tweak the lean methodology for open source but we working on bringing in relevant elements. We now have tickets for our sprints in the roadmap and are working on more transparent communication with unknown stakeholders such as the daily-updates channel on Discord.
News from the World of Gen AI
Check out this article by Benjamin Marie to learn how to fine-tune your own LLMs with QLoRa.
MidJourney released V5.2 with a bunch of new feature, including:
Improved aesthetics and sharper images
New “high-variation” mode
New “/shorten” command
New “Zoom out” feature
Google’s Deepmind looks to be working on a new ChatGPT killer called Gemini from this Wired article
This article reveals that GPT4 isn’t the singular most powerful model as everyone thought but is a mix of multiple specialized models (around 8 220B parameters models 🤯)
AWS released their whitepaper on Cloud Adoption Framework for Artificial Intelligence, Machine Learning, and Generativ