HackerNews Summary (2023-05-25 09AM)

Source: ChatGPT plugins now support Postgres & Supabase

Summary
Supabase contributed Postgres and Supabase implementations to help developers build “retrieval plugins” for ChatGPT. These plugins allow ChatGPT to access up-to-date information from a database and answer questions from external data sources, such as PDF documents, Confluence, or Notion knowledge bases, by converting them into smaller chunks and querying a vector database for relevant documents.

Fact

🔌 A Retrieval Plugin is a Python project designed to inject external data into the ChatGPT, allowing it to dynamically pull relevant information into conversations from your data sources.

🔍 The retrieval plugin converts documents into smaller chunks, converts the chunks into embeddings using OpenAI’s text-embedding-ada-002 model, stores these embeddings into a vector database, and queries the vector database for relevant documents when a question is asked.

📚 You can choose your preferred vector database provider from a list of supported options, such as Supabase and Postgres.

🧩 Supabase’s Retrieval Plugin version interacts with the vector database via PostgREST, which is helpful for using Row Level Security or if you are planning to use the data in the Retrieval store beyond ChatGPT.

👥 The Postgres implementation is great to start with because there are now a large number of providers supporting pgvector, and both versions have the same schema so you can easily switch between them.

---

Source: PostgresML raises $4.7M to launch serverless AI application databases based on Postgres – PostgresML

Summary
PostgresML has secured $4.7M funding to launch serverless AI application databases, bringing AI tasks to the Postgres database to simplify app development and reduce complexity. Its custom Postgres load balancer, PgCat, pools connections to create a mesh of Postgres clusters that appear as independent Postgres databases, enabling multi GPU inference workloads. PgCat is MIT-licensed, scalable and offers free databases with up to 5GB of data.

Fact

💾 PostgresML’s extension for Postgres brings AI tasks to the database engine, reducing complexity for app developers and saving time.

🥇 PostgresML can load state-of-the-art LLMs and datasets directly from HuggingFace, and combining open source embeddings, vector database and input text results in a single process is up to 40x faster than alternative architectures for semantic search.

🌟 By integrating leading machine learning libraries like Torch, Tensorflow, XGBoost, LightGBM and Scikit Learn, developers can train their own models for better ranking and recall.

🆓 PgCat is scalable and offers free databases with up to 5GB of data, only charging for caching custom models, data, and indexes for maximum performance.

---

Source: Andreessen Horowitz

Summary
A curated list of resources called the “AI Canon” has been compiled to share important papers, blog posts, courses and guides about modern AI. It offers a gentle introduction to transformer and latent diffusion models, technical learning resources, practical guides to building with large language models (LLMs), and analysis of the AI market.

Fact

💻 “Software 2.0” by Andrej Karpathy explains why AI is a new and powerful way to program computers.

🤖 “State of GPT” by Karpathy is an approachable explanation of how ChatGPT/GPT models work and what directions research and development may take.

🤖 “Transformers, explained” by Dale Markowitz is a shorter, more direct answer to the question “what is an LLM, and how does it work?”

🧠 Stanford CS229 is an introduction to machine learning with Andrew Ng.

📈 “Stanford CS224N” is an NLP with deep learning course with Chris Manning, covering NLP basics through the first generation of LLMs.

---