MIT's Embedded Vector DB "Caliby" Goes Open Source
**Caliby**, an embedded vector database jointly developed by Xinjing Zhou of the MIT Database Group (advised by Professor Michael Stonebraker) and Jinming Hu of Sea-Land AI, has been released as open source. It is available under the MIT license and can be installed via `pip install caliby`.
Key Features
Supports three index types — **HNSW / DiskANN / IVF+PQ** — scaling to vector collections ranging from millions to tens of millions
Unified management of text and vectors**: embeddings and raw text live in the same system, eliminating the need to shuttle data back and forth between a vector DB and a relational DB
Disk persistence**: overcomes FAISS's weakness of becoming unusable when data doesn't fit in memory; indexes are automatically restored after process restarts
CPU optimization via **SIMD acceleration** (AVX-512 / AVX2 / SSE) — no GPU required
Performance Comparison
Compared with pgvector (a PostgreSQL extension), Caliby achieves roughly 2× faster index builds and about 4.5× higher query QPS. There is no PostgreSQL instance to operate either — like DuckDB, everything runs in-process. Against FAISS, it solves the "stops working once data exceeds memory" problem through disk spilling.
Approach to AI Agents
Rather than aiming to be a general-purpose vector DB, Caliby is positioned as a data management engine **specialized for AI agent / RAG workloads**. It provides cross-session memory persistence, metadata filtering, and multi-index isolation — the challenges that arise with LangChain, CrewAI, and similar frameworks — in a single library.
> "AI agent and RAG scenarios need a lightweight, embedded data engine like DuckDB."
The development team's vision is to become "the DuckDB for AI agent data," with GPU acceleration and TypeScript bindings planned for the future.
---
AWS MCP Server Officially Released (GA)
On May 6, 2026, the **AWS MCP Server** became Generally Available (GA). AI coding assistants such as Claude Code and Cursor can now execute more than 15,000 AWS operations directly.
What Changes
Previously, working with AWS resources meant navigating the AWS console or running CLI commands. With the MCP (Model Context Protocol) Server, AI assistants can now invoke AWS APIs directly — you can instruct them in natural language to inspect infrastructure configurations, create or modify resources, check costs, and more.
A Qiita write-up evaluates the release as having fundamentally solved the pain of "opening the AWS console is a hassle."
---
ACM Sounds the Alarm on "Vibe Coding"
The Technology Policy Council of the ACM (Association for Computing Machinery) has published a report expressing **strong concerns** about "vibe coding," an AI-driven coding practice.
Key Takeaways from the Report
AI coding platforms have a structural problem: they make it easy to **skip fundamental software engineering practices such as design, testing, and review**
The report asserts that "AI does not understand," emphasizing the importance of mechanisms for humans to verify the correctness of generated code
It warns that the proliferation of vibe coding (the practice of prompting AI to write code for you) risks degrading software quality
Notably, an opinion piece titled "I Will Never Use AI to Code" earned 30 points on Hacker News the same day, suggesting that debate over the merits of AI coding is heating up in the community as well.
---
Other Notable Moves
Restrictions on free Qwen usage**: Reports are surfacing in Reddit's LocalLLaMA community that some form of limitation has been placed on free access to Qwen models. Details are unclear at this point, but it hints at a changing landscape for open-source LLM usage
MTP support in llama.cpp**: Voices in the community are calling for official llama.cpp support for Multi-Token Prediction (MTP), reflecting growing demand for faster local LLM inference
---