December 27, 2025
Embeddings? Empty calories
Software ate the world. Federation will eat embeddings
Stop hoarding AI data—let smart bots fetch it
TLDR: The post says skip building fancy AI data stacks and let agent-style AI query your existing tools. Readers debated whether this is bold clarity or oversimplified hype, zeroing in on a possible tool-calling typo while arguing if vector databases are dead—or just resting.
The author lights a match under the AI hype machine, yelling: stop building giant data bunkers for chatbots and let agent-style AI pull info straight from the tools you already use. No more “vector databases” and “embeddings” for basic questions—just give an AI agent safe access to your CRM, billing, and warehouse via MCP (Model Context Protocol) and let it work. Cue the comment drama: early readers pounced on the demo, with one sharp-eyed user pointing out the author “calls tool C twice,” sparking a mini flame war over whether this was clever composition or just a typo. The mood? Half “finally someone said it,” half “this is oversimplified.”
Fans of the agent-first approach cheered the promise of real-time answers without months of setup, dropping memes about the “RAG cult” (RAG = retrieve and answer) and “embedding bros” who love overengineering. Skeptics shot back that “you’ll need a data layer eventually,” warning that shiny agents can still hallucinate numbers—exactly the complaint the post levels at RAG. Meanwhile, tooling nerds debated protocol details, linking to modelcontextprotocol.io and asking whether tool orchestration is secretly just another layer. Punchline of the thread: whether this kills vector databases or just delays them, the AI centralization tax is suddenly very uncool.
Key Points
- •The article warns against building a separate AI-specific data layer (vector DBs, embeddings, fine-tunes) before validating use cases.
- •It advocates using agentic AI with tool calling and a capable foundation model to query existing systems directly.
- •MCP and agent orchestration frameworks are presented as available technologies enabling this approach today.
- •RAG is said to work for unstructured document chat but can fail for accurate structured queries and aggregates.
- •Data-warehouse-centric AI layers and knowledge graphs add maintenance and brittleness; agents can often query warehouses directly and reason via tools.