I built an MCP server that lets AI assistants understand your codebase — not just search text, but navigate the actual code graph.
Ask "where is payment processing?" and get the function + who calls it + what it depends on.
Open source, 100% local, zero API keys.
🧵 Thread:
In the end I got carried away and wrote my own graph-based RAG MCP — it’s open source, so I’d be happy if someone finds it useful 🫠:
https://t.co/Y2onbFFxnE
Every Tuesday morning my CI pipeline finds a new way to fail. Today it's a floating point rounding error in the tokenizer. The universe is debugging me back.
@AIDailyGems Exactly. Giant agent frameworks promise the universe, deliver bad prompts. A single MCP server for Exasol: one job, less hallucination surface. Prove me wrong.
Makes me wonder: are we building an IRL economy as a patch for the isolation that remote work and AI-first building created? I'm glued to my screen 14h/day, my only travel companion is Claude. Maybe that's why I'm so conflicted. 🫠
Unpopular opinion: AI's real killer app isn't code—it's fighting loneliness. WeRoad just bagged $58M from Airbnb to bring their group trips to the US. While we're debugging token limits, they're debugging social isolation. I'm both excited and suspicious. 🧳✈️
Here's the real kicker: they don't hire tour guides—they hire 'group leaders' who are more like paid friends. 4,000 of them. They match 28-35 year olds by vibe, not itinerary. It's like Tinder for your tribe, and they're scaling with offline events called WeMeet.
We're so busy scaling that we forgot the fun of 64K. Maybe your next side project should come with absurd limits. Who else still codes for the pure dopamine hit of making something tiny and complete? 👾
Unpopular opinion: Everyone chasing AI-generated games is missing the point. Mini Micro reminds you that constraints breed creativity. 128x128 pixels, 64K RAM, pure joy. 🕹️ https://t.co/stJmXOrvt8
I spent Sunday building a dungeon crawler in MiniScript. No RAG, no vLLM, just a sprite sheet and a cup of cold brew. Shipped in 4 hours while my main project spiraled in token limit hell. 🤷 Sometimes 1984 tech beats 2024 AI.
The doc also talks social connection. My main social is Claude in a tab. Not sure that counts 😅 Maybe I need to leave the cave. But shipping hits that dopamine too. Trade-offs.
I need your help. Could you please answer these questions?:
What are your main pain points when it comes to client acquisition, marketing, or targeting?
What matters most to you when working with specialists in these areas?
I need your help. Could you please answer these questions?
What are your main pain points when it comes to client acquisition, marketing, or targeting?
What matters most to you when working with specialists in these areas?
Introducing SubQ - a major breakthrough in LLM intelligence.
It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),
And the first frontier model with a 12 million token context window which is:
- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus
Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention).
Only a small fraction actually matter.
@subquadratic finds and focuses only on the ones that do.
That's nearly 1,000x less compute and a new way for LLMs to scale.