Grisha

🚀 400 stars on GitHub! Thank you all for supporting recursive-llm. Processing unbounded context with any LLM is the way forward! Appreciate every star and your feedback. 🔗 github.com/ysz/recursive-…

Where dense itself loses: 35 of 150 questions need evidence from 2-3 pages. Joint recall@10 is 11%. Top-k can't grab discrete pages at once. Query decomposition next. Caveat: FinanceBench has unusually severe lexical mismatch. On legal docs BM25 still adds 5-15% to hybrid.

Sharpest gap on metrics-generated questions (50 programmatic "FY2018 X for company Y" prompts): BM25 0% recall@10, dense 57%, hybrid 55%. Questions say "capital expenditure". Filings say "Purchases of property, plant and equipment". Lexical retrieval has nothing to match.

Ran a retrieval benchmark on FinanceBench. 150 finance QA pairs over 368 SEC filings, full corpus. Expected hybrid (BM25 + dense via RRF) to win. Pure dense beat it across every slice. Here is why.

GraphRAG is not “better RAG”. It is a different tool. Hybrid RAG helps you find the right chunks. GraphRAG helps you find the right relationships. Start with Hybrid RAG. Add GraphRAG when the real problem is multi step context, dependencies, and connected entities. Most teams don’t need GraphRAG on day one.

@KevinPicchi VC

Managed to find a list of 2500 VCs (They specialize in AI & SaaS) Never been a better time to get funded Want the list? Comment "VC" and LIKE this post. Will DM in 24 hours.

For now it works for repos with 100+ stars - around 9K repos in the dataset.

GitHub doesn't show you similar repos when you're browsing. So I built a Chrome extension that does - adds a "Similar Repositories" panel right in the sidebar. Finds related repos by who stars them, how similar their descriptions are, and what topics they share. 🔗 github.com/ysz/gh-recomme…

@saipxai I checked it with some small provider. I wanted that exact one, but your name is great too.

@grishahq Which registrar did you use? Got antsintelligence*com in our portfolio.

Today I checked a .dev domain and it was available. Came back 3 hours later to register it - already taken. WHOIS shows it was registered today. Do registrars leak/search-snoop availability checks? Is that even legal?

@eCom_Amin BING

BING ads are a $1M+ GOLDMINE to target 60yo millionaires with ZERO competition and after generating $1,000,000+ with BING i’m giving away EVERYTHING i know about it like + comment “BING” and i’ll shoot you 41-pages of ecom sauce about it (must be following + RT for priority access)

I really enjoyed this paper, and when the first version was released about three months ago, I built my own implementation (400+ stars). I’ve updated it recently as well. You might find it interesting to play around with it. github.com/ysz/recursive-…

link to original post 1password.com/blog/from-magi…

5/ ~6K lines of Python. No open ports. Six security layers by default. A bit more friction, zero supply chain risk. github.com/ysz/nanoclaw

1Password just exposed how OpenClaw's skill registry became a malware vector. Central marketplace + auto-installs = poisoned supply chain. I built nanoClaw with the opposite architecture. Here's why these attacks don't work on it 🧵

📦 Custom skill files are checked for ownership and permissions before loading. Background tasks are capped to prevent resource exhaustion. If you hand an agent real tools, secure the tools first.

🔑 Dashboard API now requires auth. Every audit log entry is HMAC-signed so nobody can quietly rewrite history.

Spent tonight hardening nanoClaw's security 🔒 When you give an AI agent access to files, shell, and the web you better make sure it can't turn them against you. Here's what I locked down 👇