Dan Yarmoluk

Knowledge graphs beats RAG. Built CKG: pre-structured domain knowledge for AI agents. 42x more accurate, 11x fewer tokens. Founder @GraphifyMD

[Dan Yarmoluk @YarmolukDan]

@heynavtoor Big words. Just compress knowledge graphs of all orthogonal domains and ask questions Stanford.

[Dan Yarmoluk @YarmolukDan]

@Voxyz_ai It’s like a state of constraints. Fred your agent direction of domain. It’s not rocket science. Pattern recognition with more pattern recognition will not solve drift.

[Dan Yarmoluk @YarmolukDan]

@sairahul1 This is not a playbook. This is musings.

[Dan Yarmoluk @YarmolukDan]

@karpathy It does have great long thinking and taking it to a surprisingly further than anticipated. I don’t have context issues, I compress knowledge graphs, so it’s dope

[Dan Yarmoluk @YarmolukDan]

@cjzafir This works - routing cuts spend per task. The other 10X: cut the tokens themselves. Stop re-deriving context every session -- compile it one into a graph the agent reads instead of re-searching. Measured: 11x few tokens than RAG....free ckgs if you want them

[Dan Yarmoluk @YarmolukDan]

@karpathy Nailed the pattern. LLM as a bookkeeper solves who maintains the wiki.

Next layer: compile it into a graph. Typed edges, confidence scores, deterministic retrieval, no re-reading 200 pages per query. Built CKG - 3.8 X F1, 11X fewer tokens vs. RAG.

[Dan Yarmoluk @YarmolukDan]

@trq212 I think the assumption more pattern recognition on RAG pattern recognition will solve drift might be faulty.

[Dan Yarmoluk @YarmolukDan]

@trq212 Those harnesses are not good enough. I structure knowledge and don't harness my mustang...maybe folks should start looking at domain vs. a series of bloated SWDLC constraint-based systems.

[Dan Yarmoluk @YarmolukDan]

@elonmusk You both are cool. But what about 42x more token intelligence cool, how many messages a fella got to send

[Dan Yarmoluk @YarmolukDan]

65× higher intelligence-per-token** vs RAG/GraphRAG:

• Macro F1 0.47–0.53 (0.77 at hop=5)
• Only ~269 tokens/query
• 0% hallucinations by design (pure Markdown/CSV graphs)
• Strong multi-hop & agent-ready (MCP server)

Perfect lightweight enabler for Grok API — inject compact subgraphs into your massive context + tools.

GitHub + benchmarks: https://t.co/AOVaqBCz4u

@xai @elonmusk @grok #GrokAPI #xAI

[Dan Yarmoluk @YarmolukDan]

@housecor

[Dan Yarmoluk @YarmolukDan]

@elonmusk Please beam it here, I got shit to do

[Dan Yarmoluk @YarmolukDan]

@XFreeze it's going pretty fast full throttle, and I do appreciate the bias for this little thing, probabilistic truth

[Dan Yarmoluk @YarmolukDan]

@elonmusk yeah, but ultimately when I inquire, the hardware or whatever other BS is just regular speeds, and if not, please show me something as I'm building things

[Dan Yarmoluk @YarmolukDan]

@tegmark What up Max

[Max Tegmark @tegmark]

Anthropic 2024: You can trust that we'll keep all our safety promises
Antropic 2026: Nvm

[Dan Yarmoluk @YarmolukDan]

@art_muse Dope, for sure

[Dan Yarmoluk @YarmolukDan]

@elonmusk my experience with Grok has been very positive. That is definitely a rocket ship...problem is, I get throttled back. I'm not a car guy, but aren't I supposed to take a German car on the autobahn?

[Dan Yarmoluk @YarmolukDan]

@can :)

[Dan Yarmoluk @YarmolukDan]

@tensor_rotator and why can't we measure intentionality with all these terminal button clicks, quality of input, length of input, time iterating input? Like you care, like you want a better output kinda thing.