📄 #POTD: Today's paper is a special one for us. It's the official blueprint for our new project, #SciencePedia.
It exposes a fatal flaw in how we all store knowledge: "radical compression."
We’ve all been taught to ignore the "dark matter of knowledge." Until now.
Thread 👇
It exposes a fatal flaw in how we all store knowledge: "radical compression."
We’ve all been taught to ignore the "dark matter of knowledge." Until now.
Thread 👇
Here are the links:
👉 Explore SciencePedia: bohrium.com/sciencepedia
📑 Read the Paper: j1q.cn/jSh0a2q1
Full Paper Title: "Inverse Knowledge Search over Verifiable Reasoning: Synthesizing a Scientific Encyclopedia from a Long Chains-of-Thought Knowledge Base"
👉 Explore SciencePedia: bohrium.com/sciencepedia
📑 Read the Paper: j1q.cn/jSh0a2q1
Full Paper Title: "Inverse Knowledge Search over Verifiable Reasoning: Synthesizing a Scientific Encyclopedia from a Long Chains-of-Thought Knowledge Base"
The paper explains why this matters. "Reasoning compression" is a root cause of #LLM #hallucination.
When LLMs mimic human data, they inherit this "dark matter blindness."
We had to start from first principles.
When LLMs mimic human data, they inherit this "dark matter blindness."
We had to start from first principles.
Our method: Decompress scientific reasoning.
It starts with a “Socrates Agent”.
This agent generated ~3 million first-principles questions across a curriculum of ~200 courses, from high-school to graduate level .
It starts with a “Socrates Agent”.
This agent generated ~3 million first-principles questions across a curriculum of ~200 courses, from high-school to graduate level .
Next, multiple independent "solver models" generate Long Chains-of-Thought (LCoT) for each question.
To ensure high fidelity, we only keep answers verified by "cross-model answer consensus" .
This creates a massive, verifiable LCoT knowledge base.
To ensure high fidelity, we only keep answers verified by "cross-model answer consensus" .
This creates a massive, verifiable LCoT knowledge base.
This LCoT base powers our "Brainstorm Search Engine" .
It performs "Inverse Knowledge Search".
You stop asking "What is an Instanton?"
It performs "Inverse Knowledge Search".
You stop asking "What is an Instanton?"
Instead, you retrieve all the first-principles paths that derive the concept of an Instanton.
This reveals its non-trivial connections:
Quantum tunneling in double-wells
QCD vacuum structure
Hawking radiation
4D manifolds in pure math
The "dark matter" is made visible.
This reveals its non-trivial connections:
Quantum tunneling in double-wells
QCD vacuum structure
Hawking radiation
4D manifolds in pure math
The "dark matter" is made visible.
This is #SciencePedia. Our "Plato Synthesizer" narrates LCoT chains into coherent articles.
The result? Articles with ~50% lower factualhallucination and significantly higher knowledge density than baseline LLMs .
#SciencePedia is now live with ~200,000 entries. Come explore.
The result? Articles with ~50% lower factualhallucination and significantly higher knowledge density than baseline LLMs .
#SciencePedia is now live with ~200,000 entries. Come explore.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
