Releases·Tool Release
Nordri PRTG 0.1.1: An Operable Experimental CPU First Model Based on Plastic Relational Tensor Graphs Architecture
A developer shared Nordri PRTG 0.1.1, an early CPU-only prototype that represents information as a graph of entities and relationships carrying learned mathematical properties, rather than data flowing through a fixed network. The model can be deliberately taught, perform limited reasoning over what it learned, and expose its computational path to an answer. Version 0.1.1 adds training tools, state persistence, and behavior comparisons against simpler versions.
2 engagement1 sourceSat, Sep 5, 2026, 01:28 AM
AI Overview
- Nordri represents information as a graph of things and relationships carrying learned mathematical properties, and can show its computational path to an answer.
- Version 0.1.1 adds tools for training, saving model state, and comparing behavior against simpler versions.
- A saved model contained 2,007 nodes and 1,604 active relationships; 3D images compress its higher-dimensional geometry into a partial view.
- In one test it answered 1,797 of 1,797 cases: 1,397 arithmetic problems and 400 reasoning questions on facts from artificial test worlds.
- Disabling the mathematics stored in relationships made the model decline to answer the 400 reasoning cases, evidence the relationships participate in computation.
- Developer seeks feedback from others working on Hebbian or Plastic Tensor Model architectures.
Related
Developments·Paper
BaryGraph: A Relational Geometry for Cognitive AI
A project called BaryGraph describes a recursively constructed relational vector architecture for AI memory and reasoning. Starting from a flat semantic substrate, it builds triadic objects joined by stored relational vectors, which then form higher-order structures entirely in vector space. A model navigates the resulting semantic landscape via message passing, returning bounded semantic constructions rather than fluent continuations. The work, shared on the author's site, is framed as an experimental architecture rather than a finished system or claim of consciousness.
A project called BaryGraph describes a recursively constructed relational vector architecture for AI memory and reasoning. Starting from a flat semantic substrate, it builds triadic objects joined by stored relational vectors, which then form higher-order structures entirely in vector space. A model navigates the resulting semantic landscape via message passing, returning bounded semantic constructions rather than fluent continuations. The work, shared on the author's site, is framed as an experimental architecture rather than a finished system or claim of consciousness.
3 engagement1 sourceredditThu, Sep 24, 2026, 06:52 PM
Products·launched
Weave: plain-language notes to interactive knowledge graph with MCP memory server
Weave is an open-source tool that converts plain-language notes into an interactive knowledge graph by extracting concepts and relationships, and it also provides an MCP memory server for AI agents. It solves the problem of manually organizing notes and connections, targeting users who want automated knowledge management.
Weave is an open-source tool that converts plain-language notes into an interactive knowledge graph by extracting concepts and relationships, and it also provides an MCP memory server for AI agents. It solves the problem of manually organizing notes and connections, targeting users who want automated knowledge management.
1 engagement1 sourceredditMon, Aug 24, 2026, 09:11 PM
Developments·Paper
Differentiable Clone-Structured Causal Graphs Enable End-to-End Cognitive Map Learning from Raw Images
Researchers propose a differentiable version of the Clone-Structured Causal Graph (CSCG) algorithm, enabling end-to-end learning of interpretable cognitive maps directly from raw image sequences. This removes the previous requirement for a predefined discrete alphabet and allows integration with neural network modules, advancing biologically-inspired spatial reasoning in AI.
Researchers propose a differentiable version of the Clone-Structured Causal Graph (CSCG) algorithm, enabling end-to-end learning of interpretable cognitive maps directly from raw image sequences. This removes the previous requirement for a predefined discrete alphabet and allows integration with neural network modules, advancing biologically-inspired spatial reasoning in AI.
0 engagement1 sourcearxivTue, Jul 14, 2026, 05:54 AM
Developments·Paper
Research paper finds linear representations of grammaticality in neural language models
A new paper on arXiv investigates whether neural language models encode grammaticality as a linear feature in their internal representations, moving beyond probability-based measures. The study addresses ongoing debates about models' grammatical knowledge by probing internal states rather than relying solely on output probabilities.
A new paper on arXiv investigates whether neural language models encode grammaticality as a linear feature in their internal representations, moving beyond probability-based measures. The study addresses ongoing debates about models' grammatical knowledge by probing internal states rather than relying solely on output probabilities.
0 engagement1 sourcearxivThu, Jul 16, 2026, 04:23 PM
Products·building
Visual editor that compiles drag-and-drop graphs into PyTorch code
NeuroBranch is a visual editor for designing neural network architectures by dragging typed cards onto a canvas and wiring them together. It compiles the graph into runnable PyTorch code, supports two-way sync between graph and code, and includes an AI copilot. It solves the problem of unreadable neural net code for developers who want to see and edit models visually.
NeuroBranch is a visual editor for designing neural network architectures by dragging typed cards onto a canvas and wiring them together. It compiles the graph into runnable PyTorch code, supports two-way sync between graph and code, and includes an AI copilot. It solves the problem of unreadable neural net code for developers who want to see and edit models visually.
1 engagement1 sourceredditSat, Aug 1, 2026, 07:01 AM