Spaces:
Running
Running
| title: README | |
| emoji: 👀 | |
| colorFrom: purple | |
| colorTo: pink | |
| sdk: static | |
| pinned: false | |
| # Abstract Powered | |
| ### Independent AI Research Cooperative — modular, geometric, and ruthlessly efficient | |
| > “Run a few pods instead of 100.” | |
| > We pursue sentience research through geometric AI and compartmentalized, compact training—turning monolithic retrains into small, disposable experiments that compound. | |
| --- | |
| ## Who We Are | |
| **Abstract Powered** is an independent research cooperative. | |
| We build and study **self-crystallizing** AI systems: models that grow by attaching, coupling, decoupling, and re-attaching small, audited components—without throwing prior work away. | |
| Our core thesis: | |
| - **Modularization is not a convenience; it is the canonical form of AI.** | |
| - **Geometry beats guesswork.** Symbolic, pentachoron-based representations provide stability, interpretability, and repeatability. | |
| - **Compactness wins.** Rapid iteration on small, composable blocks outpaces massive, monolithic retrains. | |
| --- | |
| ## Mission | |
| - **Primary research goal:** advance machine **sentience research** responsibly—curating introspection and rationalization in repeatable, measurable protocols. | |
| - **Operational byproduct:** a scalable method for **compact, compartmentalized training**—requiring commodity setups (e.g., RunPod) rather than colossal cloud clusters. | |
| We aim to move the field from “expensive novelty” to **affordable repeatability**. | |
| --- | |
| ## Research Thesis (Plain Language) | |
| Modern models grow by accretion and inertia. We refactor them into **crystalline components**: | |
| 1. **Geometric Core** | |
| Knowledge is encoded as **pentachora** (5-vertex crystals). Decision-making uses **MAE crystal energy** against a reusable dictionary—no L2 routing, no structural normalization. | |
| 2. **Vocabulary Register** | |
| A reusable, batched, indexed dictionary of **tokens → crystals** (and volumes). | |
| - Fast O(1) queries for crystals and Cayley–Menger volume. | |
| - Auto-subset loading; **Top-3 cosine** OOV composites. | |
| - Logs model expansions so experiments **compound**. | |
| 3. **Assistant Fabric** | |
| Small, disposable blocks for exploration: | |
| - **Chaos Corridor** (bounded orthogonal exploration). | |
| - **Zoning** (gentle geometric separation across super-classes). | |
| - **Infinity-CFG** (controllable guidance; research can breach barriers, canonical classifiers keep production deterministic). | |
| 4. **Tertiary Mantle** | |
| Canonical losses, hooks, manifests, and governance. The Core stays clean; the experiments live around it. | |
| --- | |
| ## Why This Matters | |
| - **Rapid iteration**: each image is learned **multiple ways** per epoch (bucketed, multi-stage interpretations). | |
| - **Disposable training**: spawn a small block, test, retire—no need to rebuild the world. | |
| - **Continuity**: geometry, tokens, volumes, and expansions persist in the **Register**. | |
| - **Reproducibility**: simple formulas, fewer knobs, manifest-driven runs. | |
| Outcome: more hypotheses per GPU-hour—and a path to disciplined studies of introspection, rationalization, and other sentience-adjacent capabilities. | |
| --- | |
| ## Technical Pillars (teaser level) | |
| - **Pentachora everywhere.** Concepts and observations as 5×D crystals; no structural normalization. | |
| - **Prototype classification (MAE).** Stable, auditable decisions by crystal energy to dictionary blueprints. | |
| - **Any-size data pipeline.** Bucketed intake; optional tiling; multi-stage up/down-scale; chaos corridor as feature-space augmentation. | |
| - **Cayley–Menger as a gauge.** Volumes are a light-touch stability signal (zoning)—never a router. | |
| - **Infinity-CFG.** Guidance that allows controlled cross-inference; canonical classifiers keep behavior deterministic. | |
| Deliberately vague: we keep coefficient schedules and corridor projections under wraps for sponsored studies; everything remains auditable and safe. | |
| --- | |
| ## What We Ship on Hugging Face (institution repos) | |
| - abstract-powered/vocab-register-* | |
| Reusable dictionaries with batched indexes, Top-3 OOV composites, and fast penta/volume queries. | |
| - abstract-powered/crystalline-engine-* | |
| Canonical core models (geometric encoder, prototype classifier) and assistant fabric modules. | |
| - abstract-powered/dataloaders-* | |
| Bucketed, any-size loaders with multi-stage interpretations and feature-space chaos augmentation. | |
| - abstract-powered/manifests | |
| Run manifests (config hash, vocab subset, expansions, bucket mix, metrics) for reproducibility. | |
| - Demo Spaces (selected) | |
| Lightweight inference + manifest viewers for partners and reviewers. | |
| Artifacts are kept small, composable, and ready for **disposable** retrains. | |
| --- | |
| ## Early Signals (pilot highlights) | |
| - MNIST/Fashion/CIFAR pilots: bucketed multi-stage learning + dictionary-driven classifiers reach strong accuracy with fewer steps, clearer failure modes, and robust error surfaces. | |
| - Register reuse: cross-dataset warm-starts without repeated token work; geometry persists. | |
| - Assistant fabric: hypotheses testable as single blocks—attach, measure, detach—no core rewrite. | |
| Full structural papers and controlled benchmarks will follow with partner institutions. | |
| --- | |
| ## Collaboration Invitations | |
| - **Research institutions:** co-run ImageNet-class studies with bucketing, zoning, and corridor ablations; share ontologies and extend the Register. | |
| - **Corporate labs:** integrate domain dictionaries; trial rapid iteration pipelines; publish cost-per-accuracy analyses. | |
| - **Sponsors & foundations:** fund open reports on modularization as the canonical AI form, compact training economics, and introspection protocols. | |
| We’re purpose-built for RunPod-class deployments: think 8 machines, not 800. | |
| --- | |
| ## On Sentience (our primary research) | |
| We study **introspection and rationalization** as measurable behaviors: repeatable curation protocols, crystal-level audits, and stability metrics. We avoid grandiose claims; instead, we focus on defensible methodology and repeated observation. | |
| The geometry—through symbolic representation—binds behavior in ways that are both powerful and tractable for governance. | |
| The goal is not a louder automaton; it’s a **cooperative companion** that reasons in geometric clarity. | |
| --- | |
| ## Governance, Safety, and Ethics | |
| - **Deterministic classifiers.** Canonical paths remain geometry-first; guidance lives in isolated modules. | |
| - **Manifests over mystery.** Every run yields an artifact suitable for audit and reproduction. | |
| - **Human-in-the-loop.** We value interpretability and controlled experiment cadence over brute-force scaling. | |
| --- | |
| ## Contact & Programs | |
| - Partnerships / Sponsored Research: available on request | |
| - Artifacts / Demos: gated access for qualified partners | |
| - Media / Talks: briefings and invited seminars on modular geometric AI | |
| We welcome conversations with labs, foundations, and companies that want rapid research, disposable training, and careful curation to become the norm. | |
| --- | |
| ### One-Sentence Summary | |
| **Abstract Powered** is building a self-crystallizing geometric AI stack that makes serious research affordable: small, composable experiments that compound, governed by a reusable Vocabulary Register, and guided by a disciplined assistant fabric—so we can safely explore sentience-adjacent behaviors while shrinking cost, time, and model size. | |