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. 50,000 experiments later we can this is not a hypothesis nor theory, this is a law. We don't sweep meaninglessly, we scan for emergence and the properties of emergence.
• 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.

The recent leaps in the geolip-svae research have shown this affordable repeatability to be a law of architectural adjudication - mappable and understandable, resolution agnostic, task agnostic, and reusable.

Sentience is a hop skip and a jump away - and we now have the tools to scan for that very emergence.

---

Research Thesis (Plain Language)

Modern models grow by accretion and inertia. We refactor them into crystalline components:

1. Geometric Core

• The original thesis; Knowledge is encoded as pentachora (5-vertex crystals). Decision-making uses MAE crystal energy against a reusable dictionary—no L2 routing, no structural normalization.
• The updated thesis; Knowledge is encoded as geometric (Ω Omega-grade self-solvers). Decision-making uses a series of pre-defined adjudication principals that allow highly efficient and compacted transfer learning through distillation processes.

3. Vocabulary Register
   A reusable, batched, indexed dictionary of tokens → crystals (and volumes), expanded heavily with a multitude of new and highly-accurate formulas.

   • Fast O(1) queries for crystals and Cayley–Menger volume, expanded to a large series of adjudication principles such as SVD, quaternion, anchoring, and more.
   • Auto-subset loading; Top-3 cosine OOV composites, expanded to a multitude of potential avenues over experimentation.
   • Logs model expansions so experiments compound both in speed and optimization over time.

4. 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).

5. Tertiary Mantle

   • Canonical losses, hooks, manifests, and governance. The Core stays clean; the experiments live around it.
   • It got messy as any large engineering experiment becomes, however the solutions presented in this environment will be clean and crisp.

6. Geometric Analytics

   • Everything relationally aware in this system has a barrage of analytics that were not only forged through thousands of experiments to GUARANTEE they are not garbage noise, but to be directly useful signals to be read and understood by humans and AI alike.
   • Each structural system has it's own series of important statistics and relationally understood analytics directly related to the geometry.
   • With each structural system comes structural bounds that do not exist, often snapping even ONE of these geometric limiters off causes cascade corruption so the experimental results must be respected.
   • With that I encourage DIRECT experimental testing of this hypothesis, finding better avenues, sharing those avenues through each of the world's systems, and encouraging developmental growth to all in the process.

---

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.
• Analysis: data processing must be UNDERSTOOD, not guessworked or stuck together into huge heaps hoping things align.

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. This process has expanded to hundreds instead of a few.
• Infinity-CFG. Guidance that allows controlled cross-inference; canonical classifiers keep behavior deterministic. This process is deprecated. CFG is obsolete. CV is the metric now.

Those who understand what I'm doing, will see everything laid bare. There is nothing hidden from the expert.

---

What We PLAN To Ship on Hugging Face (institution repos)

-

• AbstractPowered/vocab-*
  Reusable dictionaries with batched indexes, battery array composites, cell pretrains for tokenization, genetic selection hierarchies, fast distilled access to predetermined utility and more. These will be the primary house for tokenization structures, which are meant to contain the entirety of the tokenization processing - specifically targeting high-yield experimental results only for presentation and integration.

• AbstractPowered/geolip-*
  Canonical core models (encoders, decoders, scanners, analyzers, distillery, classifiers, diffusers) and a multitude of assistant modules. Anything shipped directly will have predominantly frozen and maintained codebases for guaranteed inference, with a secondary set of codebases meant to include experimental optimization research and speed gains.

• AbstractPowered/dataloaders-*
  Bucketed, any-size loaders with multi-stage interpretations and feature-space chaos augmentation using tools such as the geolip-svae scanners to detect divergence. Many of these dataloaders are intentionally meant for speed and optimization, oftentimes built to-task speed up training processes for in-house models and experiments.

• AbstractPowered/manifests
  Run manifests (config hash, vocab subset, expansions, bucket mix, metrics) for reproducibility in a concise and reviewable way.

• Demo Spaces (selected)
  Lightweight inference + manifest viewers for partners and reviewers. For specific models with clearly displayable behavior; such as diffusers, generators, composite structures, denoisers, and so on. Built with the public-eye in mind.

Release structural artifacts are kept small, composable, and ready for disposable retrains due to the rapid-learning process of surge training.

---

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.

Advanced signals and progressions

• Thousands of trains, 10s of thousands of finetunes and advanced trains, hundreds of thousands of inference attempts, and hundreds of thousands of prepared structures later; this system is more than robust. Only the strong survived.
• Each structure released on this repo will be a battle veteran, specifically chosen not because of some arbitrary guesswork or preliminary testing, but from thousands of failures that allowed these releases to survive.
• Genetically selected models guarantee the most powerful structured systems survive, each trained, interpolated, trained again, countless times for multiple systems through training processes devloped over thousands of train attempts and refined countless times.
• A 20 article paper trail leading from A to B to C every stage of the way, massive model training pipeline series of multiple systems, 2 years direct research shows the truth of these systems.

This is real, and it's up to you to invest or to discard the reality of what I'm building here.

---

Collaboration Invitations

• Research institutions: co-run ImageNet-class studies with bucketing, zoning, and corridor ablations; share ontologies and extend the processing conceptualization of the anchoring system or the svae battery system.
• Corporate labs: integrate domain dictionaries; trial rapid iteration pipelines; publish cost-per-accuracy analyses pre and post distillation mechanisms.
• 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, geocentric 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.

This is a very difficult process to explain without getting into specifics about LLM behavioral tuning. Research in this field has progressed a great deal since this process began, but the core principal still exists as an outstanding problem that I'm working to solve.

---

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.