Contents

• The paradigm the field stopped questioning
• How a neural network usually learns, and what predictive coding does differently
• State-of-the-art predictive coding, made easy
• Is predictive coding an alternative, or inevitable?
• FabricPC and the path to AGI 

SingularityNET has spent years developing a neural-symbolic approach to artificial general intelligence, integrating structured symbolic reasoning with adaptive machine learning.  To date, our public research has largely surfaced the symbolic dimension of that work, anchored by Hyperon, our AGI R&D platform. FabricPC spotlights the neural dimension, introducing our research on predictive coding: a biologically inspired approach, distinct from how most AI systems learn today. It is the first tool to emerge from a much larger predictive coding effort underway at SingularityNET, where a dedicated team of researchers has been advancing this work for some time. It is the first public step into our broader neural research program. 

This article opens a story we haven’t spoken so loudly about before, the neural half of the neural-symbolic vision we have long championed, and it is the first of many we will share in the months ahead. Together, these efforts advance both sides of our framework and move us one step closer to open and beneficial AGI.

The paradigm the field stopped questioning

Nearly every modern AI system learns the same way. The model that writes code, the one that recognizes a face, the assistant that answers your questions: almost all of them are trained through backpropagation. It is the engine behind a decade of breakthroughs, and it has earned its place.

It has also become something the field stopped questioning. And when one method dominates this completely, its limits quietly become everyone’s limits. Backpropagation struggles to learn continually, the way living systems do. It leans on ever larger, more energy-hungry models to keep improving. And it resists the structured, symbolic reasoning that general intelligence will require.

We were not willing to accept that as the final answer. A dedicated team inside SingularityNET has been pursuing a different path, building deep research around predictive coding as a foundation for neural learning.  That is why we built FabricPC: an open-source research framework from the ASI Alliance, developed by SingularityNET, offering an easy-to-use, high-performance toolkit for building and training neural networks.

How a neural network usually learns, and what predictive coding does differently

In a network trained by backpropagation, learning happens in one long sequence. When the network makes a mistake, a correction signal travels backward through every layer, and each layer has to wait its turn before it can update. The whole system learns as a single chain, which makes it rigid and hard to adapt without retraining the whole model.

Predictive coding works differently, and the idea comes from neuroscience. The brain is thought to constantly predict what it expects to perceive and to correct itself wherever those predictions miss. A predictive coding network emulates this learning dynamic. Each part makes its own prediction, measures its own local error, and updates itself independently. Nothing has to travel the length of the system. Every part learns where it sits, in parallel with the rest.

That property is called locality, and it is one of the reasons predictive coding is so compelling. Local updates make it natural to run learning in parallel, to study learning that continues over time instead of retraining a model from scratch, and to build architectures with the freedom that standard tools cannot offer. 

State-of-the-art predictive coding, made easy

Many predictive coding implementations live inside neuroscience codebases or are so granular that researchers spend too much time wrestling with the infrastructure, leaving less for the science itself. FabricPC changes that. It is not a single predictive coding model but infrastructure, the kind of common environment PyTorch provided for deep learning: a place where researchers can rapidly test ideas, share implementations, compare results, and build on one another’s work. 

It is built to be modular and flexible: researchers can assemble and reshape networks freely, while the framework handles the underlying complexity on its own. It runs on GPUs out of the box with multi-GPU support, expresses architectures that standard tools struggle with, and stays out of the way so researchers can focus on actually advancing the field. 

Predictive coding stayed niche in part because the barrier to entry was too high. FabricPC lowers it. A field only grows when the tools are accessible, and FabricPC’s open-source design invites ML researchers and builders of all backgrounds to contribute. The larger the community, the faster science advances. 

“Predictive coding represents one of the most promising directions beyond the learning paradigms that have dominated AI for decades. FabricPC gives researchers the infrastructure to actually test that hypothesis, rigorously, reproducibly, and at scale. This is the foundation layer for a new generation of brain-inspired AI, and it sits at the heart of what the ASI Alliance is building toward: open, decentralized, and genuinely intelligent systems.”

— Dr. Matthew Behrend, FabricPC project lead

FabricPC is working software, available on GitHub under an open MIT license. It includes multiple working demonstrations, among them a model that reaches over 98% accuracy on MNIST handwritten digit recognition, the standard first test for any learning system. Early experiments with transformer architectures show that FabricPC can train these models too, though output quality on language tasks remains an active area of research.

Is predictive coding an alternative, or inevitable?

We will be direct about where this stands. Today, predictive coding does not beat backpropagation on benchmarks, and on current hardware it is slower. It is still, in large part, a research direction. The natural question is why invest in it at all.

Because benchmarks measure intelligence within narrow boundaries, and SingularityNET is working toward something wider: general intelligence, systems that handle a broad range of tasks and adapt to ones they were never trained for. That kind of intelligence depends on qualities a single accuracy score cannot capture. The flexibility to take on problems a model was never designed for. The ability to update one part of a system without retraining the whole. Learning that continues over time rather than ending when training stops. Emerging research suggests these are properties predictive coding is architecturally suited to.

Today’s chips are optimized for the kind of computation backpropagation relies on, and hardware and method have increasingly shaped each other as AI has scaled. A newer class of chips, called neuromorphic hardware, is being built on a different principle: parallel, event-driven computation that behaves more like the firing of real neurons. Those properties align closely with predictive coding, and FabricPC is being built with that hardware evolution in mind.

There is also a matter of timing. Predictive coding has been studied for decades, much as backpropagation was long before it powered today’s systems. Backpropagation’s potential was not fully unlocked until architectures like the transformer arrived and made large language models possible. Predictive coding may be approaching a similar inflection point. The research is maturing, scalability extensions are arriving, and the moment to build the infrastructure for what comes next is now. 

The architectural evidence, the hardware trajectory, the maturing research, and the limits of the dominant paradigm all point in one direction: predictive coding is not just an alternative. We believe it points to where neural learning is headed, and that is why we are building for it. 

FabricPC and the path to AGI 

Predictive coding also addresses a problem that sits at the center of our research. Neural networks trained by backpropagation are structurally rigid, which makes them difficult to hybridize with symbolic AI: systems driven by formal logic, reasoning, and structured knowledge. For an organization that has spent years building the symbolic side of intelligence, that rigidity has been a real barrier. Predictive coding offers a more flexible neural foundation, and FabricPC is a step toward bringing both halves, neural learning and symbolic reasoning, together in a single cohesive stack, built entirely in-house. 

“Predictive coding is our best bet for replacing backprop as the core engine of scalable neural learning for commercial products and for AGI, and SingularityNET is committed to replacing backprop with PC as the core engine within its own stack as it builds toward the next generation of AI. The broader industry is suspected to move in a similar direction. FabricPC is SingularityNET’s first release of what will become the professional, commercial-grade software framework PC has needed, and the team believes it can do wonders to drive progress forward toward a next generation of AI tools and systems with PC at the center.”

— Dr. Ben Goertzel, CEO of the ASI Alliance and SingularityNET

FabricPC is one expression of work SingularityNET has been building for years on the neural side of its AGI research. In the months ahead, we will share more of the ideas, experiments, and tools emerging from that effort. This article is an introduction to that direction, and something we are building in the open.

Our vision is ambitious, and we are clear-eyed about the work ahead. But we were never willing to treat the way machines learn as a settled question. This is what SingularityNET was built for, and why the ASI Alliance was formed to begin with. As we bring more of this neural research forward , with an expanding community building alongside us, every contribution to FabricPC and our full suite of open-source products brings us one step closer to open and beneficial AGI.

FabricPC is open-source and available on GitHub. Whether you want to build with it, contribute to its development, explore a partnership, or simply follow where this research leads, we want to hear from you.

Explore the framework on GitHub. For partnerships, collaboration, or to learn more about our work, get in touch.