Rewiring AGI - Neuroscience is All You Need

We have published our perspective paper on AGI, and the limitations of deep learning. You can reach our paper from here.

Introduction

We published our paper as a pre-print under the title "Rewiring AGI—Neuroscience Is All You Need.". We argue that current deep learning models (including "reasoning" LLMs) proficient at pattern recognition but fundamentally limited in achieving true artificial general intelligence (AGI).

High Level Overview

The terms interpolation and extrapolation have been redefined within the machine learning community, and arguably lost some of their original meaning. In a nutshell they refer to the ability to generalize to new situations outside of previously encountered data.

We know these models can sometimes produce incorrect answers with confidence, and those mistakes might be magically fixed in a later release. While many factors contribute to these changes, the main point is that if a model hasn’t been trained on a specific input pattern, its responses can become unpredictable.

We say a model performs interpolation on a new input if:

1. The new input has a clear, well-defined local neighborhood within the model’s internal representation (e.g., in feature space or decision tree leaves).

2. The model’s outputs remain consistent and stable within that neighborhood.

3. The new input is not far outside the regions the model has learned from, rather, it is close to, or overlaps with, the areas (data manifold) covered by the training data.

Some people argue that current models generate novel ideas. However in reality, they are designed to predict the next token based on previous data. This is the way they are designed. And by design, these models optimized to extend input in the way most consistent with their training data. Well, that's called recalling.

Benchmarks

A true AGI would not need artificial "benchmarks" to demonstrate its abilities that are currently used. Even if benchmarks are used, they should be integrated with the ideas of Moravec's Paradox.

From our paper:

The reliability of current benchmarks is compromised by the uncertainty surrounding the training datasets of LLMs, raising concerns about potential data contamination. Furthermore, benchmarks are still in their early stages, requiring significant refinement to ensure they provide a robust and comprehensive evaluation of AGI capabilities.

Our Key Propositions & Thoughts

I'll summarise our propositions in this section.

Developing an AGI requires understanding how intelligence naturally emerges and operates, emphasizing adaptability, efficiency, and robust learning across contexts. Let's think about spiders (I find them fascinating) for a minute. Despite their tiny brains, they can survive in this 3D complex world. Also, if you take a spider and move it somewhere else, like 2 blocks away, they can still survive in that area too. This shows that spiders are able to generalise rather than rely on memorizing every detail (which our brain does not do either!) of their environment. In other words brains aren’t storing exhaustive information instead, they use general principles to navigate.

Have you ever seen a spider carrying GPUs on its back? :) Me neither. :)

To sum up, unlike narrow AI, AGI must function in:

1. Complex (like our 3D world) and

2. Changing environments

Insights from neuroscience should guide its development, where even simple organisms display remarkable abilities with limited resources. Key principles for AGI development should include but not limited to:

• Building on neuroscience foundations, prioritizing efficient structures over scale.

• Emphasizing robust generalization and adaptability rather than sheer complexity.

• Using practical, contamination-free benchmarks that reflect real-world adaptability.

• Mimicking the modular and hierarchical organization seen in biological systems.

• Enabling efficient learning from limited data and supporting abstraction for reasoning.

• Considering biologically inspired models like spiking neural networks for efficiency.

Overall, developing AGI is an iterative process focused on foundational mechanisms, adaptability, and biological inspiration. For example, current architectures are still missing several critical elements:

• Temporal dynamics, the (human) brain uses the timing of neural spikes.

• Data & Energy efficiency

• Real time adaptation by building on previously learned examples

Given these gaps, we argue that the first AGI systems will likely be seen as “dumb” by many people, since they will require time to learn and explore before reaching more advanced levels of intelligence.

Conclusion

So that’s the big picture of our perspective paper. Our current work focuses on building a model inspired by neuroscientific principles and we plan to publish our findings and ideas soon.