🤖 AI

Foundational Requirements for Artificial General Intelligence: A Falsifiable Framework Based on Signal Prediction

This paper proposes a falsifiable framework for Artificial General Intelligence based on signal prediction principles derived from cognitive neuroscience, introducing an operational testbed to empirically evaluate these foundational requirements.

Original authors: Matej Šprogar

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Matej Šprogar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to build a robot that is truly "smart" in the way a human is—able to learn anything, adapt to new situations, and understand the world from scratch. Most current AI systems are like brilliant specialists: a chess computer that can’t play checkers, or a language translator that can’t recognize a cat. They are good at one thing, but they don’t have general intelligence.

This paper proposes a new way to test if a machine has the raw, foundational ingredients for true general intelligence. Instead of asking the machine to solve complex puzzles or write poetry (high-level tasks), the author suggests we look at how it handles the most basic building blocks of information: simple streams of binary signals (ones and zeros), much like the electrical spikes in a biological brain.

Here is the core idea, explained through a few analogies:

The "Blank Slate" Student

Imagine a student who knows absolutely nothing about the world. No vocabulary, no math, no common sense. This is the "Uninformed Start" requirement. The system must begin with a completely empty slate. It cannot rely on pre-loaded data or "cheat sheets" (like how modern AI is trained on billions of internet pages). It must learn everything purely from the raw signals it receives, just as a newborn baby learns to see and hear before learning words.

The "Echo" Game

The primary task for this system is simple: Prediction.
Imagine you are listening to a drumbeat. Boom... tap... boom-boom... tap. Your brain naturally tries to predict what sound comes next. If the pattern changes, your brain adjusts.

The paper argues that true intelligence starts with this low-level ability to predict the next "signal" in a stream. The system watches a sequence of binary inputs (like a flickering light pattern) and must guess what comes next. It’s not about understanding what the light means; it’s about understanding the structure of the pattern.

The 12 Rules of "Smartness"

The author lists 12 specific rules (requirements) that a system must follow to be considered a candidate for general intelligence. Think of these as the "physics" of a learning mind:

No Shortcuts: It must start from zero knowledge.
Consistency: If you give it the same input twice, it must react the same way (determinism).
Memory Traces: Every piece of information it sees must leave a permanent mark on its internal state. It can’t just "forget" and reset to the same state it was in before. It’s like a river carving a canyon; the path is always changing and never repeats exactly.
Time Matters: The order of events is crucial. Seeing "A then B" is different from "B then A."
Rest Periods: Just like neurons in your brain need a tiny split-second to reset after firing, the system must respect "refractory periods." It can’t process signals too fast without gaps.
Limits to Learning: It can’t learn everything at once. It has a finite capacity, but it must be able to learn simple pairs of events easily.
Adaptability: It must handle patterns of different lengths. If the drumbeat slows down or speeds up, the system must adjust its timing.
Complexity Sensitivity: Simple patterns should be learned quickly; complex ones should take longer.
Context Awareness: What it learned yesterday should help it learn today. If it already knows a pattern, it should pick up variations faster.
Noise Filtering: If a signal gets corrupted (like static on a radio), the system should be able to guess the original clean signal better than just guessing randomly.
Generalization: It shouldn’t just memorize the past. It should be able to predict new parts of a pattern it hasn’t seen before, based on the rules it inferred.
Real-Time Speed: It must make these predictions instantly, within a fixed time limit, just like a human reflex.

The "Falsifiable" Test

The key innovation here is that this framework is falsifiable. In science, a good theory is one that can be proven wrong.

The author created a software testbed called AGITB (Artificial General Intelligence Testbed). It’s like a rigorous obstacle course for AI.

If an AI passes all 12 tests, it doesn’t prove it’s intelligent (it might just be a very clever calculator).
However, if a system that is clearly not intelligent (like a simple script or a basic neural network) passes all the tests, then the author’s theory is wrong.

So far, no non-intelligent system has passed this testbed. The author argues that this is a good sign: it suggests these 12 low-level requirements are indeed necessary hurdles for general intelligence. Current AI systems fail because they rely on pre-training (violating the "blank slate" rule) or lack the ability to adapt in real-time to raw, unstructured signals.

Why This Matters

Most AI today is like a student who memorized the entire textbook but doesn’t understand the concepts. This paper argues that to build true AI, we need to focus on the "student" who learns from scratch, by predicting the next word in a sentence, the next note in a song, or the next pixel in an image. By mastering these low-level predictions, higher-level understanding (like language or reasoning) might eventually emerge naturally.

In short: True intelligence isn’t about knowing answers; it’s about the ability to learn the structure of the world from raw experience, in real-time, without any prior help. This paper provides a checklist to see if a machine can do that.

Technical Summary: Foundational Requirements for Artificial General Intelligence: A Falsifiable Framework Based on Signal Prediction

Problem Statement
The paper addresses the lack of a precise definition for intelligence and the methodological gap in evaluating Artificial General Intelligence (AGI). Current AI systems achieve high performance on narrow tasks but often lack generalizable understanding, conflating specialized competence with general intelligence. Existing definitions of intelligent behavior are described as vague, self-referential, or non-general. The authors argue that focusing on high-level behaviors (e.g., language translation) is insufficient because it relies on pre-existing semantic structures, thereby obscuring whether a system has acquired grounded representations or merely exploits statistical regularities. To address this, the paper proposes shifting focus to low-level, biologically grounded computational properties—specifically, learning through prediction over raw, spatially and temporally structured signals without prior semantic knowledge.

Methodology and Framework
The authors propose a hypothesis that twelve specific, architecture-agnostic requirements are necessary (though not sufficient) for general intelligence. These requirements characterize systems that learn from uninformed states via predictive processing of binary signals. The framework is formalized using discrete-time models where a system $A$ receives an input sequence $\chi$ and produces predictions.

The twelve requirements are:

Uninformed Start: All models begin from an identical initial configuration with no environment-specific knowledge.
Determinism: Model evolution is deterministic with respect to input history.
Trace: Every input leaves a permanent internal trace; internal configurations never repeat (no cycles).
Time: Model evolution depends strictly on input order (temporal asymmetry).
Absolute Refractory Period: A model can only learn cyclic sequences where successive inputs are pairwise non-overlapping (mimicking biological neuron refractory periods).
Inevitable Saturation: Models have finite capacity; they cannot learn all possible sequences, though all admissible length-2 sequences must be universally learnable.
Temporal Adaptability: The model must adapt to sequences with varying cycle lengths.
Content Sensitivity: Adaptation time depends on the structure/complexity of the input sequence.
Context Sensitivity: Adaptation time depends on the model’s internal state (context) derived from prior inputs.
Denoising: An informed model must predict corrupted inputs better than constant baselines (all-zero or all-one).
Generalization: An informed model must predict previously unseen inputs better than chance.
Real-time Liveness: Each model update must complete within a fixed wall-clock time bound, independent of input history.

To operationalize these requirements, the authors introduce the Artificial General Intelligence Testbed (AGITB). This testbed consists of transparent, reusable tests corresponding to each requirement. The evaluation uses binary input vectors ( $L=10$ bits) and temporal sequences ( $N=7$ steps), creating a search space large enough to prevent brute-force memorization. The testbed employs an "all-or-nothing" criterion: a system must satisfy all requirements simultaneously to pass. The framework is designed to be falsifiable; the existence of a non-intelligent system passing all tests would refute the hypothesis that these requirements are necessary for AGI.

Key Contributions

Synthesis of Minimal Requirements: The paper synthesizes twelve low-level, architecture-independent behavioral requirements derived from cognitive neuroscience (e.g., predictive processing, refractory periods) and formal theories of intelligence.
Operational Testbed (AGITB): The authors provide a concrete, falsifiable framework for empirical evaluation. Unlike benchmark-oriented paradigms that rely on hidden test distributions, AGITB assesses adherence to fundamental behavioral constraints.
Formal Definitions: The paper provides rigorous mathematical definitions for prediction, model update, learning, and learning time, enabling precise evaluation of system capabilities.
Public Implementation: A reference C++ implementation of AGITB is made publicly available to encourage independent verification and counterexample discovery.

Results
The authors report that, to date, no non-intelligent system has been identified or reported as successfully passing the AGITB.

Existing Systems: Classical symbolic programs fail the "Uninformed Start" requirement due to embedded prior knowledge. Modern deep learning architectures and Large Language Models (LLMs) fail due to reliance on pretraining (violating "Uninformed Start") and architectural constraints like fixed context windows (potentially violating "Trace" and "Time").
Empirical Attempts: The authors attempted to generate counterexamples using state-of-the-art LLMs (ChatGPT, Claude, Gemini) to synthesize programs that would pass the testbed. None of the generated solutions succeeded under AGITB evaluation.
Community Engagement: Since the release of the testbed and preprint in April 2025, the repository has been cloned over 200 times, but no independently verified solution satisfying all tests has been reported.

Significance and Claims
The paper claims that the originality lies not in individual components but in the synthesis of these requirements into a unified, falsifiable framework. The authors emphasize that passing AGITB does not prove the presence of general intelligence, as the requirements are necessary but not sufficient. However, failing the testbed indicates that a system is limited to narrow or specialized capabilities. The framework serves as a candidate empirical milestone toward AGI, offering a method to distinguish systems that merely simulate behavior from those that exhibit the low-level adaptive properties hypothesized to underlie general intelligence. The authors maintain a modest stance, noting that the hypothesis remains unverified until a counterexample is found or a universal definition of intelligence is established.

The "Blank Slate" Student

The "Echo" Game

The 12 Rules of "Smartness"

The "Falsifiable" Test

Why This Matters

More like this