PART 21 / Frontier

AGI connections: evidence before promises

Which biological ideas yield testable research hypotheses about broader capabilities?

Guided illustrated lesson

What you will learn

Evaluate every proposed connection with evidence, limitations, counterarguments, and a falsifiable experiment.

Before you begin: Cognitive architectures

START WITH A QUESTION

Picture this.

A system improves from 75% success to 80%. That is interesting. But what improved, on which situations, and why? A careful answer can be more valuable than a much larger claim about intelligence.

Generalization
Performance beyond the examples or conditions used to fit and tune a system.
Percentage point
The absolute difference between two percentages.
Inductive bias
A modeling preference that makes some solutions easier to represent or learn than others.

Read for the idea first. Work through the numbers next. Try answering the questions before opening the explanations.

Turn inspiration into a question you can test

Biological systems motivate ideas about sparse computation, memory, embodiment and learning from limited experience. Translate one idea into a measurable hypothesis. For example: under a fixed training budget, a particular connectivity prior improves success on held-out navigation layouts relative to a degree-matched alternative. That sentence specifies a mechanism candidate, a comparison and a setting. 'This architecture is closer to the brain' does not yet specify a measurable advantage.

Capability, transfer and mechanism are different claims

Success on one task demonstrates a capability under those test conditions. Transfer asks how performance changes on new tasks or conditions. A mechanism claim asks what caused the improvement. These need different experiments. Copying some wiring does not copy development, physiology, learning history, body or environment. Even a convincing model of one animal behavior would not settle a general definition of intelligence. Keep the scope of the conclusion aligned with the scope of the evidence.

Read percentages carefully

The difference between 80% and 75% is 5 percentage points. Relative to the 75% baseline, it is about 6.7% improvement in that success rate. Those statements use different denominators. Neither says how many trials were run, how variable the result is, or whether the candidate had privileged input. A five-point result from 20 trials and one from thousands of comparable trials deserve different uncertainty analyses. Always retain counts and protocol beside percentages.

Build a small ladder of generalization

Begin with held-out examples from the same setting. Next change a meaningful factor such as noise, layout or delay while holding the scoring rule fixed. Then test a different task only if the claim requires it. Describe where performance degrades rather than averaging away the weak cases. A simpler baseline that matches the candidate can teach you which complexity was unnecessary. None of these individual tests is a certificate of AGI, but each can make a narrower claim more informative.

SEE THE RELATIONSHIPS

From structure to behavior.

Evidence: a labelled conceptual diagram
Figure 21.1 · Evidence. Original educational diagram; synthetic, not a measured result.
The reasoning pathBiological idea then Falsifiable hypothesis then Controlled experiment then Limited conclusion1. Biological idea2. Falsifiable hypothesis3. Controlled experiment4. Limited conclusion
A learning flow for this part. Follow the assumptions through to an observable result.

MAKE THE IDEA PRECISE

The mathematics.

difference_pp = 100 (p_candidate - p_baseline)
Symbols and units
SymbolMeaning
p_candidate, p_baselineMeasured success fractions under matched evaluation conditions
difference_ppDifference in percentage points, not relative percent improvement

Build the equation, step by step

  1. Express both success rates as proportions on the same evaluation protocol, then subtract baseline from candidate.
  2. Multiply by 100 to report a percentage-point difference. This differs from relative percentage change, which divides the difference by the baseline.
  3. An observed difference needs sample sizes, uncertainty, distribution details and a valid comparison. One metric improvement does not establish general intelligence or identify the mechanism that caused it.

WORK THROUGH THE NUMBERS

Worked example

Success fractions 0.80 and 0.75 differ by 5 percentage points. Without sample counts, uncertainty and comparable conditions, this arithmetic alone does not establish a reliable advantage. It says nothing about AGI.

A small Python experiment.

Python 3 · Standard library only · Synthetic educational example. Run this in your own notebook or a local file.

candidate, baseline = 0.80, 0.75
print(f"difference = {100*(candidate-baseline):.1f} percentage points")

Read the code as a story

The code subtracts two proportions and multiplies by 100. It reports percentage points. Dividing the difference by baseline before multiplying by 100 would report relative percentage change instead. This arithmetic is descriptive; it contains no uncertainty calculation and no test of why the values differ.

Expected output

difference = 5.0 percentage points

Change one quantity, predict the result, and then run the code again. Explain any difference before adding complexity.

Test your understanding.

Try it: Write an alternative explanation for a five-point advantage and a control that could test it.

Self-check: What is the difference between improving from 50% to 55% and improving by 5%?

The first is a gain of 5 percentage points and a relative gain of 10%. A 5% relative gain from 50% would produce 52.5%. Report both the metric definition and denominator to avoid ambiguous claims.

Where this model stops: The book presents no proven route from connectomes or SNNs to AGI.

A research question: What result would cause you to reject your preferred architecture hypothesis?

Record evidence in your mastery worksheet →

PREDICT / TRY / EXPLAIN

Rewrite an overclaim

Use paper, a calculator or the Python example. Write your prediction before checking the result.

  1. Start with the sentence: 'Our brain-inspired model is more intelligent because it scored 80% instead of 75%.'
  2. Replace it with a claim that names a task, denominator, comparison and held-out conditions. Leave unknown facts as requirements, not invented measurements.
  3. List two alternative explanations and one controlled comparison for each.
Compare with a worked solution

A defensible template is: 'On [predefined held-out task], the candidate achieved [successful trials / total trials], compared with [baseline counts], under [matched budget and inputs].' More training and better sensory input are two alternative explanations. Match training interactions for one control and match observations for another. If counts are unavailable, say the percentages alone are insufficient.

CLOSE YOUR NOTES

Can you explain it back?

These questions test the reasoning, not just a remembered phrase. Say your answer out loud or write a sentence first.

01What is the relative improvement from 75% to 80%?

Reveal the reasoning for question 1

(80−75)/75 × 100 ≈ 6.7%. The absolute gain remains 5 percentage points.

02Does success on a new layout prove the architecture caused the gain?

Reveal the reasoning for question 2

No. It supports a limited generalization claim. Isolating the architecture requires appropriate controlled comparisons, because training or input differences could explain the gain.

TAKE THE IDEA WITH YOU

Three things to remember.

  • Convert a broad analogy into a specific falsifiable comparison.
  • Separate capability, generalization and mechanism claims.
  • Report denominators and uncertainty alongside percentage changes.

Next we design a comparison that could actually distinguish the preferred explanation from its alternatives.

Continue to Part 22 →

THE NEXT LAYER

Chapter depth map.

This introduction provides a first model. The following chapters are outlines for deeper future editions; their full treatments have not yet been published.

Candidate lessons

21.1 Learning and embodiment · Outline
  1. 21.1.1 Sensorimotor grounding and embodiment
  2. 21.1.2 Continual and developmental learning
  3. 21.1.3 Plasticity, local learning and consolidation
  4. 21.1.4 Multi-timescale adaptation
  5. 21.1.5 Why task-specific success does not establish general intelligence
21.2 Computation and integration · Outline
  1. 21.2.1 Energy efficiency and sparse computation
  2. 21.2.2 Modularity and specialized circuits
  3. 21.2.3 Hierarchical control and global integration
  4. 21.2.4 Predictive processing and world models
  5. 21.2.5 Architecture versus training and environment

A claim-by-claim research discipline

21.3 Evaluate an AGI hypothesis · Outline
  1. 21.3.1 Label established fact, practice, research, hypothesis or speculation
  2. 21.3.2 State supporting evidence and its scope
  3. 21.3.3 Identify current limitations and counterarguments
  4. 21.3.4 Design tests and disconfirming outcomes
  5. 21.3.5 Separate operational capability from consciousness claims
21.4 Limits of copying biology · Outline
  1. 21.4.1 Connectomes omit dynamics, history and bodily context
  2. 21.4.2 Structural similarity does not prove functional equivalence
  3. 21.4.3 Biological mechanisms may not be optimal for silicon
  4. 21.4.4 Compare conventional baselines fairly
  5. 21.4.5 Unsupported AGI claims remain unsupported

Read with the original sources.

These readings support the concepts and deeper investigation. The worked numbers and illustrations above are teaching examples.

Complete source register →

Where next?