The 7 Deadly Sins of AGI Design
How AI Lost Its Way
Based on the article by Peter Voss
Published on November 27, 2024
The term 'AGI' has been distorted. Real AGI should learn like a smart college graduate, yet the field is dominated by a flawed, brute-force approach. A detailed analysis exposes seven key mistakes preventing meaningful progress.
The 7 Deadly Sins
1. Focus on Knowledge
The misconception that having knowledge is a measure of intelligence, rather than the ability to dynamically acquire and generalize it.
2. Narrow AI
The focus on specific abilities (e.g., chess) rather than a general intelligence applicable to a wide, dynamic range of problems.
3. External Intelligence
The reliance on human programmers to design and tune systems, meaning the intelligence is external to the AI, not autonomous.
4. Pre-Trained
The limitation where models are trained once and cannot be updated incrementally in real-time, making them static and disposable.
5. Quantity
The focus on massive quantities of training data, including low-quality information, instead of prioritizing data quality for robust intelligence.
6. Statistical
Basing models on statistical regularities rather than ontological, real-world features, leading to hallucinations and ungrounded knowledge.
7. Lack of Theory
The overarching sin of lacking a foundational theory of intelligence, opting for a brute-force 'data and compute' approach instead.
AI vs. Human Brain: A Resource Mismatch
Current AI Models
Gigawatts of Power
10s of Trillions of Words
Human Brain
~20 Watts of Power
A Few Million of Words
A Smarter Path Forward
Third Wave of AI
An approach, also called Cognitive AI, that requires orders of magnitude less data and compute and has the ability to learn incrementally in real-time.
INSA
The Integrated Neuro-Symbolic Architecture, developed by Aigo.ai, is presented as a concrete example of a 'Third Wave' path to AGI.
Real AGI
The goal should be systems that meet the criteria of autonomous, human-level learning with limited resources, not just pattern matching at scale.
Key Players & Examples
IBM Deep Blue
An early AI chess champion cited as a prime example of Narrow AI—it couldn't even play checkers.
Google's AlphaFold
A powerful protein-folding predictor, but still a manually engineered, narrow system for a specific task.
DARPA
The agency that coined the term 'Third Wave of AI' to describe the need for new, contextual AI approaches.
Aigo.ai
A company developing a Cognitive AI service based on their INSA architecture, representing the 'Third Wave' approach.
Frequently Asked Questions
What is the core difference between statistical and ontological concepts in AI?
Statistical concepts are based on word co-occurrence from training data, while ontological concepts are grounded in real-world features and attributes, allowing for a deeper, more detailed understanding.
What is the financial implication of pre-trained, disposable AI models?
They are incredibly expensive, costing hundreds of millions of dollars to train, and are inherently disposable, leading to a poor return on investment as they depreciate over just a few months.
What fields of study are essential for developing a proper theory of AGI?
The author suggests deeply exploring epistemology (theory of knowledge), cognitive psychology, philosophy of consciousness, volition, and ethics to build a foundational theory of intelligence.