marcel blattner | May, 2025
Image created by marcel blattner. 1Mio. Eigenvalues of a Bohemian Matrix with base pattern: base_pattern=[-1j,0,1j]
Introduction
The current discussion about Artificial Intelligence - especially about AGI or Human Level AI - is often based on a fundamental, but possibly false assumption: that intelligence is substrate-independent and can therefore be arbitrarily transferred from one medium (like the biological brain) to another (like digital computers). This assumption is omitted in most discussions.
The Role of the Physical Substrate
Every information-processing system is fundamentally shaped by its physical substrate. This applies to both biological and artificial systems. The human brain, for example, has evolved over millions of years to perform information processing particularly efficiently. Furthermore, our brain possesses a remarkable combination of abilities. It processes information highly in parallel across billions of neurons, achieving extraordinary energy efficiency - consuming only about 20 watts, it can handle complex cognitive tasks for which modern computer systems require many times more energy. Particularly impressive is its robust pattern recognition and categorization, which allows the brain to function reliably even with incomplete or faulty information. This ability is complemented by an adaptive learning system capable of grasping and generalizing new concepts from just a few examples (generalization).
These properties are directly linked to the biological architecture of the brain and result from the specific way biological neurons function and communicate with each other.
Digital Systems: A Fundamentally Different Architecture
In contrast, digital computer systems are based on a completely different architecture:
- Sequential processing as a basic principle
- High energy consumption for complex calculations
- Precise numerical computations
- Necessity of large amounts of data for training
- In a computer, components like memory, CPU, etc., are logically and physically separate from each other. Not so in the brain. It's all in one.
These differences are not merely superficial but reflect fundamental properties of the respective substrates. While the biological brain is based on chemical and electrical processes that operate in parallel and with inherent fuzziness, computers are based on discrete, binary states and deterministic operations.
The Implications for AGI
This substrate dependence has far-reaching implications for AGI.
- Fundamental Limitations: Every substrate comes with its own limitations. The hope that we can overcome these simply through scaling might prove naive. There is growing evidence that scaling does not produce the emergent properties hoped for about a year ago.
- Different Strengths: Instead of trying to exactly replicate human intelligence, perhaps we should leverage the specific strengths of digital substrates to develop complementary forms of intelligence. Here, it is worthwhile to study the work of Michael Levin more closely.
- Energetic Aspects: The energy efficiency of biological systems suggests that certain forms of information processing might be fundamentally tied to biological substrates.
Empirical Evidence
The thesis of substrate dependence is supported by various empirical observations:
- Despite enormous progress in computing power, AI systems remain inferior to human intelligence in many areas, especially in tasks requiring generalization and transfer learning (applying abstract concepts to fundamentally different problem domains).
- The energy efficiency of biological systems has not yet been remotely achieved.
- Certain aspects of human cognition, such as consciousness or intuitive physics, have proven particularly difficult to replicate. And no, Sora does not have physical intuition.
New Research Directions
These findings suggest that we may need to fundamentally rethink our approach to developing artificial intelligence:
- Instead of directly imitating human intelligence, we could look for ways to better utilize the specific strengths of digital substrates.
- Developing new computer architectures that implement certain biological principles could be promising.
- Hybrid systems combining biological and digital components could represent an interesting middle ground.
The assumption of the substrate-independence of intelligence appears increasingly questionable in light of the available evidence. This does not mean that developing powerful AI systems is impossible, but it suggests that we may need to adjust our expectations and approaches. Instead of directly replicating human intelligence, the more promising path might lie in understanding and optimally utilizing the specific strengths of different substrates.
The future of AI research may lie not in overcoming substrate dependence, but in consciously considering and creatively utilizing it. This could lead to new, perhaps unexpected forms of intelligence that differ fundamentally from human intelligence but could be particularly valuable precisely because of that difference.
Remark on the Claim that OpenAI has Achieved AGI
Critics have raised several points against OpenAI's claims:
- Performance Limitations: François Chollet, co-developer of ARC-AGI, points out that o3 still fails at some very simple tasks, indicating fundamental differences from human intelligence.
- Skepticism Towards Benchmarks: David Rein warns that it is difficult to determine whether the ARC-AGI test truly measures the kind of general intelligence intended. He notes that many previous benchmarks claiming to measure fundamental aspects of intelligence were inadequate.
- Sustainability Concerns: The high-performance mode of o3 takes an average of 14 minutes and likely costs thousands of dollars per task, raising questions about the model's practicality and environmental impact.
- Lack of Grounding: Critics argue that o3's reliance on natural language instructions instead of executable symbolic programs means it cannot establish direct contact with reality through execution and evaluation.
- Dependence on Human Data: The system relies on expert-labeled, human-generated chain-of-thought data, limiting its ability to autonomously acquire new skills.
- Unclear Generalization: While o3 has shown impressive results on reasoning tasks, it is unclear how well these capabilities transfer.
These criticisms highlight that while o3 represents a significant advancement, substantial hurdles remain to be overcome.