What If We Got Our AI Development Strategy Wrong?
Current AI research and development are focused on building increasingly complex architectures, such as transformers, massive neural networks with billions of parameters, and supercomputers to support them. But what if this path doesn’t lead us to true intelligence, but only to more powerful computational systems? Perhaps we should pause and ask: is this strategy really the right one?
Get Eugene’s stories in your inbox
Join Medium for free to get updates from this writer.
Current Strategy: Complex Models, Big Data
Today’s AI leaders, like OpenAI and DeepMind, develop systems that rely on enormous amounts of data and computational power. Models like GPT-4 are built on intricate transformer architectures that require substantial resources for training and deployment. These models have achieved impressive results, but they also face several issues:
- High computational demands: Such systems require vast resources, which limits their accessibility and raises questions about how sustainable this strategy is.
- Lack of adaptive learning: Transformers and deep networks lack built-in mechanisms for dynamic experience accumulation, unlike the biological brain.
An Alternative: Simplicity, Scalability, and Adaptation
Could there be a different path, one more akin to biological systems? Inspired by how the brain works, we can propose alternative principles:
- Horizontal scalability: The brain uses parallel processes, simplicity, and decentralized structures. What if we focused on creating systems that can scale horizontally using simple components instead of increasing complexity?
- Dynamic learning and experience accumulation: The biological brain learns continuously, updating synaptic connections as new information is received. Could a system based on these principles become more adaptive without requiring full retraining on all data each time?
Example of a New Approach
One possible approach is a system that learns from short sequences, where dynamic experience accumulation happens in real-time, and scalability is achieved by adding new levels of abstraction as needed, similar to how the brain functions.
Conclusion
Could it be that our current direction in AI development is leading us toward a dead end, where everything becomes more complex but doesn’t bring us closer to true thinking? It’s time to consider other strategies that might prove more efficient and closer to natural learning and thinking processes.
