AI’s Dark Secret: We Don’t Know Why It Works
The researchers behind the world’s most advanced AI systems have a confession to make: they don’t fully understand their own creations.
The people who build today’s most capable artificial intelligence can describe exactly how they train it. They can write down the architecture, the optimisation procedure and the objective the system is rewarded for meeting. What they cannot do, in any comprehensive way, is explain why it all works together in the first place. This lack of understanding is not a bug, but a feature of the complex systems that now underpin much of modern technology.
The rise of deep learning has given us AI models that can outperform humans in many tasks, from image recognition to playing complex games like Go. These models are built using neural networks, which are designed to mimic the way the human brain works. However, as the complexity of these networks grows, they become increasingly difficult to understand and interpret.
“We’re not sure why certain architectures work better than others,” says **Andrew Ng**, a pioneer in the field of AI and former head of AI at Google. “We’ve tried to figure it out, but it’s still a bit of a mystery.” Ng’s comment reflects the general sentiment among researchers in the field, who are often forced to rely on intuition and trial-and-error when designing and training AI systems.
A Black Box of Code
The problem is particularly pronounced when it comes to the use of large language models, which have revolutionised fields like natural language processing and text generation. These models are trained on vast amounts of data, which they use to learn patterns and relationships that allow them to generate human-like text. However, the exact mechanisms behind this process are still not well understood.
What this means is that much of modern technology now rests on tools that we can use far better than we can understand. This is not necessarily a bad thing – after all, we don’t need to understand the inner workings of a car in order to use it. However, it does raise important questions about accountability and transparency in AI development.
The Future of AI Research
As researchers continue to push the boundaries of what is possible with AI, it’s likely that we’ll see even more complex and opaque systems emerge. However, this also creates an opportunity for researchers to develop new techniques and tools that can help us better understand and interpret complex AI systems.
One possible solution is to develop more transparent and interpretable AI models, which can provide insights into how they arrive at their decisions. Another approach is to use techniques like explainability and feature attribution, which can help us understand why certain models make certain predictions or decisions.
Ultimately, the question of why AI works is not just a technical one, but also a philosophical one. As we continue to rely more and more on AI in our daily lives, it’s essential that we have a deeper understanding of how it works and what it means for our society.
