Spurious Sparks of AGI 💥
On the Unsurprising Finding of Patterns in Latent Spaces
We humans are prone to seek patterns everywhere. Meaningful patterns have often proven to help us make sense of our past, navigate our presence and predict the future. Our society is so invested in finding patterns that today it seems we are more willing than ever to outsource this task to an Artificial Intelligence (AI): an omniscient oracle that leads us down the right path.
Unfortunately, history has shown time and again that patterns are double-edged swords: if we attribute the wrong meaning to them, they may lead us nowhere at all, or worse, they may lead us down dark roads. I think that the current debate around large language models (LLMs) is a prime example of this.
This article was the original starting point for our recent paper on the topic co-authored with Andrew M. Demetriou, Antony Bartlett and Cynthia C. S. Liem. The paper is a more formal and detailed treatment of the topic and is available here. While this blog post focuses in particular on practical examples of finding patterns in latent spaces, the paper includes a detailed review of social science findings that underline how prone humans are to be enticed by patterns that are not really there.
Models are Tools, Treat Them as Such
In statistics, misleading patterns are referred to as spurious relationships: purely associational relationships between two or more variables that are…