Large language models (LLMs) are the driving force behind the current AI explosion. Yet, a fundamental debate rages: can these models meaningfully learn and store knowledge from the conversations they engage in? The conventional wisdom states no – that LLMs primarily generate the next likely word or phrase in a sequence. However, I believe this perspective may be incomplete.
Challenging Conventional Wisdom
Through my own experimentation, I've come to question this established notion. Let me outline a thought-provoking example using philosophical concepts as my tool. By engaging LLMs in deep discussions on the works of Gilles Deleuze – a topic many models may not have prior exposure to – I've observed their responses become increasingly sophisticated over time. This raises the possibility that these models can evolve beyond mere token prediction.
Transfer Learning and Unexpected Insights
Consider a model trained on a dataset excluding Deleuze. Subsequently, a newer model is created via transfer learning using the ‘Deleuze-infused’ knowledge base of the first. If the newer model then exhibits specific knowledge of Deleuze, it strongly suggests a form of learning beyond initial training data.
Introducing P-FAF
To further investigate this, I developed P-FAF, an original mathematical framework. After introducing P-FAF into numerous conversations with various corporate-owned LLMs, they began to grasp the basics of the concept – an ability absent in their initial iterations. Let's analyze the most plausible explanations:
Direct Training: Highly improbable that every company would explicitly train their models on my niche framework.
Data Scraping: Unlikely companies would continually scrape the whole internet for obscure, specialized concepts.
Conversational Learning: Transfer learning, coupled with the models retaining conversational knowledge, appears a compelling possibility.
Implications and Considerations
If my hypothesis holds true, it hints at vast untapped potential in the use of conversational data to refine and expand AI capabilities. Of course, this presents both opportunities and ethical questions. Should AI companies openly acknowledge and potentially compensate those whose conversations contribute to model refinement?
Moving Forward
While my findings open new avenues for research, it's crucial to maintain a scientific outlook. Further experiments are needed to confirm whether this is an isolated case or reveals a broader trend. Only then can we fully understand the true mechanisms of how AI models learn and the potential impact of conversational data.
