In this thought-provoking article, the author delves into the recent research by Anthropic, exploring the inner workings of modern LLMs and the possibility of AI consciousness. The piece begins with a light-hearted personal story about the value of exploration, setting the tone for the discussion to follow. It then shifts gears to explain how contemporary generative AI and LLMs function, using the example of completing a sentence with a single word. The author introduces the concept of a scratchpad or working storage area within LLMs, which is thought to influence the AI's output. This leads to the exploration conducted by Anthropic, who built specialized tools to find this potential scratchpad, referred to as J-space. The J-lens, a mathematical probing tool, was used to identify internal representations associated with potential words, or tokens. The findings suggest that the working storage area has a significant impact on the AI's downstream processes, raising questions about its role in AI consciousness. The author reflects on the debate within the AI community regarding consciousness, drawing parallels to the global workspace theory in human neuroscience. While the research is exciting, the author remains cautious, emphasizing the need for further exploration and a cautious approach when drawing parallels to the human mind. The article concludes with a call for mindful progress in AI research, prioritizing the well-being of humanity and our self-preservation.
