In this thought-provoking piece, Lance Eliot delves into the recent research by Anthropic, exploring the inner workings of large language models (LLMs) and the potential discovery of a 'scratchpad' or working storage area. Eliot begins by sharing a personal anecdote about a stream and dam, using it as a metaphor for the exploration and discovery process. He then shifts gears to explain how LLMs work, using the example of completing a sentence. The author suggests that LLMs might have a scratchpad, a working storage area, which could be crucial to their functioning. Eliot then discusses the 'Working Storage Conjecture' and the need for specialized tools to find this area, introducing the concept of the Jacobian lens. Anthropic's exploration, using the J-lens, led to the discovery of J-space, an internal representation of potential words. This finding has significant implications, as it suggests that the working storage area impacts the LLM's output. Eliot then delves into the experiments conducted by Anthropic, including the example of Claude, their LLM, answering a question about the number of legs an animal that spins webs has. The author speculates on the potential connection between J-space and consciousness, drawing parallels to the global workspace theory in human cognition. However, he acknowledges the debate within the AI community and the reservations of neuroscientists, who caution against anthropomorphizing AI. Eliot concludes by emphasizing the importance of cautious exploration and the need to define consciousness before making definitive claims about AI's consciousness. Overall, this article provides a thoughtful and engaging analysis of the latest AI breakthroughs, offering a unique perspective on the potential inner workings of LLMs and the implications for AI consciousness.
