If you take the bot, and use it to generate output the size of its training set, and then use that output to train a new bot, it is not as effective as the first bot, trained on human output.
This factor being greater than 1.0 is the criteria I would use before busting out the “I” word.
If you take the human, and have it write books and articles to the same length of all the books and articles it has ever read, and then use that human’s writings exclusively to train a new human, it is not as effective as the first human, who has been trained on the most noteworthy of 50 billion human’s output in 100 millennia and a multitude of teachers in complex educational system.
That’s why cultures and ideals that revolve around one book or author results in dumb people.
What is the “I” word? Iterational improvement?
I believe it’s not that simple.
In reinforcement learning, robots can perform exploration and exploitation to acquire new knowledge. If somehow this can be applied to LLMs in the deeper sense, a robot would be able to create higher quality content from initial inputs. This is an emerging field of research that uses CoT, ToT, GoT, etc.
Essentially, the outcome hinges on the robot’s capabilities. Eventually, the synthetic data generated by an LLM-equipped robot might be good enough to train another robot. This development could be a critical milestone in achieving AGI/ASI.
Hold on. First consider the no-op. Another human body experiences all of the same stimuli that I have. I can assume it will be no more or less effective than me across all metrics. Overall a no-op would have a Chatbot Intelligence Factor of 1.0. This is the case for the bot as well.
However I have to believe that I can alter some of these stimuli in a way that causes a next human to be overall more effective across all metrics, achieving a Chatbot Intelligence Factor greater than 1.0.
There is no evidence that the bot, given its training set as input, could produce an altered version of this training set that, when used to train another bot, results in a bot that is more effective across all metrics.
Therefore the bot would have a Chatbot Intelligence Factor less than 1.0, or worse than a no-op. I would be curious to know the Chatbot Intelligence Factor of any iterations of any bot.
Can you show us the evidence that you could alter some of your stimuli in a way that causes the next human to be more effective across all metrics?
If I could I would call it a “contribution to humanity” and at this point I have no evidence only belief.
So really what you’re saying is you believe you can do it, and that the AI can not do it, despite having no evidence for either perspective, which makes me wonder… what’s your point again?
Mostly to be able to claim prior art should someone else independently arrive at and publish this framework for measuring intelligence.