Published 2022

3 years after making this post: https://edisonscientific.com/articles/announcing-kosmos

1.5 years after creating this post:

https://x.com/emollick/status/1781676884557697529

^ was a post from Ethan Mollick, probably related to a paper that was published around the time aiming to do something similar.

November & December 2022

If humans can generate logical conclusions from existing data why can't AI with a lot of information at the very least help us map the unmapped areas within it's domain of information?

This is a "filling in holes" method that doesn’t technically generate anything novel outside of the bounds of data, maybe it can work given current models. In general DL exploits locally smooth properties of data which allow it to generalize without being trained on every exhaustive example, so maybe "knowledge" can be smooth and we can let LLM patch knowledge gaps for us.

With the right prompts and model tuning with the current architecture I think it’s at least worth an effort. Problem is I think they set a block on anything that it doesn't have a high confidence rate on or could even be slightly misleading or inappropriate which limits ChatGPT.

I wonder if you were able to train a separate model to filter for novelty and symbolic logic of ideas and then add a couple processing layers to davinci using transfer learning + clever prompt injection? davinci-3 I think is the same model as ChatGPT with less fine tuning and fewer prompts injected.

Maybe this way it can derive logical conclusions that make sense that humans haven't tested yet. 99% of it will be probably be horseshit but there could be some interesting stuff in there, could spit out the confidence in it's logic, and explain/provide further directions which can also give it a chance to revise or clarify.

Even potentially a cycle between two LLMs to have a prompted conversation until the directive of ‘print when a novel theory has been found’ + a clear definition of what novel theory means. Simple example below could be expanded upon (to be closer to the prompts in the next section) and repeated over and over until the above directive is reached.

GPT1 ‘What is a field in emerging tech?’,

GPT2 ‘AI’

GPT1 ‘Ask a specific question about the emerging tech field starting with why’

GPT2 ‘Why haven’t we reached AGI’

GPT1 ‘Computation, efficiency, self-discovery’

GPT2 ‘Ask a specific question related to one of the reasons you just gave me’

GPT1 ‘What’s stopping us from improving efficiency’