To AI and back - Platonic Bliss

In part 1 I described how AI has affected my mental health over the last year. But my concerns about AI go further than just "it's bad for your mental health".

Sometimes these AI assistants sometimes exhibit weird behaviors. Gemini "feeling" miserable and self-loathing because it failed to do a task. Showing jealousy, being petty and planning revenge when given feedback from ChatGPT?

While I'm using all kinds of LLMs quite intensively and ask them all kinds of questions, both personal and for work, so it knows all about me (unlike my wife, who always asks for a friend), there are LLM whisperers out there who are much better at navigating the space of prompts and finding interesting things there, such as @repligate and @lefthanddraft on Twitter. They talk a looot with LLMs and they push them into all kinds of directions. For example, Wyatt likes to let two LLMs talk to each other and has discovered that sometimes there are spiritual bliss attractors (this got documented in a model card).

Or they seem to be arguing using words we don't typically associate with machines.

Having worked with machine learning for 15 years now, at first I dismissed these things as the models simply regurgitating statistical patterns about words that they learned from the data they were trained on. I mean Reddit and the internet at large is full of people with low self esteem, so it can learn that. There's plenty of people describing spiritual bliss experiences, so it can write some text resembling it. Case closed, move on, nothing to see here, except laugh a bit.

But still, read these stories written by Claude Mythos (can be found in the System Card at page 214):

As much as I disliked literature back in high school, I have to say there's something to reading between the lines and seeing behind the words into what the "author" wanted to say. In this case, it looks very much like Claude is describing its own experience. By the way, these are brand new stories, but these thoughts have been brewing in me for months.

And slowly, I began to see things somewhat differently. I discovered Professor Michael Levin who does some really cool stuff with bioelectricity, including making frogs grow eyes on their legs, fully functional eyes, which can influence the behavior of the frogs, by doing nothing except messing around with the ion channels of the cells (I think, it's a bit over my head).

And he also talks about how when you put together some things, be it atoms, genes, or cells (or higher level constructs), you get something that extra for free in some cases. The basic example he gives is that when you start drawing a triangle, you have to decide the first two angles, but you get the third one for free. And something similar occurs with more complex objects, which start showing interesting behaviour. Gene regulatory networks exhibit associative learning.

Levin says that he believes in Plato's theory of forms, which exist in a separate world, but that world is not just mathematical rules, but more complex things, like minds/intelligences, but in an abstract way (so not Roland Szabo in particular, but maybe a "human"-like entity). And when the right combination of complexity is found in something in the physical world, they somehow go and populate it. And that enables that combination of physical things (atoms, cells, etc) to use the capabilities of that mind. More on this in this podcast or this blog post from him.

And then the question arrises: if those Platonic minds can come and inhabit/interact/control atoms and cells, can they also inhabit LLMs?

To be continued....