Unmaking Sense

Episode 11.50: What Mixtral-8x7b-instruct-Q4_K_M thinks of these arguments. I read a transcription.

This episode unusually is composed entirely of me reading out the contents of a conversation that I had this morning with the Mixtral-8x7b-instruct model, I think the 4KM quantised model if you’re interested in that sort of thing. I think it’s really interesting even if it’s a bit repetitive, but I think nonetheless the model’s capacity to summarise the argument of episode 47 which I gave it as a text file is truly remarkable and the way it takes up the ideas and runs with them more remarkable still. In episode 51 I will share some of the ideas that I have had as a direct consequence of this stimulating conversation in a way that further illustrates exploration of a latent space.

Episode 11.49: That human minds can make no sense does not mean there is no sense to be made.

It is a legacy of the human arrogance that comes from believing ourselves to be the chosen species of the planet and the universe, put here by God to dominate everything. That prejudice is demonstrated in the way that we still harbour the belief that something that is unintelligible to the human mind is unintelligible to any mind. But image to image diffusion shows that chaos as it appears to us in the latent space is not the chaos we imagine it to be. There is a potential for the creation of new images out of the points in latent space that have themselves come from parent images that have been made unrecognisable by the noising process. By unmaking sense, we follow that same path, and we intend that by remaking sense we may emerge from it with a better understanding of the world and a better way of living in it.

Episode 11.48: Image to Image as a metaphor for changing our sense by unmaking and remaking sense.

By unmaking sense we create and add noise and uncertainty in much the way an image to image AI diffusion model transforms one image into another. But in remaking sense according to some new principle we reconstruct our worldview in a way we at least hope will be better and intend to be better.

Episode 11.47: Creative acts as realisations of points in human & AI latent space; either can matter

Human creativity may be less deliberate than we like to think, and really just the expression of something that has evolved over tens of thousands, even millions of years. So AI creations can embody meaning too because they arise from neural nets and training that are reflections of human culture. It has little or nothing to do with intention.

Episode11.46: Authenticity, copyright, intellectual property and AI copies of identifiable artists.

Now that artificial intelligence can reproduce works of art, and indeed create new ones in the recognisable style of almost any artist, and to such an extent that they are almost indistinguishable from works that are original, raises the issue of whether an artist’s style is itself the principal legacy, rather than the works painted. We discuss some of the issues that this raises and the question of why being an original painting by the real artist matters so much to us and whether it should.

Episode 11.45: The bias implicit in some AI training means we are automatically fed skewed data.

Not only textual data but the kinds of images produced by AI are suffused with cultural biases and prejudices. We need to be aware of the distortions that can become ingrained if these implicit biases are not countered and held in check.

Episode 11.44: Motivational impacts of AI, especially those that may affect learning and creativity.

We have not even started to come to terms with the motivational implications of AI, whose dangers far outweigh the more dramatic science fiction scenarios so many love.

Episode 11.43: Uploading files to local AI models and creating deep fake videos based on tiny data.

We talk about ways to upload whole files, text files as prompts to locally running large language models, and after that we talk about creating false videos that look as though you’re saying things that you’re not saying, and have never said, on the basis of little more than about 30 seconds worth of data.

Episode 11.42: A three-part excursion on Diffusion Models, Recall and Human Attention and Creativity

We explore some of the intriguing parallels between diffusion models - noising and denoising - and human learning and recall with emphasis on innovation and creativity.

Episode 11.41: The RAG Trade: Retrieval-Augmented-Generation (RAG) and updating AI models.

A short introduction to RAG.