Instead, we are going to tap again into a new development in Art and Neural Networks as a metaphor of where AI progress sits today, and what is feasible in the years to come. For our 2019 “initiation” on this topic with foundational concepts, see here. Today, let’s talk about OpenAI’s CLIP model, connecting natural language inputs with image search navigation, and the generative neural art models like VQ-GAN.
Compared to GPT-3, which is really good at generating language, CLIP is really good at associating language with images through adjacent categories, rather than by training on an entire image data set.
What we know intuitively, and what the software shows, is that the pixelated image can be expanded into a cone of multiple probable outcomes. The same pixelated face can yield millions of various, uncanny permutations. These mathematical permutations of our human flesh exit in an area which is called “latent space”. The way to pick one out of the many is called “gradient descent”.
Imagine you are standing in an open field, and see many beautiful hills nearby. Or alternately, imagine you are standing on a hill, looking across the rolling valleys. You decide to pick one of these valleys, based on how popular or how close it is. This is gradient descent, and the valley is the generated face. Which way would you go?
In the long take this week, I try out a contrarian point of view on personal finance chatbots. Trim, a savings chatbot, just withdrew support from Facebook Messenger. While lots of other chatbots are still invested in conversational banking, what could we take away from the counterfactual of chatbots failing to get B2C traction? What is the impact on the rest of the platform wars waged by Amazon, Google, and Tesla for connected homes, cars, and the Internet of Things?
This week, we look at:
What it means to ask questions and find answers
From asking simple questions that result in neobanks and roboadvisors. Who will win — Schwab or Robinhood?
To asking macro questions about the finance / high-tech competition. Who will win — Goldman Sachs or Google?
To asking profound questions about the nature of the work, and the art of finding your own questions.
We can't formulate the questions for you. But we can give you a framework of needs for both the individual, and the organization.
The questions that you ask are the answers that you will get.
The image is taken from an AI paper which explains how to use generative adversarial networks (i.e., GANs) to hallucinate hyper realistic-imagery. By training on hundreds of thousands of samples, the model is able to create candidates representing things like “just a normal dude holding a normal fish nothing to see here”, and then edit out the ones that are too egregious.
The reason the stuff above is so scary is actually that you can mathematically transition in the space between images. So for example, you could move between “a normal dude” and “just a normal fish” and have nightmare fish people. Or you could create a DNA root for an image which is part dog, part car, and part jellyfish. Check out the video below and the very accessible https://www.artbreeder.com/ website to see what I mean.
