How to Be Generative by Stealing Fire from AI

Monday, August 14th, 2023

We are in the idea and content generation business.

AI is showing how easy machines can repurpose already existing content and do so quickly and creatively.

One of the ways to thrive in this world is not just leveraging those tools, not just tapping into What Does It Mean to be Human, but also being generative at scale, knowing that that content, too, will feed the machine.

One way is to "steal" a concept used by AI, the Generative Adversarial Network. This is used by AI to generate data, famously used by image generators like Midjourney, to take random input and map them into "latent space" which is a high dimensionality map into desired output (faces, illustrations).

There are two adversaries: the generator and the discriminator.

The generator tries to generate realistic images and the discriminator tries to distinguish between real and generated data.

There are two ways we can incorporate this model in our own lives.

One is the Generator is a random, chaotic part of our mind throwing all kinds of thoughts at us, many of them initially "bad" or "negative" because they are working off of a bad "seed" -- the random initial condition around which the inputs are generated.

We, in turn, have another part of ourself which must be a Discriminator -- and this is often shut down. The Discriminator needs to toss aside the generated output that doesn't match the desired output, which in the case of text-to-image AI, is defined by the prompt.

So, for example, if the prompt were, "make an portrait of the Mona Lisa," the Generator would start creating images based on the seed, and the Discriminator would determine which looks like the Mona Lisa and which doesn't till an image is generated that matches the intent of the prompt.

If we have a poorly constructed prompt for ourselves, our Discriminator will at best generate what the prompt says.

Have you ever tried this with one of these AIs and created a badly constructed prompt? The image is disappointing, as well.

So to be effectively generative, we need to be clear in our prompt, and we can learn from prompt engineering how we need to speak to ourselves: clarity, emphasis, and examples.

However, if the generator is slow -- not creating enough images -- or not responsive -- not adapting to the rate of rejection to come up with new images, the final output will also fail to meet the mark.

Humans need to optimize their generator of ideas, creativity, and responsiveness to our internal Discriminator.

The second approach is to not split ourselves into parts, but to have another person who is a trusted Discriminator to work with your Generative self.

The Discriminator helps hone the desired prompt, rejects things that falls short and allows you the freedom to be Generative to try to hit those aspects of your generative self to meet or exceed expectations of your own prompts.

The fantastical, surreal paintings of 15th-16th century Dutch artist Hieronymus Bosch visualize the creative powers of imagination through bizarre hybrid creatures conjured from his unconscious mind, representing unbridled generative idea flow.

Yet Bosch imposed order on these chaotic visions through technical skill and moral themes, just as the article advises discriminating and filtering ideas to prompt imagination toward meaningful goals.

Bosch’s unique ability to materialize his vibrant inner visions while applying a critical eye makes his stylistic genius apt for conveying the dance between unrestrained creative generation and wise discrimination.

(output by Claude.ai)