We got the AI revolution wrong

In the last ~2 years, it turned out that AI/ML is much better at doing tasks that we all assumed required a "human level of creativity", years of specialized experience in that field, or "abstract thinking is needed to do this".

Real-world usages of models

The recent advent of AI/ML is being driven by 2 major players - OpenAI and Google. OpenAI built a GPT-3 model (released in 2020) and now is building on top of it. A known example is DALL-E (creating images by text prompts), as well as developed products in partnership with external companies, such as a Github Copilot which writes code by providing a description of it.

Google bought DeepMind in 2014 and in past years has been applying its models and sprinkling its expertise on top of various areas, both internal such as database energy usage optimization, and creating chip designs (also done by Samsung), as well as external ones with Google Health initiative for example.

Midjourney is the most recent ML model to create abstract images from text input (called diffusion model), it is a relatively small initiative by the ex-co-founder of LeapMotion. It created the image above this article. Quite similar but less easy to use is Stable Diffusion model from Stability AI (requires local setup).

All the above examples were typically associated with expert domain knowledge, years of experience, and creativity of the human mind.

Of course, AI/ML (especially ML) is also great at automatization tasks that are not 'creative' - for example creating maps based on satellite images, or cutting background from images.

When the output is good enough

Computer algorithms are doing great everywhere we have some pre-existing data set, and it turns out we have a lot of it in areas like art, books, product (also fashion), or medical records. Computer models just skipped a whole chapter "do computers have creativity" and delivered the results we asked. I'm not going into the discussion if mashing past works to get new ones is "creativity", the bottom line is that output is so good that people are building products around it and are ready to pay for it. Nobody cares if creativity is needed (art) or if it requires having "10 years of work experience in the field" (coding/medical) - as long as the price/time/quality ratio is right.

First ones to be replaced by machines

I think we will see more and more AI/ML usage in the initial stage of product design, art, visual asset creation, book/code writing, and many more, as it is often a fast starting point to experiment. And then we will work on this to fine-tune it, adjust the input parameters of a model and finally fill in missing bits (ex. specs for manufacturers that are not (yet) known to the computer model).

We assumed many of our jobs require abstract thinking or specialized experience, and those will be the last to be replaced by computers. I think that has just changed.

On the bright side, there will be new jobs created - to help AI and fill missing bits of input (ex. create better text prompts), or to polish its results.

Extras

Great read in details of diffusion models and why/how they create so good images from text input.

Simple explanation between AI and ML. The title of this article is using "AI" even as probably a more proper should be "ML revolution", we never anticipated in popular magazines and books a "Machine Learning revolution" we always name them just "AI", in the rest of the text I'm trying to be more precise.