I was asked by several of my students to write about the ChatGPT, yet I kept delaying it. The reason is very simple. I have so many conflicting feelings, that I did not know where to begin. Let us start with a simple idea of the “uncanny valley”.
The current generation of AI tools is incredibly good. ChatGPT sounds like a person it tries to imitate including the style and the message. DallE2 generates images and visualizations often better than the relevant artists. The relevant diffusion models are very rich and they can generate pretty much anything with technical perfection beyond human skills. And yet they are incredibly creepy.
When I say “creepy” I refer to evolutionary mechanisms all of us have. We are naturally uncomfortable with mimicry. This is a survival instinct from our ancestors and poisonous creatures deep in the jungles we can easily imagine. If we see, hear or read something and we understand that it is not honest and genuine, we get instinctive rejection.
Less effective AI mechanisms are more limited and do not try to copy human styles so well. So we respond to them more positively.
Another common evolutionary mechanism is vomiting when we feel sensory disparity. Usually, this used to happen as a result of poisoning. So when we get vertigo, e.g. vision disagreeing with vestibular apparatus we tend to become dizzy. Nowadays many people do this recreationally.
Diffusion models tend to fill in the gaps by hallucinating the most appropriate combination of features. The result may be right or wrong or both, but it is usually glorious. Consider this the ultimate form of improvisation. The network will take whatever input we provide and use it for a glorious improvisation. At the current level of technology fact-checking is limited. Yet, the network will get a lot of facts right due to the contexts it learned from.
So basically, if you put an expert like me in an improvisation studio and supply me with LSD, the output will be roughly compatible with a diffusion model AI.
This is what the politicians do
Some of the best salesmen and politicians kind of hallucinate the same way the AI does. A funny story. I had a problem with my AirPods. So I went to a laboratory. The technician was so unfocused that he said to me “There is a problem in the brain of the device.” When I asked him what the brain of the device was, he became very puzzled. I did not work with him after that.
So it is normal for salespeople to hallucinate aloud. Donald Trump is a master of this technique. I am still not sure which part of what he said he actually believed in, which part was a clear manipulation, and which part was improvisation. I am quite sure that AI can out-trump the ex-president.
I was really impressed by some coaching skills presented by AI. When coupled with a person who knows what he is doing, the hallucinations provided by AI are often translated into actionable ideas that actually work very well.
The diffusion models are very good at preserving the context and filling in the gaps. Thus if there are gaps in our logic or if we diverge from the context, the diffusion models can help us. Moreover, the current generation of diffusion models is so rich that it can fill in the context correctly – using cues of what experts do in similar situations.
AI is also incredibly creative since it can easily cross the gap between the expression techniques and knowledge domains. Thus we can get unexpected knowledge transfer and fresh new ideas.
Will AI take our jobs?
There are many salesmen, coaches, and support people that have very little understanding of what they are doing and hallucinate answers based on the knowledge base of the relevant organization. These jobs are likely to go to AI. Branding experts and illustrators often start by hallucinating various combinations of ideas. AI can make this initial brainstorming very easy, increasing the productivity of creative people. To be honest, AI can also fill in trivial engineering jobs by implementing very basic solutions and design patterns. So yes, some jobs will be lost, especially the entry-level positions usually preserved to interns and outsourced to poor countries.
15 years ago AI started winning Trivia contests. I hope still we are 10 years away from AI that will be able to check the facts and remove less effective hallucinations. Once this step will be complete, AI will be able to take the jobs of real experts providing real solutions, for example proving mathematical theorems. Around 2050 AI is expected to be smarter than a human in pretty much any expert or leader position. And this future really frightens me.
How can I use diffusion models in my work?
Really smart people are looking not for new threats but for new opportunities. AI offers an abundance of such opportunities, allowing each of us to transfer our will into a text of an image. This means that we need to work on our creativity and accuracy in formulating keyword requests, and then AI will provide us with illustrations for brainstorming, creative feedback, and texts clarifying the context.
The biggest change will probably happen in the school home works where kids do not come up with non-trivial ideas. AI will be able to do the project for the kid, just like parents have to do nowadays. I think AI will easily write down the trivial and entertaining blog posts, or at least fill in the missing data. It can also write and review code. This may for example deal with common coding mistakes and more boring coding tasks.
Using DallE2 for visualiziations.
In visualization for memorization and guided visualization for self-help, we need to create a lot of visualizations. We can ask DallE to do this for us. Clearly, if we use complex and abstract words, it will fail. But if we can simplify our request, AI may actually help us formulate our visual dictionary and provide cool funny visualizations.
We can also ask AI to create summaries and reviews on some subjects and this way we can learn the context before diving into the details.
None of the new AI models is revolutionary, but the combined innovation presented by these models generates a technological breakthrough. It is uncanny, threatening, and also inspiring, educating, and enabling. So far the will is still in our hands, and AI is just a smart tool. As such, we need to learn how to benefit from its capabilities.