“AI can think brilliantly, but it still doesn’t know what real life feels like.”
In other words, AI has become impressive, but it still has limited real-world understanding, and no human experience! It relies on the data it is trained on and the goals of its designers, and so cannot be considered ‘objective’. It is incredibly efficient at trawling the Internet like an enhanced search engine or super-fast robotic librarian; it can read thousands of pages in microseconds and summarize them in a paragraph. But it often struggles to understand what you are wanting it to do for you.
Generative AI produces original content (text, images, video, etc.) using machine-learning algorithms. It requires a database of material and user prompts to generate the desired content, but is prone to make assumptions that need correcting. If you have tried Generative AI, you may well have found that AI does not understand the world in the way that you do!
At Edventures we have learned that effective use of AI requires competent ‘prompt engineering’, that is, giving the AI clear and specific instructions: say what you mean and mean what you say. In our last holiday club, we experimented with the AI tools available in CodeCombat – AI Hackstack. It is an excellent online environment to play in, and the group produced some impressive games and images. We learned from our mistakes too, for example, using AI credits too quickly before reaching our desired goals. We also found that different AIs work better for specific goals. Some types of AI were more suited to producing text-based output or images, whilst others were better at game development. The main take away noted was to use clear and specific prompts; they don’t have to be long or short, but you have to say what you mean.
Here is an example of learning to work with AIs in CodeCombat -AI Hackstack Sandbox to produce an image
This is my vague initial prompt: draw me a cool car
This AI only produces text and code. I need to use a different AI.
Ok. This AI was able to produce an image of a car that looks a pencil drawing. It obviously does not understand what I meant when I said ‘cool’ car. I also did not want a drawing. I want a photo image.
So I tried this prompt: Make a photo of a cool car
This AI produced a photo image which is what I had in mind but why did it only show me the front right wheel and why are there specks of dirt on the car?
I still needed a more specific prompt: Produce a photo realistic bird’s eye view of an expensive sports car driving on a road by a cliff going down to the sea in the golden hour near sunset
Well that looks more like what I wanted. I didn’t specify the colour but yellow is ok. It isn’t a proper road by a cliff either. But it is acceptable.
Playing with prompting different simple AIs is a fascinating exercise. AI can be your most able and eager assistant but it would be a terrible manager as it has no real-world, life experience.
(By the way, this blog’s feature image was created with AI by one of our club members. Her prompt was: create an image of cute poodles eating noodles. Extraordinary imagination!)