The 2023 Alpha Sources AI advent calendar
AI has taken everyone and everything by storm this year, most successfully embodied by the services created by Open AI, chiefly its digital assistant ChatGPT and image generation service DALL-E. All technology firms, if not companies more generally, are now looking for a way to integrate AI into their core business, less they want to get left at the proverbial station by the departing train. I am moderately curious and open-minded to new technologies so I have dutifully tried to update my knowledge. I have read around, and listened to conversations with Open AI’s main protagonists and proponents of the technology, as well as those who believe that AI could well be the end of humanity. I am still undecided about how important this technology is. After all, a healthy dose of cynicism and scepticism have to be applied to everything that is presented by Silicon Valley as the next big ting. It’s possible that we’ll be praying at the altar of AI, or even AGI, in short order, but it’s equally possible that no one will remember OpenAI in a year’s time. What I will say, however, is that the sheer brute force with which ChatGPT and DALL-E can generate content that passes for acceptable, or even high-quality, written and visual material is remarkable. And remember that both of these services are still, to a large extent, constrained by what they can do in their public versions, at least for now.
One of the more remarkable twists in this story has been the spectacular factual gaffs made by ChatGPT. Earlier this year, a lawyer in the U.S. used ChatGPT to do legal research, which the AI promptly did by coming up with six case precedents that didn’t exist. Needless to say, that didn’t go down well with the court. This has now come to be known as ChatGPT “hallucinating”, and as it it turns out, it does this quite a lot. I have noticed similar errors when using ChatGPT to come up descriptions of concepts and theories related to my work on demographics and economics. It is a great aid at first glance, but it less great once you realise that the sources and facts provided by ChatGPT, to substantiate the argument it sets out in text, often aren’t true. What’s worse, they look true, so true in fact that even experts in the field can easily be taken for a ride. But such hallucinations can be a force for good in other contexts. Marc Andreesen, an American business man and software engineer, puts it well when he describes how many initially thought that AI would displace blue collar workers or simple white collar clerical tasks, but that it seems instead that it could be more creative tasks that are put to the sword, at least initially. I think Sam Altman, one of the chief creators of ChatGT, has made a similar point. In an economic sense, this means that AI could be shifting the traditionally understood spectrum of value added/productivity—blue collar to white collar, non-creative to creative etc—in a way that no-one had anticipated.
Taking AI for a spin
The realisation that AI could be a serious challenge to creative content creation by humans inspired me to do a little Christmas project; the Alpha Sources AI advent calendar. Specifically, this will be an advent calendar of stories, not necessarily about Christmas. In fact, most of them won’t be about Christmas at all. It starts the 1st of December and ends on the 24th. I am northern European, so Christmas eve is the big day for us. Apologies in advance to my English and North American readers expecting an entry on December 25th. You can follow it all on the blog, and I will be using my social media presence to highlight the project too.
The rules are simple. For each day, I have generated an image with DALL-E2 and a corresponding fictional story with ChatGPT 3.5, in 400 words. I have transparently added the prompts used to generate the DALL-E2 and ChatGPT content after each story. I was not able to subscribe to ChatGPT 4.0, and DALL-E3 in time for this project. Some of the pictures that I am seeing from the latter suggest that I would have been able to boost the quality of the visual elements, significantly. Maybe next year. Some of the stories are stand-alone while others are chapters in two or three-part series that run through the calendar. I have tried to stretch the use of genres as much as possible, though the public versions of these services invariably set some hard stops. Satire, regrettably, is all but impossible, especially in terms of images, as is the use of actual individuals. Violence, gore and general bodily nastiness are difficult to generate too, again especially with AI image creators. This constrains the creative exploration of key genres such as crime, war, conflict, horror and related themes. I suspect the same applies to sexual and romantic themes, but I didn’t explore either.
I have tried to let the AI do as much heavy lifting as possible, but with an important distinction. The pictures generated are all unedited versions from DALL-E2, in some case after several variations, while I have edited the text generated by ChatGPT for the stories, in some cases substantially. I am a writer after all, not a visual artist. I have standards, even if they are, for this kind of project, relatively low.
I don’t think (creative) writers or visual artists need to lay down their pens and brushes just yet. DALL-E2 comes up with all kinds of rubbish along the lines of people with “six fingers”, arms not properly attached to the body etc. It has been substantially easier to do abstract pictures and visualisation rather than specific motives, no matter the prompt. In general, DALL-E2 struggles with details, though as I said above it seems as if DALL-E3 is a fundamental step up in that regard. ChatGPT, for its part, too easily resorts to clichés and a general trivial cadence in its writing. It does not, as it turns out, have a voice, though if you force it to re-do prompts it does seem to get better over subsequent iterations, or at least, more creative. That said AI is is a force to be reckoned with in both disciplines. Specifically for writing, the best way to articulate what ChatGPT does—and what it will become better at over time—is to first to see writing as the act of solving a problem, and secondly, to realise that ChatGPT helps to solve such problems. An author might need his or her main character to move from A to B—emotionally or spatially—and to do so in a particular manner, or she might need to describe this character and his or her demeanour in a certain way with certain adjectives. A writer of economic research might need to describe an idea or phenomena succinctly to a non-technical audience. Or a writer of manuals for toasters might need to describe a feature of the product. In all these cases, and countless more, ChatGPT can help. It is unlikely to produce the final version of any individual text, especially not in a first go, but it can certainly set you on the way.
I hope you enjoy the calendar.
Have a Great December!