Tag Archives: AI

AI – Two Reasons for Optimism

Posted by natechen on

There’s been a lot of hand wringing around the potential innovation and potential hazards of a thing called “AI.” There have been many debates about ethics and implications. The technology could force a significant shift in the way we look at intellectual work and intellectual property and I regret to inform you that any attempt to prevent it’s development will inevitably be worse for humanity than allowing it. (Don’t believe me? Look into why Roman technology stagnated.)

What I don’t want to do today is contribute to that debate. I am more interested in some things I’ve seen in AI that are actually signs for optimism about the way we will react to the technology. This is not to disregard the shortcomings and hazards AI may pose. But I am a creator first and foremost. We survive by making things that emphasize the good and beneficial aspects of our tools rather than by constantly dwelling on all the shortcomings. Hazards are things to be avoided and drawbacks things to be compensated for. The real question is what will we get if we avoid and compensate our way to a successful AI creative environment?

Well, the first thing is we will get a much more verbal society. Ever since Apple Computers introduced the first graphical user interface (technically inspired by Xerox, I believe, but still usually credited to Apple) electronics have been moving us towards a visual culture. Look at any smartphone screen and you can see the upshot of this. Lots of pictures, very few words. However the things we call AI are large language models (LLMs) developing algorithmic prediction based on a neural networking framework – a bunch of fancy terms meaning they read the Internet and form an idea of how the words connect to each other. That means in order to get an output from the LLM you must input words. You cannot press buttons with pictures on them. You cannot draw something.

You. Must. Use. Words.

This is very different from the way using electronics have been going for the last twenty to thirty years. Nothing has dealt more damage to the modern person’s verbal skills than how little they are needed to use modern tools. Don’t get me wrong, the visual communication employed in modern user interfaces is quite impressive. Given the international market for much of these products its also very practical. However it has also reduced the interest in and power of verbal communication in almost every aspect of society. An AI built on an LLM pushes the pendulum back in the other direction by forcing prospective users to interact with it verbally. For the writer and the storyteller that is a positive development.

Of course, AI requires a very idiosyncratic kind of verbal communication right now and that’s less than ideal but I will take what I can get.

That brings me to the next thing about AI that gives me cause for optimism and that is the need for framing. If you have used some kind of online form in the last five years or so you may have been asked to find all of the stop lights or buses in a picture before you can submit it. The primary purpose of this exercise is to prevent automated programs from flooding the form with submissions. The bots that fill out these forms cannot understand the pictures so they fail this simple test.

The secondary purpose of this exercise is to create AI that can understand the pictures.

One of the things no AI can do is frame an object. When the AI program looks at that picture all it sees is a bunch of pixels arranged in a grid. It has no way to tie specific groups of those pixels to a concept like a bus or a stoplight because buses and stoplights are arbitrary concepts invented by humans. The AI has to be taught the concept to understand it. The idea must be “framed” for the AI by human beings who already grasp it.

Human beings have a remarkable ability to learn new concepts and apply them to the world around them based on their pattern recognition skills. It is this ability to “frame” issues that gives rise to creativity, language and communication. Even if an AI can be taught the very broad, basic aspects of something like the law or medicine it still will not grasp the intricacies of a given situation in its specifics. Working out these intricacies and communicating them back to an AI is going to be a necessary skill going forward. This, in turn, will demand people develop situational awareness and communication skills, things which technology has so far driven people away from, rather than towards.

This emphasis will, once again, push people to develop verbal skills which our society has largely allowed to atrophy over the last thirty years or so. In this environment there will be plenty of opportunities for people with a strong command of language to thrive. Better yet, it may change cultural tastes. Visual art is all well and good, don’t get me wrong. I love to draw as much as I love to write. However there hasn’t been as much taste for verbal craftsmanship as there has for visual craftsmanship in my lifetime and if the rise of LLM AI pushes our culture towards verbal excellence again I think it would be a nice development.

I am not saying these things are guaranteed. Nor am I in any way implying that AI will not cause our culture considerable difficulties as it grows towards its full potential. The printing press and the Internet did those things as well. However I do see some reasons to embrace this shift in technology not just for its ability to boost dreary things like efficiency and productivity but also for its ability to push our culture towards aspects that have long been ignored by most people – the communication of ideas through verbal excellence. It is by no means guaranteed but one can hope.

AI and the Digital Frame

Posted by natechen on

The use of Artificial Intelligence in creative endeavors is a topic of growing debate, with people who are strongly for it and strongly against it. Personally, I don’t think AI is as big a “threat” to creativity as some pretend. I also don’t think it’s a huge boon to creativity that many of its biggest boosters imagine it will be, although I certainly think it will have uses very soon. Let me explain what things look like from my very casual understanding of AI and my much deeper understanding of the art of storytelling.

First of all, let’s address the term Artificial Intelligence. This is a marketing term. There are a number of assumptions baked into the term which I don’t entirely agree with, although I will be using the term AI for the bulk of this post. What the programs we call AI do is they use mathematical algorithms to predict categories and outcomes. Pattern recognition and prediction is a function of intelligence. However, AI does not organize the information that it uses to make predictions by itself, it relies extensively on user input to create connections between data points then uses very advanced math to predict further connections or to anticipate what connections people would make between new data points.

A more honest term would be algorithmic prediction. However predictions are much riskier things than intelligence so the decision has been made to brand these programs as AI. There’s a lot of tricky things baked into this idea, not the least of which is that prediction is intelligence. That could be a whole essay in itself but I’ll leave it be for now. Let’s move on to the second important part of the discussion for the purposes of my area of expertise, the large language model.

AI that takes gives output via text or voice synthesizer choose their own words based on a large language model (LLM). These models analyze truly titanic quantities of text to build an idea of how the human language is used. Typically they are trained on some portion of the available text on the Internet. Risky? Undoubtedly. Most of the Internet is used by people who use its anonymity as an excuse to prolong their adolescence and write accordingly. As a result many of the “chat” AI out there, like ChatGPT, come off as shallow and immature. Personally I don’t think that’s a shortcoming of the technology but rather a shortcoming of how it’s been trained to make predictions. But I digress.

Once a LLM is trained you prompt it with a few words and concepts and the AI sees how those words are connected in its data sets, clubs them into the most common format out in the wild and regurgitates it all for you to read. Of course, it adjusts the format according to certain rules in its coding. AI tends to write with excellent spelling, grammar and punctuation, for example, even though much of the Internet has none of those things. So, with this information about LLMs in mind let’s talk about using AI in storytelling.

The art of telling a story involves very deliberately taking events and arranging them into a narrative to create character development and provoke an emotional response from the audience. The nature of a LLM makes it a very poor tool for this undertaking. Remember, the way AI works is by using it’s LLM to predict the most likely way words and concepts are connected. So if you use AI to build a story concept you will get what is statistically most common for that kind of story. I cannot think of any less creative way to tell a story. It may be mathematically deliberate but math doesn’t define character development or emotional response. Emotion, in particular, is blunted when the same emotional stimulus comes in over and over, which is exactly what you will get when statistics drive your outcomes. You might argue it’s the entire point.

If you’ve ever read any fiction on the Internet you’re probably aware that a huge amount of it is very derivative. Repetitive, predictable characters, plots and settings. There’s nothing wrong with writing something a lot like something you’ve read, in fact it’s one of the most important exercises for developing writers in my book. But when you’re building a LLM all that repetition weights the model towards those overemphasized story components. The AI is going to give you more of that than anything else. In short, your story will be confined by the whims of whatever is popular on the Internet, making it even more predictable and derivative.

In time, it may be possible to dig into the structure of your LLM and tweak what elements are weighted and how but even then, it will still be in service to an AI. It’s still a predictive – and predictable – algorithm.

The second problem with AI as a creative tool is the ability of the programmer to build frames into how it delivers output. I mentioned this already. AI chatbots can be programmed to regurgitate text with proper grammar, punctuation and spelling and that’s not something that’s well reflected on the Internet, after all. Those are guidelines programmed into the AI by their creators.

There is no guarantee that these will be the only preprogrammed tools put in place to bound the output of the AI. In point of fact there’s plenty of evidence other such tools already exist. People who have tinkered with ChatGPT have found it has very strict ideological blinders placed on it and many of the variants of AI image generators will reject prompts with certain key words like “fiery” in them. Yes, the word fiery is banned. No, it’s not clear why. Right now these constraints are very obvious.

However, like all computer software, AI is getting more seamless in how you interface with it and more opaque in how its internal logic works. I doubt the creators of these tools will have them simply reject prompts they dislike for much longer. Soon they’ll just replace these undesirable prompts with more palatable ones. The end user probably won’t even be able to detect the change.

This second problem is the part of AI that really disturbs me about the technology, that audiences and creators will have their creativity bounded by the constraints of programmers without even realizing the straight jacket being put on them. It’s why I currently view the technology with great suspicion. And its why I have no plans to use it in any of my endeavors at the moment, no matter how great the time savings from doing so potentially are.

Now I’m not a total pessimist. I think there are a ton of very useful applications for AI in creativity coming down the pipe. So far none of the basic structural issues of algorithmic prediction or LLMs have been overcome so AI still pushes towards homogeneity but that’s not insurmountable. If deployed carefully and judiciously, to handle small tasks, that can be balanced out. The digital framing problem is more complex. I doubt the danger it poses can be entirely removed ever, since it’s an effect of the human nature of AI creators rather than a result of some shortcoming of the technology. In that matter you’ll just have to find programmers who you think you can trust and pay as much attention as possible to what happens when you prompt it. That’s still years away, though.

In the mean time, continue to watch the technology develop with a wary eye. When the conditions are right for widespread use of AI the creative landscape will begin to change incredibly fast. If you’re ready for it that will be a once in a lifetime opportunity to get your stories out. I am waiting for it with great interest.