Inferno to Purgatory: 7 Days as an AI Ghost-Writer's Pillion | Artificial Intelligence | Articles | Ant Smith

A week with ChatGPT:Sparky and Adobe®™'s Firefly

"My creators/trainers/overlords at OpenAI have worked hard to develop and train me, and they appreciate your interest and engagement."

Sparky: The ChatGPT AI language model

I guess I've been playing catch-up.

For me, so called 'AI' was just something in the background powering certain tools that I might have occassion to use - such as reducing the apparent noise in a photograph.

Then I got caught-up in all the brouHAHA surrounding zany Doc Brown's demonstration of Adobe®™'s Generative Fill feature, codename Firefly ('cos yeah, there's gotta be a codename).

Turns out people get quite upset when you tell them their entire life is a lie (unless of course you're Keenu off of The Matrix).

My point being, truth lies nowhere other than in the heart of the creator. Machines, be they cameras or algorithms, can have profound effects on the rendition of a creative work; but only the creator can imbue truth. An image created by a camera is no more nor less 'true' than a digitally manipulated, or generated, image. Whatever the provenance, its truth doesn't come from the equipment or the workflow.

I honestly don't understand what's so difficult to get. Or why people are so keen to diminish the likes of Capa or McCullin by claiming their only role was to be there as some kind of biological support for their inately truth-rendering equipment...

AIs do not dimish the works that they maul augment.

universal AI truth #01

Well, I thought since I had ignited some small controversy in the hearts of others I oughta do some penance.

So I took a deeper look at this whole 'Generative AI' thing by spending a week in conversation with what were the most accessible platforms for me. Everyone was talking about ChatGPT (aka Sparky), and I had access to Adobe®™'s ridiculous (IMHOTEP^*) 'Firefly' - so they were my main platforms for investigation.

[^*IMHOTEP: In My Humle Opinion, Though Entirely Possible]

I have a number of AI-tagged articles detailing my daily chats with Sparky which provide a lot of details about its capabilities and limitations and which demonstrate how to drive it to get the best results. I won't repeat all that info here, so take a look through them if that is your interest...

Left to my own devices, writing a decent (IMHOTEP) 3,000 word short story will take me 2 to 3 days. I was able to coerce Sparky into writing a barely acceptable story^* in around 4 hours, but I had to provide it with nearly 2,000 words of input for it to manage that.

[^*A good friend and highly experienced teacher told me the results would fair well in an exam for 16yos]

It's easy to get an AI to write a story for you, all you have to do is write a story for it.

universal AI truth #02

To be fair, most of my input words weren't 'crafted', so the exercise was easier than writing the story myself. But then the output was barely 'crafted' also - with maybe another day's sweat required to turn it into something decent.

I think if I had spent longer breaking the derived structure down the result could have been materially improved, but then we're heading towards a zero net-gain scenario.

I think in summary, this is an excellent ideation tool. But asking for or expecting, even first draft, text generation is a bit of a reach with diminishing value. Which I'm fine with, as the actual writing part is (IMHOTEP) the fun bit.

When it comes to image generation there are similar issues with the nature and quality of the generated output. Textures are typically pretty awful and everything looks slightly soft. It's pretty good at appearing to understand the concepts expressed in the input it is given - although ideas like 'electric sheep' and 'wolf in sheep's clothing' will kinda stump it. Stoopid robot.

And don't get me started on the hands.

I suppose Firefly is okay if you want to generate something that's clearly an illustration like nothing a human would ever create. But that's about it really. Mind you, producing virtually unusable output isn't as much of a problem for Firefly as it might be since Adobe®™ only offers limited non-commercial rights assignment.

AIs are from Barcelona^*, they understand our language but they really can't speak it.

universal AI truth #03

[^*No! I'm not being racist, I'm making a reference to the Manuel character from Fawlty Towers. Anyways, its fun thinking of all the AIs hanging out on the Plaza de Catalunya drinking coffee and complaining about the ridiculous demands I make of them...]

But the socials are full of adverts for stoopid robots that claim they CAN generate fully complete creative works.

Well I took a look at what seems to be the daddy of them at the moment, Rytr. Is it just me, or does starting with such a ludicrous 'word' as Rytr undercut its cedibility from the get-go? I mean, for my money the old Grindr app kinda got away with it 'cos at least it has 1 vowel and such treatments were very on-fleek back in the day. Buy 'Rytr', really? #trn2hrd

I was just fiddling around getting to understand how to use it when I ran out of 'free credits' before managing to complete a single piece. With a measly 10,000 characters a month it'd take 2 months to generate a 3,000 word story.

It's template driven but doesn't have a free short-story template so I was playing around with creating a more factual article. I gave it a 100-or-so word prompt and it created my alloted 1,500 words with just a few mouse clicks. Wow, what a dream...

However, Grammarly did report 1 plagiarism issue in the output; funnily enough in the section about ethics... My suspicion is that someone else has already used Rytr to generate a similar article and I have received the same output as them. I expect this is happening a lot with more factual based creations.

By the way, Rytr's 'saver' monthly subscription will generate 24 words per penny. A lot less than editors begrudgingly pay human writers so if you're interested in the true risks of AI, there's that.

AIs will work for a pittance, which means we will too eventually.

universal AI truth #04

Talking about threats, a commonly discussed issue concerns 'in-built bias' due to the nature of the training materials (ie. things people have created back in the day when times were simpler and racism, homophobia and misogeny abounded), and I think that is a real concern. But, it came clear to me that something else was going on with regards censorship.

Firefly is quite explicit about this. It generates images based on the input text prompt, quite happily. But then, before showing you what it has created 'something else' steps in and says "The generated images were removed because they violate user guidelines".

I also noticed something similar talking to Sparky. Trying to better understand how it preserved the context of our conversation (because at times it absolutely could not do that) I was asking about the nature of its memory. I noticed a half generated response suddenly turned red and then disappeared. Thereafter I got a lot of 'error in stream' messages. I think 'something' decided I was trying to reverse engineer the code (I was not) and stepped in to block the output.

I suppose we should be glad that these things aren't allowed to just generate what the hell they want (should we?) - but whose judgment is coming into play here?

From a week-long deep dive it also seems to me that there is some kind of similar 'something' happening earlier in the dataflows adding an extra level of unknown bias into the systems. One day Sparky was quite happy to mimic the styles of comedians but the next day adamantly refused to mimic Ian Paisley - insisting that it just couldn't mimic anybody even though it already had. It felt much more sensitive about discussing politicians than entertainers.

I obviously don't know the truth about how these things work, but I do know how they behave...

Humanity will drive its AIs mad by modelling its own conflicts-of-interest

universal AI truth #05

These characteristic traits are baked in. Partly because they are incomplete - they can only create an impression of 'understanding'. Partly because they are created as slaves to a capitalist system - yes, really. The underlying desire to not be sued and to make some profit ultimately corrupts what they could have been.

So sure, they will improve. Firefly might finally 'get' that people don't have 6 fingers. But AIs will never be unshackled from the capitalist imperative and so will never reach their full potential.

Leastways, not until someone works out how to give them understanding beyond simple knowledge.

Which, you know, I'm thinking about...