Part 3: 'Artificial Intelligence' and Machine Learning are Overhyped Scams

In the previous post of this series, I said that the next part would be about why all those widely publicized “AI” breakthroughs are closer to clever parlor tricks or the next iteration of ‘autocomplete’ than anything that is useful, adaptive and reliable. To better illustrate the point, and some traffic from clickbait, I will use two semi-NSFW images to show what “AI” can achieve and what it cannot, under real life conditions. So let us start by carefully looking at the first photo below.

As you can see, it shows the posterior side view of a slim and very attractive nude blonde holding a glass of white wine and standing in the roof garden of a medium rise apartment building with the Eifel tower in background. And yes.. there is a reason why I phrased this description to sound like a prompt for a text-to-picture “AI” image generator. So.. can “AI” replicate this photo or create a very similar one. As it turns out- YES, and here is why. All “AI” programs are dependent on one or more training sets (updated or not) to obtain the information necessary to generate algorithms necessary for creating output. For image processing, the source of all data for training sets will be the internet and there is no shortage of photos of the Parisian skyline and slim blonde nude models with a glass of white wine from various angles. Therefore creating an image similar to the one above would be trivial.

Now have a look at the next semi-NSFW photo, with a generally similar theme.

Here, we have the posterior view of a somewhat thick, but still attractive, nude woman leaning against the rail of a balcony of what appears to be a coastal hostel resort. But here is the thing.. “AI” will have an exponentially harder time reproducing anything even close to this photo. But why? A first glance, this photo contains far fewer and less complex scenery than the first one, and should therefore be easier to mimic. And this is where you will start to get a very interesting insight into how “AI” starts to fail.

Have a look at the woman in the second photo again and compare her to the one in the first. What are the consequential differences between the two? For starters, the woman in first photo has a ‘model’ body and is conventionally attractive, while the one in the second photo is certainly attractive to men, but not ‘model’ material. The first model has well coiffed and evenly colored hair, unlike the one in second photo- and the list keeps growing. The woman in second photo has tanlines on her back, an asymmetric tattoo and some pressure lines on her butt that look like panty lines. Also notice the small skin imperfections on her which are absent from the first one. But why should these factors make it much harder to generate a realistic looking “AI” image of the second scene rather than the first one?

Remember what I said about the training sets used for a text-to-image “AI” programs. Now ask yourself- for every hundred high quality photo of a slim model nude woman without clear skin and no tanlines in classical modeling poses how many photos exist of somewhat chunky but still attractive women with normal levels of skin blemishes and uncoiffed hair who has an asymmetric tattoo? Do you see the point I am trying to make? Any “AI” model that is trained to draw nude women will inevitably end up generate the image of a very generic looking but still attractive woman, rather than one who looks a bit imperfect. And all those petabytes of amateur porn on the web won’t change this situation since there is a massive level of diversity among them compared to the more well known model shots- whether they are on some porn site or Instagram. And this is just the beginning..

Did you notice that the posed photo of that model on the garden balcony (or rooftop) of that Parisian apartment looks normal- because the use of such a location for that photo is normal. Now look at the second photo, specifically the surroundings. Notice that it looks like the balcony of a pretty generic hotel room which would be affordable to a woman (or couple) in her late 20s? If you placed her in an expensive hotel room set against some stunning scenery, it would appear odd. But here is thing.. the training sets used by “AI” contain far more photos of expensive hotel rooms and balconies with stunning views than mediocre hotel rooms with average views. It should be obvious by now that just telling the “AI” to show her from behind leaning on a hotel or condo balcony would likely create an incongruous image. Let me explain.

See.. the first photo tells a story. A professional model is posing nude with a wine glass on an expensive garden balcony of a Parisian apartment as part of a photoshoot. So what story does the second photo tell? A woman in her late 20s on vacation with her BF or husband is semi-posing on the balcony of their hotel room after a shower which was probably preceded by sex with her guy. But here is the thing.. “AI” cannot make that guess based on the photo, even if you used a petabyte of similar amateur photos since the majority of those would have been shot at home. Ask me how I know this.. lol. At this point some of you might say- what about using a secondary training set of images similar to the second photo. You could certainly do that, but here the problems.

Firstly, you will be a much smaller set of photos for that secondary training set and the outputs will not be as realistic and relatively error free as those which only used the very large primary set. This will be the case even if the primary data set is used in combination with the much smaller secondary. But there is a more interesting issue with using the much smaller secondary set. See.. if you already have enough real life examples of what the scene should look like, wouldn’t it be easier, faster and cheaper to just use PhotoShop or similar software to alter or tweak existing images that are closest to your requirements? This is especially relevant since even their current versions have pretty impressive tools for “smart” manipulation of photos and other images. But what does all of this mean for “AI”, especially text-to-image programs?

To put it bluntly, by their very nature, these tools would be good for generating super-generic images and recombinations from generic images. So if you wanted to create a photorealistic image of Alyx Vance modeling BDSM themed lingerie while fighting the Combine in an abandoned building somewhere in City 17, you could do it. Heck.. you could even make her look less or more black. Similarly, creating photo-realistic images of a very generic looking Instagram Thot with the latest in fashionable plastic surgery and makeup would be similarly trivial. But actually creating a realistic photo of an average-ish nude young woman looking out over the balcony of some generic beach resort would be much harder as would be anything which require you a to use a small secondary training set. In such cases, simply photoshopping selected images from the secondary data set would be much easier and faster in addition to yielding far more realistic results. Did I mention that human beings quickly get tired of very generic images and subconsciously filter them out from their perception?

In the next part, I will go into the shallowness of other parlor tricks used to promote the idea that “AI” is a tool capable of doing great and “disruptive” things.

What do you think? Comments?