Artificial Intelligence (AI for short) has been all the talk lately. From tech giants such as Google and Microsoft, to relative newcomers such as OpenAI, everyone is trying to put AI into everything, from code editors to operating systems to web search engines to even toilets (just in case you thought “everything” means anything other than absolutely everything, nothing is safe from having AI put inside of it).
But how does “AI” actually work, and is it actually helpful to shove it into everything?
First some semantics
The term “AI” is actually quite vague, with many different definitions for it depending on the context, but if we avoid the philosophical questions associated with the problem, we end up with 2 main things people mean when they say “AI”.
Faux-intelligent behaviour from strict rules
Video games make very heavy use of this. This kind of AI controls any non-player character (NPC) in the game, and, whatever their role, the “AI” generally consists of a very complex web of rules which result in seemingly intelligent behaviour. Those rules are not perfect, often times breaking down when faced by unforeseen scenarios, such as when the player is using cheats, glitches out of bounds, or the NPC is faced with modded content which behaves nothing like what it was designed to deal with.
Needless to say, this is not the kind of AI that all of the hype is built around.
Machine learning
This is generally what sits at the core of all the “AI” that has been making waves in the last few years. There are many specific ways to do it, but to simplify, most of them work in one of two ways:
- “Fuck around and find out” (e.g. evolutionary and reinforcement learning algorithms)
- “Hey, can I copy your homework?” (train on a big ass dataset that you totally own)
Okay, I’m shitposting with these names but they’re surprisingly accurate when you really think about it.
Fancy brute forcing
As already stated, the former group is primarily made up of evolutionary algorithms (such as NEAT), and reinforcement learning algorithms (such as Q-learning or PPO). The specifics of each of those are quite complex, but the gist if it is you put the computer behind the wheel without telling it which way is up, let it just fuck around, you tell it how well it did, and repeat this until the AI learns how to do what you’ve been training it to do based on the scores you give it. A bit of fancy brute forcing if you will.
Such algorithms tend to train quite fast, (in fact you can train AI right on your own computer in games such as Keiwan’s Evolution or Code Bullet’s Creature Creator), but they tend to be quite limited in scope. Great for academic purposes, optimising single tasks, or making clickbaity YouTube videos (no shade thrown at CB, he’s fucking hilarious), but not all that great if you plan to make these puppies write code or chat like a real human.
Learning from your elders
A major issue you may have noticed with the former approach is that it’s the equivalent of teaching someone physics by forcing them to rediscover Newton’s equations and why they break down outside Earth’s conditions instead of just telling them how Einstein discovered general relativity and why it’s so much more accurate.
So the fix is easy, you train the AI on what other people have made! Most definitely with consent from the people who made the stuff you’re training on, and definitely not by stealing it and then not paying them a dime when the AI you trained on their shit makes you money!
But I’m getting ahead of myself. Once again I’m simplifying, but on a technical level machine learning algorithms trained on existing datasets work a bit like overengineered compression algorithms. The computer will compress each entry in the data set into a relatively small set of parameters, with the goal being that if you give the AI the same parameters it found for a specific image, you get an image that’s similar enough to the source image within a margin of error set by the programmer. A bit like looking out the window, you see a car, a tree, and the neighbour’s house, and then when someone tells you to draw something with those elements you draw something similar to what you saw out the window. Of course, the AI actually just uses some numbers and not a complex set of concepts such as “car”, “tree”, and “house”, but you get the idea.
CodeParade’s waking nightmare Fursona
Generator showcases this perfectly. Unlike more
popular image generation models such DALL-E which use text prompts to specify
what you want, CodeParade opted to let you set the parameters directly. This
lets you break the generation and get results worthy of H.P.Lovecraft’s books,
but also lets you peek behind the curtain as to how these AI work. It’s not
magic, it’s just numbers fed into a very complex rule set (one generated by the
AI model during training), and when the numbers are set correctly you get
something that might look nice.
But this approach still has issues. For starters, you’re effectively just taking a whole bunch of stuff, turning it into mush, and then molding that mush into shapes that roughly resemble what you started with, or a combination of them. You can’t make something new, just splice together what’s already been made. And there are more, but I’m leaving them for later.
Issues with modern AI
Now that we’ve gone over how AIs work at a high level, we can finally get into the issues with them. There are a number of them, but I’d say the more important ones are as follows
- Unethical dataset gathering
- Resource efficiency is for losers
- My source is I made it the fuck up
The (lack of) ethics in gathering the datasets
As I’ve mentioned earlier, the AI that’s making waves needs large datasets to train on. As I’ve alluded to even earlier (or rather spelled out), gathering that data isn’t exactly done ethically. Most of the data is lifted from the web without the consent of whoever made it, used to train the AI, and then the original author of the data used in training never gets any sort of credit or compensation. This is somewhat forgivable for a student project, but big companies such as OpenAI, Google, or Microsoft all either plan to, or already actively profit from the use of AI made with stolen content.
GitHub Copilot has already been proven to outright steal code from various projects hosted on GitHub without ever crediting the source (despite most licenses accepted on GitHub requiring proper crediting), and potentially breaking free software licenses such as the GNU General Public License on the behalf of users of Copilot. The authors of the original code tho? We’re left with nothing, not even a footnote acknowledging we’ve been stolen from. Some have filed a lawsuit over this unethical use of their code, but considering GitHub is part of Microsoft, a company with practically infinite money, I’d be surprised if anything good comes from it.
Reddit has also been selling fees for access to its vast library of content to companies involved in the AI business, with Google being probably the largest. And speaking of Google, they have their own infinite source of content to train AIs on thanks to YouTube.
Of course, in none of those cases do the original creators of the content being stolen get any sort of compensation. It’s particularly bad when you look at how these companies even run their respective mines of content. Reddit doesn’t do fuck all beyond reviewing and occasionally banning subreddits and users who engage in very blatant breaking of the rules or law, leaving most moderation to unpaid volunteers, and YouTube very infamously which wrongly flags fair use as content theft (quite hypocritical if you ask me), only fixing things if the people behind the channel being wronged stir up enough shit on social media to get the company’s attention.
It’s like NFTs all over again!
Above I mentioned how some machine learning algorithms are actually quite efficient and can run on modest hardware rather well. But this is not the case for the large models being marketed today. These models require entire data centers working at full throttle 24/7. An entire building of computers running at full capacity means a few things: a lot of electricity being used, a lot of wear and tear on the hardware, leading to more frequent hardware replacements, and the need for some heavy duty cooling.
This study by Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat, estimates the power usage of OpenAI’s GPT-3 (an older version of the same model that powers ChatGPT and Microsoft Copilot) to just under 1.29 Gigawatts-hour, over 120 times more than the average power used by a US household over a year according to the US Energy Information Administration.
But wait, there’s more! Computers are not 100% efficient. A lot of that power ends up being turned into waste heat that needs to be taken away from the electronics, and water often ends up being the preferred choice for cooling down entire data centers. But how much water is being used? Well, Microsoft used over 6.4 billion litters of water to cool their data centers used for training their Copilot AI in 2022 alone, and that number has only gone up since then as they not only train new models but operate existing ones.
All the water used for cooling also can’t be just dumped back into a river, or even stored somewhere until it cools down to be recirculated. It usually has to be treated and cleaned, which along with pumping all of that water in the first place ends up adding even more to the already astronomical costs and resource usage.
Fluent Bullshit
Another common issue with AI is the results often times only make sense when you zoom all the way in, and make absolutely no sense when you look at the whole thing. Large Language Models (LLMs) can be asked to write large texts, but if you read those texts you quickly see a text with proper grammar which sounds like the ramblings of a madman. They change subjects, make stuff up, or completely ignore basic logic. From saying no country in Africa has a name starting with K, although Kenya is getting close to making up superbowl results, AIs are not afraid to spew bullshit that looks coherent. The main reason for this is because LLMs in their current forms are trained to generate text, not facts. They don’t know what a super bowl score is, but they do know what it’s supposed to look like, so they can promptly just make shit up that looks legit.
Oh, and the Kenya one? A common internet joke goes something like this
“There are no countries in Africa that start with K”
“What about Kenya?”
“Kenya suck deeze nuts”
At some point ChatGPT found that joke, but didn’t know it’s a joke, instead taking it all as fact and embellishing some more to make it look like it spelled the fact by itself instead of just mindlessly lifting text from elsewhere.
You can see similar issues in AI generated images, with characters ending up with 6 fingers, arms bent at impossible angles, or lighting making little sense, to name just a couple.
Attempts at generating music with AI ran into this problem as well, as any sort of structure that we’re used to hearing even in more free form genres is completely missing.
But is it actually useful?
Not really. AIs have more than once proven themselves to be unreliable, usually requiring a human to double check and/or correct the results, and quite often those corrections end up throwing away most if not all of what the AI offered.
Some businesses are hoping to replace workers with AI anyway, after all, a ChatGPT subscription is far cheaper than the wages of an entire team of programmers, writers, musicians, artists etc, but so far, attempts to make games, YouTube videos, or writing articles using AI have all been pretty bad. Meanwhile the people whose works were used to train these AIs, who would’ve been paid to do the work if the AI wasn’t around, and who would do the work better than the AI if given the chance? They’re left unable to make ends meet, because why pay \$200 for a person to do some coding, painting, music, or writing for you, when you can instead pay \$20 a month for an AI to do it.
Some people also claim AI could be used as an unbiased way to predict where crime might happen, but how would such an AI even work. Well of course, it looks at historic crime data to extrapolate. You see where this is going, right?
In the United States, the long lasting effects of historic redlining have led to a very strong correlation between poor neighbourhoods and neighbourhoods with majority black residents. Due to this and the racism that enabled redlining in the first place, those neighbourhoods were also overpoliced, which led to a lot more crime being found than in richer, white neighbourhoods. So what do you get when you train an AI on data that resulted from historic racism? You get an AI that perpetuates this systemic racism! But since the racist guy at the top isn’t a racist guy but a computer doing what it was told to (extrapolating from historic data), there’s no one to hold responsible this time. Racism with the perfect cover!
Closing thoughts
AI is probably one of the most academically interesting things to happen in recent years, but when put in practice it’s probably one of the biggest disasters in recent years, surpassing even cryptocurrencies in its harmful effects on the environment, all with added negative effects on society.
Is there anything we can do about it? Two things I can think of: avoid using any sort of AI where possible, and be loud about all the ways it’s fucking us over until we’re heard.
Enjoyed the article? Spotted a mistake? Send me your feedback on Fedi: @Reiddragon@fedi.reimu.info