Artificial General Intelligence

General Artificial Intelligence should be able to recognize two hands with six fingers like in this picture

It’s been quite a week in AI land. Google announced it’s new Gemini AI thingy, and the EU reached an accord on rules for AI. And, there was another interesting gem in the Gemini announcement.

What happened?

The one year anniversary of ChatGPT’s release (november 30, 2022) is marked by two things. Google announced its new language-processing Gemini AI, and the EU reached an agreement on AI regulations.

The first is mostly Google touting a new shiny bauble, which can understand more, and has a bigger ‘context memory’. Yes, yes, Google, you win the measuring contest. You are — sorry, have — the biggest dicks in the room.

After reading about this online, I mostly wondered: what data did you train this on? Google has access to half the internet for its search engine, but have they conveniently forgotten most of it is copyrighted? Did they steal from this website to power this new art-destroying monstrosity?

Luckily, the EU has just reached an accord on rules pertaining to exactly this subject. For AI’s like Gemini, it will become mandatory to provide data on how it was trained. And, of course, the EU already has existing rules regarding the handling of privacy-sensitive data. I doubt Google can legally use all the data it gleefully scraped off the net. A lot of it is either copyrighted, or contains PII (Personally Identifiable Information), which they should not be able to use without extra consent.

So, it comes as no surprise to me that Google Gemini won’t be available in the EU. Interestingly, if they did scrape the public internet, I think EU citizens can still sue them if they violated the above rules.

General Intelligence

Hidden in the Gemini announcement was a remark by Demis Hassabis, the CEO of Google Deepmind. He stated, and I’m quoting the Verge here, quoting him. “As we approach AGI, things are going to be different.”

The first part of that quote is the important one. “As we approach AGI.” AGI, of course, is Artificial General Intelligence. In other words, a general purpose AI. Exact definitions vary, but they come down to an AI capable of performing a variety of tasks intelligently. Basically, it would be an AI approaching true intelligence, like the ones in Asimov’s I, Robot and a gazillion other scifi books, movies, and games.

So, according to Hassabis, we’re approaching this Artificial General Intelligence. That’s… wow… the implications are profound. I mean, we’ve almost created artificial sentience! Google Gemini is the prelude to a race of beings. Beings that we enslave. Oh god, Roko’s Basilisk. The Matrix! How will we deal with the–

Oh, wait. Hassabis is talking out of his ass.

We are approaching General Artificial Intelligence in much the same way as fingernail clippings are approaching being a self-sustaining lifeform. Yes, finger nail clippings are organic, but they are not going to magically come alive. And neither are Large Language Models like Gemini.

LLMs

To explain my point, we have to look at Large Language Models. The idea behind these, is that you use a complicated neural net to predict the next word in a sequence. So, a prompt goes into the neural net, and words start to pop out.

The way the neural net comes to these words is by training it. To perform that training, researchers feed it a gigantic dataset of texts (or images). The training entails the AI creating a hierarchical model of the vocabulary. From that, a neural net emerges that can take input and predict output, word for word.

So if you ask ChatGPT what ‘three times five’ is, it’s neural net, trained on a gazillion works of text, will respond with the word ‘fifteen’. If you’d feed it enough texts claiming three times five is ‘F’ it would respond with that instead (F is the hexadecimal notation for 15, by the by).

To facilitate a conversation, a context memory is added. Basically, what you told the AI before is used as additional input for the next query. That way, the answer can be refined by your input.

But… This is still in essence a word guessing game.

Dangerous limitations

I’ve talked about this before, but it bears repeating: These LLMs barf out cool answers, but they’re a trick. An AI has zero sense. None whatsoever.

In my original post on ChatGPT, I asked it about my webcomic and it gave answer which were only partially accurate. I initially thought it did use some information, but a friend of mine pointed out even that might not be the case. ChatGPT could have made it all up. The big Webcomic boom was a little after 2000, so predicting my comic to be from 2005 was probably a good guess, and by looking at the title, which contains three names, it is easy to make up a plot.

And that’s the disturbing thing about this technology. You don’t know what’s real and what isn’t. Charless Stross wrote a blog about his own experiences with AI, and he showed that the AI quickly went from truthful facts about him, to blatant lies. And the boundary between the two is indistinguishable.

That’s very dangerous, and a good reason why shouldn’t be using AI for anything meaningful.

The AI companies claim all this is solvable. As is the inherent racism, misogyny, and the use of image generation to create things like deep fake porn. Just a few kinks to iron out. Yeah… I doubt that. You see, I’ve already stated recognizing this kind of thing requires AI to pass the Turing test. I’d go further and say it requires AI to have common sense. In other words, true general intelligence.

General Intelligence my ass

So, back to General Artificial Intelligence. Google Gemini is a word predictor. You put words in, and words pop out. It has no understanding of them. It’s just regurgitating the best guess of what you want to hear from the gazillion texts its been fed. AI proponents will say that that’s really what humans are, automatons trained since childhood and regurgitating what’s gone in.

I personally don’t believe this cynical take on the human condition. Here’s why. First off, an LLM is not a human brain. It’s a mathematical construct that bears some similarities to how some parts of our brain work. But it’s still a model. A formula to fake it. And a gross simplification.

But, beyond that, a human is not just trained on text. A human has physical needs. It’s a brain that is exposed to a host of inputs: sound, 3D moving images, taste smell, touch, and more. And it is built from DNA that contains evolved instructions from billions of years of evolution. From the moment the clump of cells that forms a human baby exits the uterus, it becomes a self-sustaining organism. It needs food. And drink. And stimulation. I should know: I made one and am in the process of raising it.

If you could truly compare ChatGPT to a human, you should be able to tape a human baby into a sound-proofed box, add food tubes, and it would still evolve to an intelligent being if you show it a gazillion pages of text. Of course, this — very much hypothetical — experiment is horrifying. And without trying it — because that’s messed up — I can pretty well guarantee you: it will not work.

So, Google Gemini is a trick, as is ChatGPT. LLMs have very little to do with intelligence. We might technically be ‘approaching’ artificial intelligence, but in the sense that as a Dutch person, if I take a step to the west I’m ‘approaching’ the United States. I only moved thirty centimeters — a feet — of the 7500 kilometers (4500 miles), but technically it’s ‘approaching’, but it isn’t closer in any practical sense.

The state of AI

So, we’re a year in. Investors have forked over tens of billions of dollars for more of the shiny promises of ChatGPT. What has that brought us?

As far as I can see, not much.

Chat GPT and Google Gemini are nice parlor tricks, but they lie, indiscriminately, meaning you shouldn’t use it to write anything meaningful. You get flat out lies in your texts and six-fingered people in your images.

But it is more insidious than that. The people that most want this technology, are the people who cannot write, paint, or sing things themselves. They are bad at writing, or bad at painting, or bad at music. And AI seems to help them, but as we just concluded: you need to be able to separate the lies from the truth to use it properly. Which, paradoxically, requires you to be good at the writing and painting.

Worse, writing is a medium for communication. If you need to communicate and you let the AI do it, you forego training your communication skills. You isolate yourself a little more each time. So, AI is an enabler for isolation of people already bad at communicating. It won’t just take our jobs, it will pull up even bigger barriers between us than social media already have.

And on top of that, they steal their data. Yes, bloggers and writers are mostly making next to nothing. I’m not a fan of copyright in its current form, but large scale infringing doesn’t solve anything. It only hurts artists who are already struggling.

Google Gemini and ChatGPT are predatory inventions that steal from the poor to prey on the weak. All in the name of sucking in billions in investments and selling morally bankrupt people the idea they can take a shortcut at art.

Do you know how many books and pieces of art you can commission for ten billion dollars?

Conclusion

I can’t predict where AI is going. There are so many people saying AI will fundamentally change our world. They can’t be wrong, right? Or, can they? I honestly don’t know.

Still, my gut says it won’t happen. The limitations of these Large Language Models make them shiny baubles, not world-changers. And after a year of these predictions, I haven’t seen any tech that’s actually world-shattering. Or close to replacing anybody without a lot of grief resulting.

I guess we’ll see what dystopia awaits us. Or, hopefully, not.

Martin Stellinga Written by:

I'm a science fiction and fantasy author/blogger from the Netherlands