The algorithm analyzes vast datasets of text written by and for humans and is then able to calculate that, given words or phrases, what other words and phrases are statistically likely to proceed and follow them. Thus, it can generate texts that reflect statistical probabilities about word combinations extracted from language humans have composed. Seed it with a couple of sentences and it will give you more sentences that appear as if a human wrote them.
Not since Alpha Go has a machine learning application generated so much excitement and consternation. Here is an example of what it can do taken from the article “Why GPT-3 is the best and the worst of AI right now,” by Douglas Heaven in the MIT Technology Review.
Human entered: In a shocking finding, scientists discovered a herd of unicorns living in a remote, previously unexplored valley in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.
GPT-3 continued: They also were found to have perfectly coiffed hair, and wore what appeared to be Dior makeup.
“We were shocked to discover the unicorns,” said anthropologist Daniel St. Maurice. “They were like nothing we had ever seen before. We had heard legends of the unicorns, but never thought they actually existed.”
When the scientists first arrived in the valley, the unicorns were surprised and startled by the presence of humans, but were also excited. The unicorns welcomed the researchers and explained that they had been waiting for them for a very long time.
More from the article:
As you can see, GPT-3 is capable of producing complex sentences that read as though they could have been produced by a human. The example sentences include cultural references and a believable account of how the scientists would react. Machines that can use language in this way are important for several reasons. Language is crucial to making sense of the everyday world: humans use it to communicate, to share ideas and describe concepts. An AI that mastered language would acquire a better understanding of the world in the process.
But Heaven also points out:
GPT-3 often repeats or contradicts itself in passages of text more than a few hundred words long. It comes out with howlers. GPT-3 hides its stupidity behind a silver tongue, but it typically takes a few goes to get it to generate something that doesn’t show the cracks.
Never-the-less, whenever we first see machines seem to do something that hitherto only humans can do, we are unsettled and have to reorient ourselves. What is really going on here? Is this a breakthrough or a clever parlor trick?
Heaven’s assessment is that this is important as he goes on to compare what GPT-3 does with what humans do with language with the assumption that it represents a point on an upwards curve of capability leading to AIs that “would acquire a better understanding of the world in the process.”
I cannot share his enthusiasm. Language as humans use it is communication. I start with an idea or concept I want to share with you. It is a process of encoding and decoding. I imagine pretty little unicorns wearing makeup with elaborately done up hair and I encode it using words:
“The unicorns were found to have perfectly coiffed hair, and wore what appeared to be Dior makeup.”
Now you have a similar picture. But what GPT-3 does is not communication in any sense but the statistical generation of text. It had no starting concept to be communicated, no mind even to contain a concept. Zero understanding of the world, just as all the GPTs before it had zero understanding. Where is the progress? From the standpoint of understanding, knowledge and comprehension, the hallmarks of what we all agree is intelligence, machine learning, in all its iterations, is flatlined. But this is not just true of language models but of all machine leaning applications. They have many narrow applications but the notion that if you build them big enough, they are suddenly going to become generally intelligent remains the very essence of magical thinking.
Trying to build intelligent machines by scaling up language models is like [building] a high-altitude airplane to go to the moon," he says. "You might beat altitude records, but going to the moon will require a completely different approach.
My own emotional response to the images of unicorns that formed in my head upon reading the GPT-3 text was one of distaste. I felt like I was confronted with the “living dead,” words came out but there was nothing inside, except maybe the statistical amalgamation of words from a vast number of human minds, like ghosts trapped in the dataset, who once had something to say about research projects or unicorns.
So yes, these statistical language models like GPT-3 are definitely in the category of parlor tricks in my opinion, but maybe still useful? My initial reaction was that giving the illusion of communication without there actually being any is basically a deception and only good for deceiving. Certainly, the technology does have massive potential for deception from cheating on term papers to generating fake news.
However, a TechCrunch article by Alex Wilhelm, Okay, the GPT-3 hype seems pretty reasonable, has caused me to reconsider – a little. Wilhelm provides concrete examples of using GPT-3 to help him find additional and maybe even better combinations of words to express ideas he was already trying to communicate. Seen in that light it is in the same category with spell and grammar checker applications.
I see that applications like that could be helpful but with one fundamental caveat. Whatever comes out of these language models is a reflection of the statistical norm, it will never come up with anything that stands above that norm, never articulate a new idea (except by accident), never rise to the level of excellence. As a training tool it might help someone up to the level of average but never beyond.
What price this automated generator of mediocre prose so prone to abuse? From Heaven’s article:
researchers at the University of Copenhagen in Denmark estimate that training GPT-3 would have had roughly the same carbon footprint as driving a car the distance to the moon and back, if it had been trained in a data center fully powered by fossil fuels. And the costs of such training—estimated by some experts to be at least $10 million in GPT-3’s case.
None-the-less, there seems to be a voracious appetite for machine generated text. From a March 29th 2021 article by James Vincent in the Verge, OpenAI’s text-generating system GPT-3 is now spewing out 4.5 billion words a day:
The best-known AI text-generator is OpenAI’s GPT-3, which the company recently announced is now being used in more than 300 different apps, by “tens of thousands” of developers, and producing 4.5 billion words per day. That’s a lot of robot verbiage. This may be an arbitrary milestone for OpenAI to celebrate, but it’s also a useful indicator of the growing scale, impact, and commercial potential of AI text generation.
4.5 million words a day. I am speechless (momentarily). Do we really want to go down this road?
Even though essentially a deception, I can see that the use of text generators is acceptable when a human certifies that the text produced conveys what he/or she desired to communicate when read by another human. Like those disclaimers at the end of political ads where the candidate says, “I approved this message.” As far as passing off machine generated text as written by a person without a disclaimer – that is pure deception and should not ever be considered acceptable.
But it seems we have already gone past the point of no return. We have entered an era where almost anything we read could be partly generated by mindless algorithms, texts analogous to those “original” oil paintings” where the bulk of the painting is copied by people with minimal training and the original artist only comes in to fill out the final details and sign it. Not the best art.
Sharing knowledge through language is the cornerstone of civilization but to communicate knowledge from one mind to another it has to already exist in the first mind. That is not what is going on here.
The only connection the images and thoughts that occur in our minds after reading GPT-3 piece to other human minds is statistical. As a question answerer that could be useful so long as your question is about a matter of fact that the statistically normal person would agree about. That is no doubt where Google is coming from in the first place, how to generate more natural sounding answers to search queries. But what if the inputs to GPT-3 are about other than matters of fact?
Vincent goes on to say:
“Another worry about the rise of text-generating systems relates to issues of output quality. Like many algorithms, text generators have the capacity to absorb and amplify harmful biases. They’re also often astoundingly dumb. In tests of a medical chatbot built using GPT-3, the model responded to a “suicidal” patient by encouraging them to kill themselves. These problems aren’t insurmountable, but they’re certainly worth flagging in a world where algorithms are already creating mistaken arrests, unfair school grades, and biased medical bills.”
Of course, I agree with the concerns expressed but calling the difference between intelligent communication between human minds and what is going on when you read algorithm-generated text a matter of “output quality” is missing the point. It is like saying speech by humans and parrots differ only by “output quality.” They are fundamentally different processes. And where do we get the idea that the problems are not insurmountable? Magical thinking again?
So, this so-called artificial intelligence, far from augmenting our own intelligence is dumbing us down in an unprecedented fashion. Publishing and reading these text outputs drags us down to the lowest common denominator of statistical probability while running rough-shod over individuals who may not be or see things from the “norm.”
Civilization moves forward not by more of the same but by something different. The best ideas are new and exceptional and not easily expressed even by their creators. In fact:
The best ideas are barely possible to express at all.
He is right. It seems to me, that we might be better served if we strove for excellence and individuality in our communications, to seek clarity in our ideas and eloquence in our expression.
One Response
Brilliant and prescient essay! Will follow up; I am doing some research for an upcoming article on algorithms and mediocrity. Very interesting to look under the hood on Chat GBT and how statistical probability fuels the AI learning (hate that phrase as it is a such a misnomer it feels propagandist).