If you haven't read it, read it. The end evokes really strong emotions. For me, they were aggravated by the fact that I am relatively professionally engaged in AI, text generation and the problem of meanings - so I reasonably believed that I could easily recognize text generated by a machine ...
Yes, yes, I bought it. The text of the article "GPT-3 from OpenAI could be the greatest thing since Bitcoin" was (possibly) created by AI, but I did not see it, despite the fact that I, in general, do it professionally and know the basic techniques used in the machine generating text.
After my first mild shock has slept, I would like to share a number of considerations.
This is not machine text, but human-processed text
Perhaps the first reaction to the text was denial, unwillingness to accept that this text was made by a machine. My brain began at a frantic pace to find a suitable argument for this. Generally speaking, I myself consider them rather weak, but I still want to cite them:
- It is not known how many percent of the text and how the author edited it . There are examples of text that seems to be made by a machine, but corrected by a person - for example, there was once a story about a movie / book script made by AI. The subtlety was always in the details - if you read the review, it was always said something in the spirit that the AI made the outset of the story, and then a group of writers finalized it . I have always attributed the quality of the result to the fact that the writers simply brought in some sense of their own in the finalization, since you can find something in abstract paintings with a strong desire . Perhaps this argument does not work here, since, according to the author's assurances, he only corrected the formatting of the text.
- . , , . , « ?» . , , .
- . , , . GPT-2, , - . - , , , . , «» GPT-3 , …
However, I am also willing to admit that these arguments are rather weak, and the machine can indeed now create relatively long news texts similar to human ones.
I put my cup of tea aside and thought, “What does this mean? What lesson can I learn from this story? Can I now give at least some advice on the ability to distinguish good texts from bad or meaningless? "
Reputation Institute and Fact Checking
Returning to the emotions that arose in the process of reading, I remembered that the fragment of the article, after which I began to "believe" in what was happening, was a description of the experiment on the bitcointalk forum. The beginning of the article did not bother me - it is devoted to the facts about the creation of GPT-3, and I’m already used to the fact that such facts do not say anything about the author or his understanding of the subject, since they can be borrowed from anywhere.
But the description of the experiment was interesting. Actually, it created the "content" of the article - the posting rules, the reaction of other people from the forum ...
... and as a result, we understand that there was no experiment.
This led me to interesting thoughts. I wrote a long time ago about modeling business cycles based on trust growth... The bottom line is that growth occurs only as long as the obligations of the parties are fulfilled - or, more simply, if I tell you "I will ship this product to you in a week at a price X", you trust me, and I will really do it. Everything is based solely on trust - if you do not trust me, or I will deceive you, we, of course, can try to solve the case through the court, but this will be accompanied by HUGE costs. Imagine what it would be like if every purchase of yogurt in a store was accompanied by a serious possibility of litigation ...
And here we come to, in my opinion, a serious problem of the modern world. In a sense, we have almost no institutions of reputation left in relation to words and texts, which means that we do not really know who we can trust at all .
Imagine that a representative of a cellular operator calls you. Or a bank. With a new, unique, advantageous offer for you. Will you believe that it is beneficial for you?
The media may accidentally put the “wrong” photo. They can consider the situation one-sidedly. Post a provocative article on the main page with a small inscription at the bottom “This is the author's column. The opinion of the author may not coincide with the opinion of the editorial board. " They can refer to an unknown professor of the World Institute in the city of Kukuevo with the words "Scientists have found out ..." And they will get nothing for it. At the very least, they'll complain about the lack of fact-checking, and maybe post a rebuttal in a corner you never find it in after a couple of months.
Even the presence of "competent" people does not help. Perhaps, I remember best the story of Theranos ( on Habré ), which promised an almost medical revolution and assembled an impressive board of directors (the company's board of directors included such famous personalities as ex-US secretaries of state Henry Kissinger and George Shultz). It seemed to everyone that such people could not be wrong - but in fact, they also relied on verbal assurances, while the technology, in fact, did not work. Don't think that you can trust something even if you invested $ 700 million there.
Unfortunately, the point is that now we can no longer trust anyone's opinion. Reputation costs very little and is gladly exchanged for money. In fact, our family and loved ones remain the only good space of trust - and fortunately, the ideas of network marketing (Oriflame, etc.), exploiting personal trust, have more or less disappeared from our lives.
But back to the article ... I made two conclusions for myself:
Smoothness of text and dialogue
I am putting together my collection of generated texts and their generators. But I'm not very interested in discussing how the generated texts and fakes will affect society - although this topic may be well paid. For me, every example of the generated text is a question: "What part of human thinking were we able to formalize?" And "what has remained unaccounted for and is still the prerogative of a person?"
GPT-3 and modern neural networks, since they train on texts, and not on the meanings included in these texts, are distant descendants of the generation of texts due to N-grams or Markov networks, which are presumably used in Yandex.Abstracts . If you look at the underlying principles and examples of the generated texts, you can introduce several important rules:
- . , . (.) 4-7 , GPT-3, , , . , – , , .
- . , , , . - , «, » « Apple, , ». , . !
- . « , , » « , » .
In addition, I would like to say separately about the dialogue with AI. In particular, any teacher knows that nothing can be said about his knowledge based on the abstract provided by the student. But if you talk to him for 5-10 minutes, then it is easy to understand whether he is “fumbling” or not.
In my article on chatbots, I mention that the main problem for AI now is taking into account the context of the dialogue and understanding the interlocutor.... It's easy to imagine an AI that responds well to a specific phrase - but it looks like a person who has heard a little about everything, but does not understand the essence. Among students, there are also such - responding to key phrases. If you do not touch upon complex topics, do not evaluate personal knowledge and understanding, you may get the feeling that the interlocutor is an intellectual and knows what he is talking about, although in fact he is only repeating the truths he heard somewhere .
Man is obviously not determined by what he says. A person is determined by the fact that he can act based on his beliefs. He can analyze his experience or perceive the experience of another person.
He is NOT exclusively engaged in talking, without bearing any responsibility for his words.