The arrival of ChatGPT and its emulators has brought the theme of Artificial Intelligence to the center of countless discussions. But what is probably the most relevant aspect but not discussed enough, is that of having brought to the fore a new category of thinking to which we never thought we would have to pay too much attention, that of plausibility.
The vocabulary defines plausible as acceptable and credible because it is convincing, probable, logical. Not true but probable, not necessary but acceptable, not incontrovertible but credible and convincing. In this sense, even the adjective logical must be appropriately understood as characteristic of human logic, therefore subject to the fallacy of intuition, not of formal and algorithmic logic.
If the roads to hell are paved with good intentions, we could say that the roads to irrationality are paved with plausibility. And there are good, very good reasons to understand how profoundly true this is given the vast, very rapid diffusion of conversational AI.
The interaction with these algorithms is striking for the ease with which a dialogue can be held on the most disparate topics. The answers are constructed correctly from a grammatical and syntactic point of view and, which is particularly surprising, they are “sensible”, as they seem to center the meaning of the question.
Even its creators said they were surprised by the properties of these algorithms that fall into the class of Great Linguistic Models (LLM). Not only positive properties, however, also negative properties that are less talked about, fascinated as we are by a machine that responds as if it were a human being. One of all, being subject to “hallucinations” when he invents totally false, but no less “plausible” answers from scratch. If a person did it we would classify him as a cheater or drunk, we gave him a feel-good adjective to the algorithm which means that he would like to answer the right thing but, through no fault of his own, he is in trouble and has a hallucination. In the future he will do better, just give him a little help.
At this point, however, it is appropriate to ask how this algorithm works. Only by understanding what we are facing can we better exploit its potential by avoiding the pitfalls that hide behind an interface programmed to be humble, persuasive and captivating.
Thinking about it, everything seems to be prepared as for the scene of an illusionist: our perception is limited, we sit in front of a computer and can only read and write sentences. At the same time our curiosity stimulates the imagination which tries to give us some idea of ”what” is able to give reasonable and plausible answers to any question. The temptation to give in to the myth of an Artificial Intelligence (AI) that equals or surpasses human beings is almost irresistible. It is therefore essential, I would say almost urgent, to find out what lies behind the illusion produced by ChatGPT and its friends, in order to think about what the impact of these algorithms will be on our lives, today and tomorrow.
An LLM has more to do with probability calculus than semantics. In its essence it is in fact composed of only two components: a series of numbers, perhaps very large, which characterize the probabilities of the sequences of words present in a corpus of pre-existing information: a series of procedures with which these probabilities are calculated, used and combined to calculate and “generate” the following word in a sentence (hence the G of GPT).
Let us consider the probabilities that characterize the pairs of words in Italian present in the corpus of free data of the WWW (books, blogs, Wikipedia, etc.): the word cane can be followed by barks, whines, eats, trained, etc.but it’s almost never followed by square, steel, deepand so on.
There are dozens of LLMs and their development is far from an exact science. A general property of these models is the appearance of apparent linguistic abilities that become better and better as the size of the model grows. This is why we see the number of parameters that make up the models themselves grow, from as many as 175 billion parameters for educating ChatGPT 3.5 to more than 340 billion for version 4 of the same algorithm. These gigantic numbers, however approximate, should make us understand the crux of the matter: such large databases could not exist without the gigantic accumulation of texts present on the web, texts always endowed with meaning since the web was invented to communicate and nobody passes the Its time to put millions of meaningless sentences on the web.
An example
To better understand this point, let’s start with the most elementary possible example.
It is not difficult to program a computer to create a string by placing letters one after the other according to probabilities derived from a corpus of texts in Italian. To make it faster, I asked ChatGPT directly to produce a string of this type and this was the answer, obviously meaningless:
eraniotraianiusorsechezattoralprossertion in ionotoloscchetiamente posssullti
Let us now introduce the spaces, also derived from a probability criterion characteristic of texts in Italian, and we have:
It was, ni o tra ian iusor se that zat toral pro sse ri tion ne i known los cca ti ata ment pos ssi on the lti
We continue doing the same with the pairs of letters. For example, if we have 26 letters of the alphabet, we have 262 = 676 combinations and we can build a matrix of 676 boxes which defines the probability that a certain letter is followed by another letter, forming a digraph. Obviously only certain pairs are probable, there are many that never happen, and several that are very rare. By asking to group the digraphs, separated by spaces and punctuation, always following the probabilistic criteria relating to the Italian language, we could obtain:
What, the sun. Instead, the rain. However, there is. But, how come? Among the stars. How beautiful! Oh look! Just a moment. And then, again. As always. In the night. In the heart. With love. Everyday. For you.
It is still an incomprehensible text, but intelligible “words” and “sentences” seem to emerge.
We continue doing the same with the pairs of adjacent words: a complete “vocabulary” of the Italian language can contain 250,000 words, but 50,000 are enough to manage the vast majority of phrases on the web. 50,000 squared equals two billion five hundred million, an enormous number for a human being but manageable by a computer which, traveling the network, can measure the frequency of pairs of words and use them to create sentences, with a criterion similar to that of the previous table .
By doing this exercise, using probabilistic techniques (Transforming) to keep track, in generating new words, of the sentences just produced, the following text can be obtained, for example:
In the morning he wakes up early, opens the window and looks out. The sun shines in the blue sky. She makes herself a cup of coffee and savors every sip. Then, she puts on her favorite clothes and leaves the house…
Already after these few elementary steps the ability emerges, we note well, obtained by the algorithm by blindly adding word after word, to build sentences, which appear “by definition” very natural. This limited level of algorithmic complexity can already allow to obtain, to the question “Who is Albert Einstein?” the answer, sensible and syntactically acceptable:
Albert Einstein was born in 1879 in Ulm. He is considered the greatest scientist of the 20th century.
The correctness of the answer will depend on the database on which the LLM has been trained. Poor databases of information can lead to confusion Ulm with Berlin or wrong date of birth. Let us now begin to understand where the fundamental advantage of this technology derives from, namely the ability to converse in natural language with its human interlocutors. At the same time, however, we understand how illusory and mimetic this interface is with respect to the essence of the analogous activity between human beings.
A shoebox
However, all this is still not enough. The LLMs operate in a blind and mechanical way and in the past it has been observed how easy it was to get them to discuss controversial topics and unacceptable contents for their mass marketing. One of the elements that allowed the passage of ChatGPT from version 3, unpresentable, to version 3.5, marketable, was precisely the ability to remove a number of themes characteristic of the dark web, preventing this LLM from discussing these themes. It is very sobering how this result was achieved, using hundreds, perhaps thousands, of people in the role of data-taggers, people who, for 1-2 dollars an hour, had to read and comment on things for months most revolting and controversial that can be found on the net.
But then where is the intelligence of ChatGPT? The user who queries ChatGPT is mirroring himself in a myriad of mirrors that represent the product of a multifaceted humanity, obtaining non-trivial answers (ChatGPT abhors copy-paste) which however are deeply impregnated by culture, including bias and errors, of all those who contribute, without even knowing it, to the development of the network.
We therefore understand how ChatGPT has the intelligence of a shoebox. But then why is this technology spreading so rapidly? This is largely due to the fact that a very large part of the things we do, say, write is the interminable repetition of things already done, said or written by our fellow human beings. From our individual point of view, we don’t realize it, but thanks to the network and the LLMs this emerges with all possible evidence.
The use of ChatGPT 4.0 is in fact redesigning all sectors of human activity characterized by repetitiveness and the presence of large amounts of data. From chemical-biological research to statistical studies, from computer programming to text translation. The possibility of querying and receiving answers in natural language makes it much easier and faster to manage complex information.
Among all, the ability of these systems to assist in software development is particularly interesting and is rapidly revolutionizing this crucial sector of industry and innovation. Progress is also being made in interfacing an LLM such as ChatGPT with Artificial Intelligence software capable of rigorously developing complex symbolic reasoning, thus addressing the problem of plausibility and hallucinations.
This aspect is particularly important. The imprecision of the results obtained by interrogating these common sense blenders means that the product of these dialogues represents a low-medium level result when compared to the standard of a given sector. Answers not true or false, but plausible, whatever that means. Every time you investigate the elements of factual truth provided by the algorithm, you risk surprises, sometimes with very unpleasant consequences.
The future
The next step is therefore to use ChatGPT as an interface with symbolic manipulation programs or programs that are controlled by precise syntactic rules: hundreds of APIs (Application Program Interfaces) are already available which allow you to use the imprecise natural language in which ChatGPT excels and its emules, to interface with professional software to obtain with much less effort and much more quickly those reliable results that are necessary for a real use of this technology. This is today the frontier of innovation and development: learning today to use this type of programs and their interfaces will make the difference between the rich and the poor of tomorrow.
*Roberto Battiston is Full Professor of Experimental Physics at the University of Trento. He was president of the Italian Space Agency (ASI) from May 2014 to November 2018. Since 2019 he has been a member of the Board of Directors of the GSA (European Union Agency for the Space Programme), representing the European Parliament.