GPT-3 language model matches humans in psychological tests
Is GPT-3 ready to replace humans?
The scholars at the German Max Planck Institute recently examined the general smartness of the language model GPT-3, which is AI - based.
Through the usage of testing tools, the scholars examined things like causal thinking, discussion. Then they compared the conclusions with those of humans.
Their results show very different results. In particular, while GPT-3 is of competition to humans in some areas, it is less productive in others. The reason for the latter might be due to minimal or lack of cooperation with the real world.
AI is efficient in training to respond, when it is given information in natural language and is capable of generating big variations of texts. Up-to-date, in this sense, the most efficient network is GPT-3. This language model was presented to the public no earlier than 2020 by Open AI, specializing in AI research.
GPT-3 can be asked to make various texts. It takes those from the big amounts of information, downloaded from the internet. It can write texts, similar to human-produced texts, but it is also good at solving math problems or programming.
The Linda issue: to err is not only human
The information about its abilities irreversibly poses questions like, can AI have potentially human cognitive functions.
To learn more, the Max Planck scholars decided to put GPT-3 through a number of psychological examinations that look at different areas of common intelligence. A few scholars also monitored the skills of AI when it comes to decision making, search of information, logical reasoning, the skill to challenge its own instinct.
When the scientists looked at the answering results of GPT-3 and those of humans, they considered correction of answers and AI’s similarities in mistakes to those of humans.
An issue, known as the Linda problem rose. It comes from cognitive psychology and it is also considered to be a classical one.
In this test, the subject communicate with the made-up youthful female by the name of Linda. Linda has concerns about social justice and is against nuclear power. The subjects then are given certain data and are asked to conclude between two remarks: Who is linda? Is she a bank teller or is she a bank teller and a feminist movement activist?
Most people would lean to the feminist side of the question. It is interesting, since the feminist side of the question makes it less likely probability-wise. The same goes for GPT-3, which “adopts” the human logic - it does not think logically, but reproduces deception, that people fall into.
Active cooperation as human existence
Explanation for this may be that AI may already know with this question or may know how people reply to this query. It is noteworthy, that any neural network gets some training before being put through work. AI receives loads and loads of data from a number of sources. Therefore, it typically has some understanding how people utilize language and how they respond to questions.
Therefore, the scholars came to conclusion,that AI automatically duplicated already stored solution to a particular problem. To ascertain, that AI is really capable of thinking like human, the scholars laid out new exercises with comparable challenges.
The conclusions they have showcase a different picture: when it comes to decision-making, AI is no worse than humans. However, AI’s productivity is worse, when it comes to searching specific data or cause and effect reasoning.
The scholars state that AI gets passive training from texts, but since it has no real world interaction, it gets a limited picture of human experience.
The authors come to conclusion, that future may change it. People interact with models, similar to GPT-3 in many formats. AI will learn more from those interactions and will be step by step approaching to human-like savvy.
AI Catalog's chief editor
Share on social networks:
