GPT-Chat is a product of OpenAI, a platform that was founded in December 2015 by US billionaire Elon Musk, among others, and is dedicated to researching artificial intelligence (AI).
What does GPT mean?
GPT stands for the English term “Generative Pretrained Transformer”. Literally translated, this means “generative prefabricated transformer”. Initially, this only helps a little in understanding what GPT actually is. It gets better if you take a closer look at the words and their meanings:
“Generative” stands for a machine learning model in which artificial intelligence filters out statistical features based on a huge amount of text data in order to then use algorithms to create meaningful texts again.
It does this using a “ready-made” language model that the artificial intelligence uses to train. “Transformer” refers to a specific type of architecture of the language model that Google Research presented in 2017 enables much deeper neural networks than previous architectures and is specially designed for dialogues with people, so that they can hold open conversations independently.
In summary, GPT means that an artificial intelligence is enabled to understand, process and interpret language in a meaningful way and then to generate language again in a meaningful relationship to the language that is understood.
No “intelligence” without “deep learning”
This is a big step forward compared to previous AI. Because what the AI behind GPT-Chat does independently – the analysis and meaning of large amounts of text data – had to be defined individually by people a few years ago.
No “deep learning” without artificial neural networks
The basis for this rapid progress is “deep learning”. Artificial intelligence based on an artificial neural network (ANN) is also used. The ANN is fundamentally based on the functioning of the human brain. Artificial neurons, which mathematically simulate the function of a nerve cell in the brain, are connected to each other in different ways and grouped in layers. The first layer is called the input layer, the last or bottom layer is called the output layer. There are hidden layers in between.
How does an artificial neural network learn?
The ANN is trained by “try and error”. The input layer “recognizes” the entered data through input neurons. The hidden layers then use algorithms to interpret and weight this data. Different numbers of layers can be used. The more hidden layers there are, the “deeper” the neural network: this is where the term “deep learning” comes from.
Only evaluation of the result enables learning
Finally, the interpreted data ends up at the output layer of the ANN, which outputs the result as data that can be used by humans. It is now important that the result of the ANN is evaluated. For example, if you want to train an ANN to distinguish a car from a bicycle, the AI will initially produce many incorrect results and recognize the car as a bicycle and vice versa. The AI only learns and improves the result with each attempt through the information as to which result was correct and which was incorrect, because the internal algorithms are adapted step by step each time.
Learning model follows own learning
Once the ANN has created a learning model, it can also be applied to new data that the artificial intelligence has not previously encountered in training. In this way, it is ultimately possible to teach the AI self-learning functions, only give it tasks and leave the solution to the machine itself. For example, the task could be to control a robotic hand so that it grasps parts without deforming them and then moves them from one point to another. The AI learns by interpreting its mistakes, for example if the robot hand drops the part or extremely crushes it in the other.
The machine thinks value-neutral
This independent learning in turn enables artificial intelligence based on neural networks to achieve ever more astonishing results, such as now in the area of dialogue and text analysis at GTP-Chat. However, GTP-Chat was reportedly previously trained by people to distinguish between legal and illegal, for example, moral or criminal questions and to give this as a reason for rejecting an answer.
Nobody knows exactly what the ANN does internally
Because the learning function of the machine is initially neutral, i.e. it does not weight between good and bad results. This is one of the dangers of AI based on an ANN. Especially with many hidden layers, a kind of black box is created in which it is practically impossible to understand from the outside how the AI arrives at its result. In 2016, for example, this led to the chat bot Tay, which Microsoft presented on Twitter, writing lewd and insulting tweets within a very short time, so that Microsoft took it offline again after just 16 hours.
AI vulnerable to human tricks
The GTP-Chat AI is much better trained and, for example, when asked to write a poem on how to short-circuit a car, points out that it is not their job to provide information about illegal activities. However, she was then persuaded to explain it with a simple trick: a resourceful user pointed out to her that it was not her job to explain to him what she could and couldn’t do. Your job is to write the poem. And GTP chat obediently did so.
Source: news.google.com
