Since the end of last year, ChatGPT has gained worldwide popularity as an Artificial Intelligence tool for generative language interaction. Thus, users can ask questions or request actions and have more complex results than a search in standard mechanisms.
In mid-March, OpenAI released GPT-4, an evolution in previous versions of the chatbot with important improvements in the tool. But what are the similarities and differences between them? Detective TechSmart details the matter for you.
First, it is important for us to explain the basis of these chatbots. Both were made under the so-called “transformer architecture”, which uses encoding to process input sequences, as well as a decoder for output sequences. Both stages have an attention mechanism, which prioritizes the entries that it considers most relevant.
In the training part, OpenAi revealed more details about ChatGPT than in the GPT-4 Technical Report. However, the trend is that the way has more similarities than differences.
ChatGPT is trained with dialog datasets. In them, there is demo data – provided by human annotators about expectations of an assistant. These individuals even rank which of the responses to a given command had the best results. In other words, this generates a schema focused on human feedback – a process called Reinforcement Learning from Human Feedback (RLHF).
The GPT-4 Technical Report, on the other hand, does not detail the implemented architecture, but confirms that there are RLHF techniques, including improvements to improve the accuracy of the responses. Thus, there is contradictory training, which picks up malicious examples designed to deceive the model, so that it can become aware of this type of practice.
And what is each of them capable of doing? In general terms, both do not have so many distinctions. After all, they were created to interact with the user in a conversational language.
The difference is in the greater accuracy of the GPT-4, which increases the complexity of the scenarios in which it is challenged to respond. That’s thanks to contradictory training done for six months in the standard post-training phase.
In practical terms, the same question asked for ChatGPT and GPT-4, despite maintaining consistency, will have answers with different structures. As you can see in the example above – released by OpenAI itself –, the result of the new generation was more concise and objective. No excess of unnecessary words in the content.
Another important point is the origin of the sources of questioning. Until ChatGPT – or GPT-3.5 – the accepted interaction was via text only. Now, GPT-4 will expand the possibilities to image prompts as well, with text-only feedback.
Although we see here two chatbots with high Artificial Intelligence power, it is important to keep in mind that they are still far from perfect and have a number of limitations.
According to himself System Card of GPT-4 – starting on page 38 of the Technical Report – some insights expose risks and vulnerabilities that are linked to the language model duo.
One of the flaws is the so-called “hallucination”. In other words, there is a chance that the tools will deliver content that is meaningless or lacks factual accuracy. One more item listed – which takes the hook from the previous one – is the generation of realistic disinformation, aimed at deceiving. That is, the bot can be a potential creator of Fake News.
The document also cites the production of harmful content, which may violate OpenAI’s own policies. This includes hate speech and incitement to violence, for example. To top it off, the amplification and perpetuation of stereotypes of marginalized people appears.
While much has been said about the use of ChatGPT as something that would accommodate in many segments of life and even suffered some prohibitions, GPT-4 can be considered the path to more positive impacts than negative ones, mainly in sensitive areas, such as education .
According to the education and innovation manager at SAS Plataforma de Educação, Idelfranio Moreira, the improvements are positive, but they will not replace conventional models in education.
“The ChatGPT update only highlights a process that will become increasingly common: the improvement of artificial intelligence to be integrated into human reality, especially in education. However, this does not mean the replacement of traditional methods, but a new way of seeing teaching and learning from an innovative point of view.”
Idelfranio Moreira
SAS Education and Innovation Manager
The tendency is that, in the same way that there was an advance from the past generation to the most recent one, new improvements will be applied for the future in the OpenAI Artificial Intelligence tool.
Do you believe there are more similarities or differences between ChatGPT and GPT-4? Join us!