Don’t ‘divine’ ChatGPT
Editor’s note: ChatGPT continues to be a topic of interest to the community in recent days. VietNamNet newspaper would like to introduce the article of Dr. Nguyen Huu Huan – University of Economics Ho Chi Minh City – member of the research and development team of Milu Vietnam chatbot – so that readers can have more perspective on computer programs using intelligence this artificial intelligence.
Don’t “divine” ChatGPT
Chatbots Artificial intelligence, ChatGPT developed by OpenAI, is breathing life and work of mankind in the world. Although chatbots or artificial intelligence robots are not new and have been researched and developed by many big technology companies for a long time, ChatGPT has a different direction when chatting like a human, with a brain that can answer. answered almost all questions in every field (although the error rate was still high).
However, the outstanding advantage of ChatGPT is quick reply, natural conversation. It is like a virtual assistant that supports text editing, searching, analyzing information, creating new ideas, even composing music, writing poetry…
This chatbot is a collection of a complex series of NLP (Natural language processing) algorithms and the structures of artificial intelligence models such as Neural Network (artificial neural network); Reinforcement learning or transformation algorithm. Thanks to the above combination, ChatGPT can think and chat like a real person with an infinite store of knowledge. Chatbots can self-learn based on available data sets or on the network without having to be trained like chatbots of old algorithms.
Thanks to the ability to self-study (unsupervised learning), ChatGPT will become more and more intelligent and perfect as the data set increases through more and more users interacting. However, more data is not good, more correct data is good. Currently, ChatGPT users often give false information, to test its intelligence, confuse ChatGPT, which can also create misinformation. OpenAI of course will have to use filters, check the data before systematizing the correct information for ChatGPT to learn.
From another perspective, we should not “divine” ChatGPT because the current rate of incorrect answers is quite high. The limitation of the current machine learning model is that it aggregates a huge amount of information, then, when a user asks a question, ChatGPT will aggregate relevant content while its data, producing results. words in natural language.
In case an uncommon question is asked or a data warehouse is not available, ChatGPT will return the same result. Of course, half-truths are not facts. For example, when asked ChatGPT, who are Quang Trung and Nguyen Hue? This chatbot confused quite a lot when it came to Quang Trung as Nguyen Nhac or Quang Trung as the winner of the war with his own Tay Son army…
The founder of ChatGPT himself also admitted, most of its answers are wrong.
Trade-offs in artificial intelligence
Over the past 2 years, our research team has also developed the Milu chatbot on top of OpenAI’s GPT3.5; Facebook’s BlenderBot; Google LaMDA. In fact, it is too difficult to access new algorithms and many barriers. For hardware, to run AI (artificial intelligence) models, the system is very expensive. Imagine, for a brain to be smart and process information well, in addition to brain structure (algorithm), we must have a very strong infrastructure for it, to have a super AI with models. Complexity requires a corresponding supercomputer. Investment resources are limited, so instead, we try to structure the algorithm optimally to save resources as much as possible.
Milu is not inferior to ChatGPT in terms of processing capacity and speed, users do not need to use VPN or virtual network, foreign phone number to register because this is a Vietnamese application. However, domestic chatbots are still standing on the shoulders of “tech giants” in terms of data.
From the actual experience of the research team, we see that ChatGPT is doing very well, but in order to provide accurate information analysis, ChatGPT is far from reaching the capacity of Google. Note, if you become a chatbot stereotyped in the form of a Google search engine, the information will be accurate, but it will lose creativity.
AI developers have to trade off accuracy and creativity in a chatbot. This is like humans with the left brain representing accuracy, logic and the right brain representing creativity and thinking. The stronger part of the brain, the person will lean more in that direction, the rest will be weaker.
To create a chatbot that is both complete, currently, researchers are offering many solutions such as the AI check AI model, that is, after the AI answers, there is another logically strong AI that will check the available information. exactly no, from there, give the final answer. Another solution is to reduce the approximate range of the AI, which limits creativity, so the answer will often be “I don’t know” when the chatbot can’t find the information.
In addition, another way is to combine the chatbot with an external knowledge database to ensure that the answer information is accurate. However, building such a large, prestigious database takes time and money. To build ChatGPT, investors have poured billions of dollars into the research and development process, if building an accurate and detailed database for AI easy to understand and update information, a good amount of money. The extra time it takes is much more expensive.
Therefore, if ChatGPT will become a super AI in the future, a metadata database like the one above is needed or the algorithm will be the deciding factor. Must solve the trade-off problem between accuracy and creativity, forming a super intelligence that can answer and fulfill all human requests. With this race, Google, Facebook as well as large technology corporations are certainly not outside. OpenAI has the advantage of being the first, but who is the king will have to wait for the answer.
Dr. Nguyen Huu Huan – University of Economics Ho Chi Minh City
VietNamNet Newspaper would like to thank and invite readers to continue to participate in writing articles to share about ChatGPT. Articles should be sent to: [email protected].