Text | Create an overview of Pablo Andres in Paris
Edit |Meister Xia
guide reading
The development of artificial intelligence is changing with each passing day, and generative AI is one of them. It can create all kinds of words, images or sounds according to the input instructions. What is the principle of this technology? What are its applications and limitations? What challenges and risks will it bring? This article will introduce you to the basic concept, latest progress and future trend of generative AI, and let you know this innovative field that is changing our society.
At a glance:
- Generative AI refers to artificial intelligence that can learn the information in the database and generate words, images or audio according to the tips input by users.
- Generative AI is still in the toddler stage, and there are many areas that need to be improved, such as unreliable or biased answers.
- ChatGPT and Bloom belong to the application model of generative AI. The potential application of Qian Qian has yet to be explored.
- Any technology has advantages and disadvantages. Generative AI consumes a lot of energy and is at risk of being abused.
ChatGPT has become a hot topic in today’s society, and a new round of AI technology revolution has begun to emerge. Generative AI has seemingly endless potential application scenarios, and it has not surprisingly triggered endless debates. However, the public’s understanding of the principle of generative AI seems to need to be deepened.
Generative AI refers to artificial intelligence that can learn the information in the database and generate words, images or audio according to the tips input by users. Eric Moulines, a professor of statistical machine learning at Université Polytechnique de Paris, explained: "In the process of learning, AI will summarize laws from data and generate original content based on this."
At present, the two mainstream artificial intelligence models are GPT (Generative Pre-training Converter) and diffusion model. Hatim Bourfone, an artificial intelligence research engineer at the Institute of Intensive Scientific Computing Development and Resources (IDRIS) of the French National Centre for Scientific Research (CNRS), added: "Artificial intelligence will understand the contextual meaning of the input text through a’ attention mechanism’. The content of its output consists of the vocabulary it learned in the training stage before, and AI will judge which word to use next according to the probability of each word. " With different database training algorithms, AI suitable for different scenarios can be obtained.
Bourfoune’s team participated in the development of an AI called Bloom, which is an academic paper translator for researchers. Pierre Cornette, an IDRIS researcher, said, "The main task of the Bloom model is to learn many foreign languages. We input a lot of text into it, and then let it judge which word should appear below according to the previous content. If something goes wrong, we will correct it. "
New technology of toddler.
Moulines explained: "The history of the first generation of generative AI models is less than ten years. The first breakthrough of generative AI technology occurred in 2017: the advent of converter technology improved the attention mechanism. By 2021, commercial generative AI will be listed, and the speed is staggering, much faster than other deep learning models. " Nevertheless, we must realize that AI such as ChatGPT is still in the toddler stage, and there are many places to be improved.
Moulines admits that the credibility of the answers given by GPT is not high enough, and it is still a problem: "ChatGPT doesn’t know what’ credibility’ is and doesn’t know how to evaluate the accuracy of his answers." That’s why sometimes ChatGPT "talks nonsense in a serious way". "Because ChatGPT generates word sequences purely based on probabilistic reasoning, it will generate seemingly credible but actually false content."
In addition to "making up", there are some other defects of generative AI that need our attention. In the process of deep learning, AI will absorb a lot of existing texts and internalize their prejudices. Moulines said: "If you ask ChatGPT geopolitical questions, the answers are all the positions of western countries. China users will definitely not agree with such an answer! "
02 endless potential applications
The charm of generative AI is that it can develop endless models with different functions by using diverse learning databases. Cornette said: "Generative AI is like a high-powered engine, which can be mounted on a tractor to maximize its traction, and can also be installed in a racing car to let it fly by." If ChatGPT is compared to a racing car, GPT-4 is its engine. "The engine is the core technology. Drivers don’t need to know the principle of the engine, but they can also go fast on the field. "
Bloom can also reflect the extensive application potential of generative AI model. Bourfoune said: "A year ago, Bloom was the only model that was completely open to academic circles." Anyone can download Bloom and use it for their own research. After the training of multilingual scientific papers database, Bloom can now help scholars understand foreign language papers easily. Cornette added: "Bloom’s development team also launched a project called Bigcode for automatic generation of computer code. As long as the function of the code is simply described, Bigcode can write specific code in the programming language specified by the user. "
ChatGPT is very popular now, which shows that ordinary users have realized its practical value. In order to compete with Google, Bing has integrated the chat function of GPT into its search engine, which can overcome the shortcoming of "nonsense" of generative AI to some extent: the source of information will be marked in the answer given by Bing Chat, which is convenient for users to understand and verify the reliability of the content. Recently, Adobe integrated the generative AI model into Photoshop, Illustrator and other software, showing another novel application.
03 exciting future
Judging from the current application, generative AI will surely usher in an exciting future, but some people worry that this technology may be abused. Bourfoune admits: "Any technology has its advantages and disadvantages. This is why OpenAI has set up multiple security barriers. " OpenAI’s content policy also takes these security factors into consideration, so OpenAI chooses to remain silent on many issues concerning the operation of ChatGPT.
For the generative AI technology that is still in its infancy, Moulines said: "We clearly know that technology is still at the starting line in our research. We are all surprised that the generated AI can be used in practice. " However, there are still many gaps in technology such as legal supervision to be filled. Since generative AI generates content based on the existing content database, it may "plagiarize" other people’s works without mentioning the original author’s name. "To create a new work with existing content, you must declare the original source. AI’s behavior is suspected of infringement. "
Although generative AI has various limitations, its potential is still huge. Moulines said: "I am very excited to think of possible breakthroughs in this field in the future. The development of generative AI is unstoppable, and derivative applications will appear like mushrooms after rain. Now everyone is scrambling to develop new technologies and making rapid progress. " Bloom is a derivative application, which can not only promote cross-language communication between scientists, but also translate papers into rare small languages, promote the dissemination of scientific research results, and is expected to be used to preserve endangered languages.
However, in the excitement, we can’t ignore the carbon footprint of generative AI. Moulines explained: "These models need to store a lot of data, so they need a lot of memory. According to our estimation, the energy consumed by OpenAI is equivalent to the energy consumption of the entire national grid in Belgium. " In the future, energy consumption may be the biggest obstacle to the development of generative AI.
关于作者