In 2023, the world witnessed the global fire of ChatGPT. The introduction of a new generation of AI represented by generative AI has changed the trajectory of artificial intelligence (AI) technology and applications, accelerated the process of human-AI interaction, and is a new milestone in the history of AI development.What trends will the development of AI technology and applications show in 2024? Let's look forward to these noteworthy major trends together.
On March 10, 2023, Nanjing National Pilot Zone of Artificial Intelligence Innovation and Application was inaugurated.
Trend 1: Moving from AI macromodels to generalized AI
In 2023, ChatGPT developer OpenAI was put in the spotlight like never before, and has put the development of a successor to GPT-4 in the spotlight. According to sources.OpenAI is training the next generation of artificial intelligence, tentatively called "Q*" (pronounced Q-star). The next generation of OpenAI products may be released in the new year.
According to the media blitz, "Q*" may be theThe first artificial intelligence trained from scratch.. It is characterized by intelligence that does not derive from human activity data and the ability to modify its code to adapt to more complex learning tasks. The former makes the development of AI capabilities increasingly opaque, while the latter has always been seen as a necessary condition for the birth of an AI "singularity". In the field of AI development, the "singularity" refers to the ability of a machine to iterate on itself and then grow rapidly in a short period of time, leading to a situation that is beyond human control.
Although some reports say that "Q*" can only solve elementary school math problems and is still far from the "singularity". However, given that the speed of AI iteration in virtual environments is likely to be far faster than imagined, it is still possible that in the near future it will be able to autonomously develop AI that can outperform the human level in various fields.In 2023, OpenAI predicts that AI that surpasses the human level in all aspects will be available within a decade; NVIDIA founder Jen-Hsun Huang says that generalized AI could surpass humans within five years.
Once general-purpose AI is realized, it can be used to solve a wide range of complex scientific challenges.Examples include the search for aliens and extra-terrestrial habitable galaxies, the control of artificial fusion, the screening of nano or superconducting materials, and the development of anti-cancer drugs. These problems usually take human researchers decades to find new solutions, and the amount of research in some cutting-edge fields has exceeded human limits. Whereas a generalized AI has almost unlimited time and energy in its own virtual world, which makes it a potential replacement for human researchers in tasks that are partly easy to virtualize. But then, how human beings can supervise these AIs that exceed human beings from the level of intelligence to ensure that they will not jeopardize human beings is another issue worth thinking about. Of course, we should not overestimate some of the statements made by the Silicon Valley giants, because in the history of AI development, theThere have been three "AI winters", many of which are examples of grand technological visions that have fizzled out due to various constraints.. But at present, it is certain that the large model technology still has a lot of room for upward mobility. In addition to GPT-4, Google's "Gemini" (Gemini), Anthropic's Claude2, are currently second only to the GPT-4 large model, the domestic Baidu "Wenxin Yiyin" and Alibaba "Tongyi Thousand Questions", are also the best domestic large model. Baidu's "Wenxin Yiyin" and Ali's "Tongyi Thousand Questions" are also the best of the domestic big models. Whether they will release more revolutionary products in the new year is also worth looking forward to.
Trend 2: Synthetic Data Breaks the AI Training Data Bottleneck
The data bottleneck refers to the limited availability of high-quality data that can be used to train AI, and synthetic data is expected to break this bottleneck.
Synthetic data is data that is synthesized by machine learning models using mathematical and statistical science principles based on mimicking real data. There is a relatively easy to understand analogy for what synthetic data is:It's like writing specialized textbooks for AI.For example, although the dialogues in an English textbook may contain fictitious names such as "Xiaoming" and "Xiaohong", it does not affect the students' mastery of the English language, so in a sense, the textbook can be regarded as a kind of "synthetic data" that has been compiled, filtered, and processed by the students. Thus, in a sense, for students, textbooks can be seen as a kind of "synthetic data" that has been compiled, filtered and processed.
It has been shown that the scale of the model has to reach at least 62 billion parameters before it is possible to train the "chain of thought" capability, i.e., to perform step-by-step logical reasoning.But the awkward reality is that there is not as much non-repeated, quality data generated by humans to date that is available for training. Using generative AI such as ChatGPT to produce high-quality synthetic data in unprecedented quantities, future AIs will achieve higher performance as a result.
In addition to the demand for large amounts of high-quality data that has led to the pursuit of synthetic data, considerations of data security are also important. In recent years, countries have introduced stricter data security protection laws, making it more cumbersome to objectively train artificial intelligence using human-generated data. Not only may personal information be implicit in this data, much of it is also protected by copyright.At a time when Internet privacy and copyright protection have not yet formed a unified standard and perfect structure, the use of Internet data for training can easily lead to a large number of legal disputes. If we consider desensitizing these data, we will face the challenge of screening and recognition accuracy. The dilemma is that synthetic data is the most cost-effective option.
In addition, training with human data can lead to the AI learning harmful content. Some, such as the use of everyday objects to make bombs and regulate chemicals, and others include many bad habits that AI should not have, such as laziness in task execution, lying to please users, prejudice, and discrimination, like humans.If synthetic data is used instead, so that the AI is trained with as little exposure to harmful content as possible, it is expected to overcome the above drawbacks incidental to training with human data.
From the above analysis, it can be seen that synthetic data can be said to be quite groundbreaking, and is expected to solve the previous problem of the development of artificial intelligence and data privacy protection can not be reconciled. But at the same time.How to ensure that relevant companies and institutions produce synthetic data responsibly, and how to produce synthetic data training sets that are compatible with national cultures and values, and comparable in scale and technology to those available in the West centered on English-language web materialsThe Chinese government will also be faced with a challenging topic. In addition to this, one of the major changes brought about by synthetic data is that theBig data from human society may no longer be necessary for AI training. In the future digital world, the generation, storage, and use of human data will still follow the laws and order of human society, including maintaining national data security, keeping commercial data secret, and respecting personal data privacy, while the synthetic data required for AI trainingAdoption of another set of standards for management.
On September 17, 2023, the 2023 Nanjing Artificial Intelligence Industry Development Conference, jointly organized by the Nanjing Municipal People's Government and the China Academy of Information and Communication Research, opened.
Trend #3: Quantum computers may be the first to be used in artificial intelligence
As the most cutting-edge application in the development of electronic computers today, AI has always had the worry of a lack of arithmetic power.A few months after the introduction of ChatGPT, OpenAI President Ortman publicly stated that it was not encouraging more users to sign up for OpenAI.2023 In November, OpenAI even announced that it was suspending the registration of new users of the ChatGPT Plus paid subscription to ensure that existing users have a high-quality experience. Clearly, as the most powerful AI in the world, theChatGPT has encountered bottlenecks in terms of arithmetic power and other aspects. In this context, discussing the application of quantum computers in AI becomes a promising future solution.
First of all, most of the algorithms in the field of artificial intelligence belong to the category of parallel computing. For example, when AlphaGo is playing Go, it needs to simultaneously consider the opponent's response moves after landing in different positions, and find the move that is most likely to win the game. This requires the computer to optimize the efficiency of parallel computing to achieve. Quantum computers are good at parallel computing because they can calculate and store both "0" and "1" states simultaneously without consuming additional computational resources as electronic computers do, such as connecting multiple computational units in series or juxtaposing computational tasks in time. The computer does not need to consume additional computational resources as an electronic computer does, such as by connecting multiple computational units in series or by juxtaposing computational tasks in time.The more complex the computational task, the more advantageous quantum computing becomes.
Second, the hardware required to run ChatGPT, againis also well suited to be imported into the current bulky quantum computers.Both need to be installed in highly integrated computing centers, supported by a team of specialized technology management.
What is a quantum computer? Quantum computers are physical devices that follow the laws of quantum mechanics to perform high-speed mathematical and logical operations, store and process quantum information. It is not only large in size, but also as the core components of the "quantum chip", usually need to be placed in the very low temperature close to absolute zero (-273.15 degrees Celsius), the use of this very low temperature in some of the microscopic particles show the quantum characteristics of the information operation and processing, and the results of the operation can only be present for a few milliseconds of time.
If quantum computers are "big and hard to maintain", why are they being developed? The reason is.Quantum computers contain so much arithmetic potential that some algorithms have already demonstrated "absolute crushing" in terms of speed relative to electronic computers, i.e."Quantum Superiority."But realizing "quantum superiority" is only a starting point. But the realization of "quantum superiority" is only a starting point. The current quantum computer can only complete some of the specialized quantum field of computing tasks, want to really use this "quantum superiority", first of all to make its quantum bits enough to realize the general-purpose computing and programmable. Moreover, theAfter general-purpose computing is achieved, quantum computers still need to maintain an advantage over electronic computers, which is called "quantum dominance."The
In 2022, researchers from Google, Microsoft, Caltech and other institutionsProof of principle that "quantum dominance" does exist in predicting observable variables, quantum principal component analysis, and quantum machine learning... Quantum machine learning, which is actually quantum computing applied to artificial intelligence, also reflects the future trend of merging the two cutting-edge technologies of quantum computing and artificial intelligence.
Theoretically proven, the practical need to further expand the application prospects of quantum computing. After launching the commercial quantum computer "Quantum System One" in 2019, the U.S. quantum computing giant IBM launched "Quantum System Two" in December 2023.The biggest breakthrough of the new system is the possibility of modular expansion.It is the company's first modular quantum computer. "Quantum System Two has more than 1,000 quantum bits, and IBM has announced plans to build a 100,000-qubit quantum computer within 10 years. These ever-increasing quantum bits are not just for the race; they are essential for general-purpose computing and programmability. And because of this, theThe modularity of quantum computers marks them as more practical.Research on quantum machine learning algorithms has become a new research hotspot. However, future quantum computersIt will not completely replace electronic computers, it is more likely that quantum computers and electronic computers in different application scenarios to play their respective strengths, to achieve synergistic developmentThe first is to improve arithmetic power significantly while balancing cost and feasibility.
At the 2023 World Intelligent Manufacturing Conference held in Nanjing, a separate industrial robotics exhibition area was set up for the first time, focusing on nearly 300 leading robot products such as welding robots and heavy-duty robots.
Trend #4: "Shockwaves" from AI Agents and Code-Free Software Development
In terms of AI applications, the "shockwave" of AI agents and code-free software development is something to watch out for in 2024.
One is the impact of AI agents on the structure of the labor force.
By now, at least 200 million people around the world are using the big models of artificial intelligence. But people are no longer satisfied with sitting in front of the computer and "chatting" with AI, but have begun to developTools that can automatically send prompts to the AI as needed for a task.. When the auto-prompting tool is combined with the big model two, the AI agent is born.
In April 2023, OpenAI co-founder Brockman gave a live demonstration of GPT's "automated mode". In the demonstration, the AI agent almost "organized" a dinner party: not only generated a recommended menu for the dinner party and a graphic invitation according to the requirements, but also automatically added the ingredients to be purchased for the menu to the shopping cart of the fresh food e-commerce app, and automatically posted a social networking post about the dinner party.
AI agents can also automatically create websites based on relatively vague demand prompts, automatically complete a variety of text and table processing work that requires the use of Office software, and even automatically generate analysis papers based on the existing thesis data summarization, and so on.
Bill Gates recently issued a long article to explain the future of AI agents, said AI agents will revolutionize the way people use computers, bringing since the invention of the keyboard, screen and mouse on the most significant innovation in the way humans interact with computers.
AI is seen as an expansive tool for augmenting human information gathering, analysis and processing, enabling new levels of human work. At the same time, however, AI agents are impacting many existing jobs, theBecause businesses may try to hire fewer people to accomplish the same tasksThis destruction of existing economic structures brought about by innovation has been called "creative destruction" by the American economist Schumpeter. This destruction of existing economic structures brought about by innovation has been called "creative destruction" by the American economist Schumpeter. As AI agents replace a large number of tasks that require fewer computer skills, this forced re-employment of the workforce will have to adapt to new labor market demands, which is destined to be a long and painful process.
The second is the impact of code-free software development on innovation in the digital economy.
While generative AI may eliminate a number of traditional digital jobs, it closes a door and opens a window to "code-free software development". Programming aids based on AI macromodels have already reached a new stage, capable of generating software or web code based on very vague instructions from the user. For example, in the GPT-4 demo in 2023, the demonstrator simply handwrote a very scribbled schematic of the structure on A4 paper, and GPT-4 automatically generated web pages that could be physically accessed based on it. This certainly lowered the threshold for developing IT services significantly.As long as a person has a creative enough digital service "idea" that can satisfy the needs of many people, it can become a windfall of Internet innovation, and the era of "everyone can innovate" has arrived.
In this regard, governments need to change their mindset, balancing market regulation with the promotion of innovation.On the one hand, it is necessary to lower the threshold of registration and financing in the process of digital innovation, to solve the pain points in the process of SMEs' development and growth, and to adapt the employment and innovation policies to the new demand of "everyone can innovate"; on the other hand, it is necessary to explore new copyright and patent protection policies that are more conducive to the protection of innovative "ideas". On the other hand, it is necessary to explore new policies on copyright and patent protection that are more conducive to the protection of innovative ideas.This motivates those who are able to come up with innovative "ideas" on a continuous basis.
In summary, looking forward to 2024, whether it is the iterative development of AI technology itself, its reshaping of data value, or its application penetration into various industries and fields, the impact of AI can be said to be ubiquitous, empowering scientific research, innovation and the economy, while bringing new challenges and risks. We should look at the many changes brought by AI with an open mind, and cautiously study and respond to the new issues and risks it may bring.
