It is different from the artificial intelligence of other types of technical types，Genesis artificial intelligence has realized artificial intelligence from perception and understanding the world to generating the creation world.，Represents the paradigm transformation of artificial intelligence technology research and development and landing applications。The big model brings "new paradigm、New Era and New Opportunity "，Human is from data pan -general、The algorithm is pan -to the "New Age" in the model of the model。OpenAI chief scientist Illia Suzkwei will call the generative pre -training model as "World Model"。He believes that Internet text is the mapping of the world，Large natural language models learned the entire world through learning a large amount of Internet text。"World Model" can generate everything，Whether it is text、Picture、Audio、Video，Still code、Scheme、Design, etc.，All with humans、The mapping of all things in the world that is closely related to life can be learned and generated。Just as the power generation station in the electrical era caused the outbreak of power applications，Information Age Personal Computer has caused an outbreak of operating systems and Internet applications，The birth of the generation of artificial intelligence will lead artificial intelligence technology to reach a strange point，and cause the outbreak of artificial intelligence applications。Artificial Intelligence 2.0 era dominated by the generated artificial intelligence，The inner technical mechanism will be on the data layer、Legal governance of the algorithm layer and the application layer brings a series of fundamental and systemic challenges。

bet365 live casino games  

Data is the basic nutrition for artificial intelligence model training and optimization，is the "foundation and life pulse" of big model training。Compared with traditional small and medium models，Large natural language model consists of 100 billion -level parameters，Need a large scale、Diversified text data develops pre -training。This makes its requirements for data collection and processing higher，This has also triggered a series of governance risks。

First，Pre -training links need to use knowledge and information in various fields that can be obtained.，Although the diversity and coverage of the data is added，Improving the model to deal with various languages、Domain and theme ability，But this data contains the personal information of massive users and even minor users。Model training through crawling public data has become an industry practice，But it may cause huge privacy risks。This risk will spread to the downstream scene，Data deletion measures afterwards are not available。The existing privacy remedy plan represented by deletion right is difficult to cope with the continuous privacy invasion effect brought by artificial intelligence models。Training data has already become a permanent mark for machine learning model，"Algorithm Shadow" will continue to exist。

Next，In deployment link，Since the big model does not have the characteristics of "final training"，This means that the data generated by the user using a model or the application containing its plug -in will be iterated to the model optimization link。For example，In the vertical deployment and plug -in mode of large models，Click when the user interacts with the model、Input、Use、Mobile、button、Search and geographical bet365 live casino games location information will be integrated into the model iteration in real time，Beyond the coverage of the current personal information protection system，The effective problem of triggering the "arrow shooting target"。

Last，In the operation link，Large model training data is difficult to solve technical security risks。Large natural language processing models will have more training data than small and medium models.。When the artificial intelligence model becomes larger and larger，Its vulnerability will be more prominent，The problem of privacy leaks becomes more common。A recent study shows，Modify through machine learning inference，Extracting attacks on training data can easily collect sensitive information and steal intellectual property。This can be seen，Compared with medium and small models，The technical dependence of the big model on the data makes it face a more difficult data governance challenge。

Legal governance challenges of the algorithm layer 

First，At present, most artificial intelligence systems are based on deep learning，This leads to a strong homogeneity feature in the field of artificial intelligence。The existing large natural language models are mostly generated based on the Transformer architecture，This makes this type of model show significant homogeneous characteristics in terms of technical characteristics。Although homogeneity can help large -scale production，But it may also cause all the artificial intelligence systems based on these models to have the same potential risks and problems。The current artificial intelligence treatment is mostly from a single algorithm and model as the perspective，The ability of systematic and ecological governance has not yet been available。This makes the homogeneous risk of large models a blind spot for supervision and governance，The effective response to such ecological risks is particularly difficult。

2，Massive parameters and scale of large models make it emerge in multiple abilities，This emergence makes it difficult to predict its performance，It is difficult to respond to the static governance method based on risk assessment。Big model training is based on the same or similar architecture training，Use the same strategy when processing text，It may cause shared prejudice or blind spots，Cross -cultural、cross -domain、Chain risk of cross -background。

third，Once the algorithm discrimination is discriminatory，It is more challenging to relieve and governance。Study shows，Machine learning and natural language processing model embedded in pre -training words can reflect and spread social prejudice in the training data corpus，Using such machine learning methods may further expand prejudice and discriminate against users，In particular, it may derive greater risks for disadvantaged groups。Compared with traditional models，Discrimination of generating artificial intelligence may be hidden in a certain character、Text fragment or summary and interpretation of specific sentence segments，Therefore, it is even more hidden。It cannot update its training data in real time due to the generated artificial intelligence，There is a lag that may cause it when it generates the view of the current events and trends，Exhausting historical data on the adverse effects of current events and trends。For this kind of prejudice，It is difficult to find effective response methods。other，From the perspective of technical mechanism，From the pre -training of large models to deployment and operation,，Discrimination and prejudice may be introduced in all links。Therefore，The governance of discrimination against the big model algorithm will become the primary difficulty Bet365 app download and challenge on the Policy agenda of the 2.0 era of artificial intelligence。

Legal governance challenges of the application layer 

First，The "hallucination" problem of large natural language processing models has no targeted good response strategies。The so -called "machine hallucinations"，refers to the content of the big model that generates meaningless or is not faithful to the source input information。Corresponding to the application layer，means a large model generated false or meaningless content。Machine hallucinations may be due to the deviation of the data set，It may be over -fit due to the model、The prompt is wrong or represents imperfect learning、Decoding errors and the problems of preferences such as priority parameters。

Next，The industrial chain of generating artificial intelligence has the characteristics of "model is service"，This causes any risk in the upstream to spread in series to midstream and downstream。Artificial intelligence technology paradigm changes while breaking through the bottleneck of the traditional industry，It also broaden the connotation and border of the "platform"。This means that platform companies in specific applications may no longer have definition of algorithm model、The final influence of design and deployment，The main responsibility of the algorithm adapted with it、A series of system design such as algorithm risk monitoring may be on the end。

Last，Compared with small and medium models，Insufficient security and reliability of big models。For example，Study shows，When Twitter、Network blog、News reports and research papers inject the false information of "water is toxic" into the large model，Large models can generate rational fiction text。The false information accepted by the model will also be scattered in the application with the help of the semantic diffusion process，Memory related to the pollution model。This indicates，False information can threaten and penetrate the entire life cycle of large models，Even if the training data is credible and correct，False information has produced a global negative effect，instead of only related interference information。Governance of this level，Although my country、EU and other regulators clearly require large model training data collection laws、Accurate and true，But the complex features of the model mechanism determine that the treatment of focusing on the data layer is difficult to effectively radiate to the application layer。

The "critical moment" of the technology of technology and the "critical moment" of governance 

Current，Chinese technology companies in chips、Equipment and other hardware is limited，Frequency frequency and volume also have a certain gap with foreign technology companies。When domestic technology companies are devoted to the development of artificial intelligence development、Actively deploy and invest a lot of resources，Adhering to the concept of two -wheel drive of innovation and governance，Actively exploring the legal governance framework with it has important theoretical and practical significance。Genesis artificial intelligence not only has the potential to empower thousands of industry，It is also related to national security and social order。The economic and public value of artificial intelligence determines that my country's artificial intelligence governance cannot be limited to the technical surface，It is necessary to jump into an important part of national governance and social governance in the process of interacting with national security and social order。It can be said，Genetic artificial intelligence not only brings "Wonderful moment" of artificial intelligence technology，It also brings the "critical moment" of artificial Bet365 lotto review intelligence governance。

From the perspective of the operating characteristics of the generated artificial intelligence，EU、The existing algorithm governance framework in the United States and our country may all show the limitations of governance in different dimensions。The core reason is，The current mainstream artificial intelligence governance is mainly for traditional artificial intelligence models，with a general potential、A new generation of artificial intelligence with large models as cores is difficult to fully adapt。For example，The EU establishes a risk -based artificial intelligence governance framework in the draft of the "Artificial Intelligence Act"，After evaluating the artificial intelligence system, it is divided into minimum risks、limited risk、Four levels of high risk and unacceptable risks，Differentiated supervision of each level。When faced with generalability、Cross -mode、When the emergence of generating artificial intelligence，Risk -based governance paradigm may encounter failure risk。This risk -based governance paradigm needs to be risk -leveling the artificial intelligence system according to the application scenario，When the emergence of generating artificial intelligence。The application of generating artificial intelligence has dynamic characteristics，It is the overall reconstruction of the artificial intelligence value industry chain。The existing risk classification method is difficult to automatically convey the dynamic conversion between categories with the extension of the generation artificial intelligence technology，It is difficult to use the method of classification of high -risk fields to play the governance efficiency of front planning and continuous monitoring。

For a long time，The United States is driving in autonomous driving、Algorithm recommendation、Face recognition、Deep synthesis and other applications Layout related legislation，A paradigm of artificial intelligence governance based on application scenarios。But，Facing the generated artificial intelligence，It also highlights multiple limitations。Genesis artificial intelligence has infrastructure status，Many artificial intelligence generation content industry application scenarios are many，It is difficult for upstream and downstream developers to control the risk of the entire system。Applicable artificial intelligence governance focuses on downstream，It is difficult to effectively radiate upstream and midstream technology applications，It is more difficult to apply effective governance to the entire ecology of generating artificial intelligence。

different from the governance path of the European Union and the United States，my country's governance of artificial intelligence relies on the main responsibility of the algorithm of the algorithm is gradually expanded。Algorithm accountability system based on the main body of the platform also needs to be changing。"The main responsibility of algorithm" is mainly based on algorithm recommendation service providers and deep synthetic service providers，Requires it to actively fulfill the obligation of active and inaction，Ber the corresponding unfavorable consequences when performing poorly。But in terms of generating artificial intelligence，Its design and operation links may bear the entity of "main responsibility" to be diversified、Diversified、Dynamic、Scenic characteristics，It is difficult to simply delineate the boundary of the responsibility to assume the subject of the subject。The risk of the big model may not only come from the R & D person，It may also be derived from the deployee and even the end user。Deploy in the big model、Among the three core operating modes based on application interface calls and plug -in -based development，The subject that should bear the responsibility has both overlap and difference。

From this to it，Whether Bet365 app download it is risk -based governance，It is still based on the treatment of subject and application，All in the development stage of artificial intelligence -specific models as the underlying architecture。In the "General Model" era，Facing the strong generalization capabilities of its display，and a new pattern of large -scale collaboration deployment of upstream and downstream industries，The mainstream governance paradigm may face different degrees of challenges。In response to the development of artificial intelligence, the development of artificial intelligence，The forward -looking layout and accelerate the artificial intelligence governance framework that is adapted.。

Anxiety and concerns of the loss of control of artificial intelligence development rate，Legislators of various countries respond quickly，Try to iterate and innovate the existing governance framework。July 10, 2023，The Interim Measures for the Management of Genesis Artificial Intelligence Service Management (hereinafter referred to as the "Interim Measures") will be announced to the society，Become the first legislation for generating artificial intelligence in the world。Full cycle and full chain for generating artificial intelligence，"Temporary Measures" is trained for the legitimacy of training data、Artificial labeling specifications、Make clear regulations for generating content reliability and security management obligations，Collaborative linkage with the "Recommended Management Regulations on Internet Information Service Algorithm", "Internet Information Services In -depth Synthetic Management Regulations"，The "three -driving carriage" that constitutes the governance of artificial intelligence in my country。my country's artificial intelligence governance is gradually governing from factor、Platform governance towards the governance of the artificial intelligence industry facing the value chain。But，With the subversive development of artificial intelligence technology，The artificial intelligence governance ideas embedded in the Internet information content governance structure have been difficult to cope with its parallel fusion、New technology application risks that are constantly emerging。For effective response challenges，Artificial intelligence governance needs to be in the concept of governance、Governance Technology、Systematic innovation in terms of governance tools and governance capabilities。

First，The concept of the principle of the accuracy of artificial intelligence is clear from the concept。Artificial intelligence accuracy is not only an important bottom line for generating artificial intelligence deployment and operation，It is also an important prerequisite for building trusted artificial intelligence。But，What needs to be clear is，Statistical mechanism based on artificial intelligence model，The accuracy principle of artificial intelligence under the visual threshold should not be required to output the statistical accuracy of 100%of the output of artificial intelligence。From the perspective of "Technical -Society" coupling interaction，It should face the risk characteristics of artificial intelligence，Explore the rules of risk definition of generating artificial intelligence，Revisiting the cognition of the accuracy of artificial intelligence，Explore the connection conversion mechanism between the accuracy and the accuracy of the governance。Current，Whether it is the EU's "Artificial Intelligence Act", the latest folding draft，It is still the "Artificial Intelligence Risk Management Framework" from the National Standard and Technology Research Institute，All emphasized that artificial intelligence governance cannot pursue "zero risk" governance goals。Two wheel drives that intend to achieve governance and development，Need to face up to the risk threshold of artificial intelligence，Explore the precise governance system based on the control level。bet365 best casino games Therefore，Based on the "Interim Measures"，Data accuracy that matches the characteristics of the technical characteristics of the same generation artificial intelligence and industrial characteristics、Algorithm model accuracy、Output accuracy and explanation accuracy standard，Construct a scientific and accurate risk governance framework。

2，Explore creating innovative and friendly regulatory infrastructure，Regulatory technology is used as a linkage multi -subject、Balanced technological innovation and public interest guarantee for the underlying support。For a large scale、100 billionth number、Super Extensibility and the generated artificial intelligence of super application scenarios for，Start it afterwards、Traditional supervision characterized by scattered promotion is facing response time stagnation、Insufficient elasticity、Unclear supervision scale、Challenges of high -domain coordination costs。Therefore，should be actively creating "govern AI、Intelligent regulatory technology system with algorithm system，In -depth exploration of the effectiveness of the technical model at the level of artificial intelligence supervision，Convert ethical specifications and compliance points into program language，Internal ethical technology constraints into the entire process of artificial intelligence design and operation，Explore intelligent supervision paradigm。

third，Exploring and constructing a classified regulatory tool system，Create a bottom grip for precise dynamic supervision。Combined with the development characteristics of the artificial intelligence industry in my country，Classification of artificial intelligence models from technical complexity、Classification classification of artificial intelligence applications from legal risk and classification of artificial intelligence enterprises from legal subject categories。Artificial intelligence supervision with classified classification as the kernel，Reflecting the precise capture of the artificial intelligence application scenarios and the precise adaptation of information and regulatory tools，It is the precise refinement and scientific combination of facts and specifications。Under the classification and hierarchical supervision framework，It should also explore the impact assessment of the construction algorithm、Algorithm safety certification、Algorithm audit、Algorithm filing four -in -one regulatory tool system。

Fourth，Comprehensive improvement of regulators' sensitivity to artificial intelligence technology and technical regulation ability。"Level" management mode does not match the application logic of the decentralization of artificial intelligence，Regulatory disorderly competition and regulatory vacuum chaos often occurs alternately。With the development of artificial intelligence technology, entering the fast lane，Artificial intelligence generation content technology due to its rapid response ability、Rich knowledge output、Multiple application scenarios, etc.，Make a challenge to the supervisor's supervision ability。Supervisor's understanding of artificial intelligence technology and cognitive efficiency and response disposal ability directly affects the scientificization of supervision、Refined、Agility。Therefore，It should fully absorb professional and technical talents to enrich the regulatory team，Change supervision thinking，Efficient information sharing mechanism between creating regulatory agencies，Enhancement matching with artificial intelligence technology innovation、Compatible response ability and management ability。

Study from a large model to large application: explore the way of artificial intelligence cure 

Although the generation of artificial intelligence technology already has bet365 best casino games a number of stunning abilities，But there is still a series of technology and application limitations。The current artificial intelligence training technology can only achieve model generation，but cannot expand model editing。How to promote the transformation of "big model" to "big application"，Successfully realized large -scale commercialization，Become a key test and challenge to become the future。The realization of this goal requires the continuous improvement of the pre -training large model，On the other hand, it is necessary to think deeply about legal response。The governance of artificial intelligence is not the day and night，The governance of emerging technology cannot float on the surface、Go forward in isolation。On the basis of deep understanding of artificial intelligence technical mechanisms，Legal logic、Industrial logic and technical logic deep fusion，Deep into the underlying architecture of the artificial intelligence industry，Explore the way to treat the root cause。

(Author is an associate professor at the School of Law of the University of Foreign Economics and Trade)