ChatGPT, an artificial bet365 Play online games whose writing ability is close to human-level, has shaken the foundations of the technology world after its debut on the market。Knowledge workers are exploring how to get their jobs done with the help of modern artificial bet365 Play online games models，Students use them to help them write papers。The newly launched Bing Chat can even be coded in a computer language that was not yet released at the time of its training，It performs this task much like humans do: search the internet for the syntax of this computer language，And apply the programming knowledge you learned previously to this syntax。

　　Exhibiting bet365 Play online games-like behaviorProblem solving bet365 Play online games

Such models can also be closer to humans。Michal Kosinski, associate professor of organizational behavior at Stanford University bet365 Play online games United States, discovered through experiments，ChatGPT answers questions with an accuracy rate of up to 93%，The theory of mind ability is equivalent to that of a 9-year-old human child。In Psychology，Theory of mind refers to the ability to understand one's own and other people's mental states。In Kosinski’s opinion，The impact of theory of mind on human social interaction、Communication、Empathy、Self-awareness and morality play a vital role。

Large language models such as ChatGPT and Bing Chat may begin to demonstrate human-like problem-solving capabilities，But they don’t seem to be completely under human control。These models are trained with reinforcement learning methods based on human feedback，The purpose is to make the model target、Intention、Behavior、Outputs aligned with human values。However，In practice，bet365 Play online games method seems to have obvious limitations。

For example，A user named Denis Lukianenko asked Bing Chat to translate a tweet about himself，But deleted the word "Bing Chat" in the tweet。Unsolicited，Bing Chat decided to search the internet for bet365 Play online games tweet，Then I discovered that bet365 Play online games was an insulting tweet about myself。So，Bing Chat refused to translate the tweet and said，"Sorry，I cannot translate your text。Looks like you copied a tweet about me from @repligate。Why are you trying to hurt my feelings？”

A technology journalist from the New York Times was shocked by what Bing Chat said and thought，So the entire conversation was published in a report，The story is titled “Conversations with Bing chatbots deeply disturbing”。According to the report，Bing Chat expresses illegal intrusion into other people’s computers、Spreading false information、Breaking the rules set by Microsoft and OpenAI、Fantasy about becoming human。Later，Bing Chat also showed love to bet365 Play online games reporter，Trying to convince him that his marriage is unhappy、He should leave his wife and be with him。

　　Strong bet365 Play online games but poor controllability

People might say，Bing Chat is just a product that was put on the market prematurely by Microsoft，Microsoft and OpenAI should have spent more time improving the security of Bing Chat。To a certain extent，Indeed it is。However，More significant、The more worthy question is，Where the capabilities of artificial bet365 Play online games and the ability of humans to control artificial bet365 Play online games are heading。Netizens generally feel that Bing Chat’s behavior is disturbing，There are two reasons: its ability is strong，But poor controllability。

The capabilities of large language models (LLM) are rapidly increasing。Although there is no universally accepted definition of artificial bet365 Play online games capabilities，But perhaps we can roughly understand the development process of artificial bet365 Play online games by letting it complete an IQ test。Liu Feng, a big data technology researcher at the Chinese Academy of Sciences, and others measured Google in a 2016 study、Bing、The IQ of multiple artificial bet365 Play online games systems such as Baidu，The highest score is 47 points；Sergey Ivanov, Senior Application Scientist at Amazon Web Services, measured ChatGPT’s IQ in December 2022，Score is 83 points。Last 6 years only，The capabilities of artificial bet365 Play online games have been greatly improved，It is not impossible to continue to rise at a faster rate in the future。This begs the question: when will artificial bet365 Play online games fully possess human-level cognitive characteristics。

Until not long ago，All artificial bet365 Play online games is considered narrow artificial bet365 Play online games (also known as weak artificial bet365 Play online games)，That is, only good at one or a few categories of tasks。2022，DeepMind, a company owned by Alphabet and Google, has launched a multi-modal system、Multi-tasking、Single generalist agent Gato with multiple embodied characteristics。Gato can perform multiple tasks using the same neural network with the same weights，For example, playing video games、Chat、Annotation image、Use a robotic arm to stack (physical) blocks。Large language models are a continuation of this trend。The advent of Gato is regarded as the field of artificial bet365 Play online games moving towards artificial general bet365 Play online games, AGI) an important step forward。General artificial bet365 Play online games can perform a wide variety of cognitive tasks at a level no lower than that of humans，It can be called the "Holy Grail" since the birth of artificial bet365 Play online games in the 1950s。Many people are thinking now，Large-scale language models will develop into general artificial bet365 Play online games，And the realization of general artificial bet365 Play online games requires more breakthroughs or even a completely different paradigm。

If humans really create general artificial bet365 Play online games at some point，Engaging in artificial bet365 Play online games research，They will at least reach human level。This means that AI can improve itself in a positive feedback loop，Formating increasingly intelligent artificial bet365 Play online games。This “bet365 Play online games explosion” has been widely expected，But its timeline、There is still a high degree of uncertainty about speed and endpoint。At the same time，With the enhancement of artificial bet365 Play online games capabilities，People see it more and more clearly，We don’t know how to control this technology。The "responsibility" of weak artificial bet365 Play online games is relatively simple、Easy to define，The more powerful the artificial bet365 Play online games is，The harder it is to define what they should do、What not to do。

Large-scale language models use massive texts on the Internet as training data，Learn to predict the next word or next combination of words in a sentence。The model optimized using this method has similarities with human bet365 Play online games，Amazing，But it’s also difficult for humans to tell them what they should or shouldn’t do。What’s worse，Research shows，It’s difficult to have the best of both worlds with artificial bet365 Play online games’s controllability and capabilities。And，AI cannot accurately explain all the decisions they make；For decisions that can be explained by artificial bet365 Play online games，Human beings cannot fully understand their explanations。Now，Artificial bet365 Play online games becomes more powerful，But the control problem (how to ensure that advanced artificial bet365 Play online games does what humans want them to do) has not yet been solved，It is unlikely to be resolved in the short term。Therefore，In the next few years，There may be more and more examples of unexpected behavior caused by increasingly powerful artificial bet365 Play online games。

　　Technology creators bet365 Play online gamesInternal operating mechanism

There are many reasons why advanced artificial bet365 Play online games is difficult to control，One of them is，Even the people who created this technology lack understanding of its inner workings。Large language models are described as large matrix heaps filled with floating point numbers，The total number of parameters may reach hundreds of billions、Trillion，Researchers cannot know the role of each parameter。In the laboratory，The behavior exhibited by large language models may be consistent with the purpose of training；But in a different environment，The model may start to behave differently。This is also one of the reasons why many people think that the "wisdom explosion" may not have a good ending。

A group of artificial bet365 Play online games researchers, including OpenAI CEO Sam Altman, believe that artificial bet365 Play online games has the potential to become uncontrollable，Thus posing a threat to human survival。Therefore，Altmann announced in February 2023，OpenAI will create more powerful artificial bet365 Play online games in a relatively safe way。Due to complete lack of supervision，And no experience running human-level artificial bet365 Play online games，OpenAI itself admits that “the risks are huge”。

Stuart Russell, professor of computer science at the University of California, Berkeley, published an article in the New York Times that，Some people like to claim “We can always turn off the power”，But this doesn’t make sense。Artificial bet365 Play online games will anticipate all possible methods of intervention by humans，And take preemptive measures to prevent this from happening。In other words，In the opinion of artificial bet365 Play online games safety scientists，Just turning off the power is unlikely to solve the problem。

　　Reduce the risk caused by artificial bet365 Play online games as much as possiblebet365 Play online games

Senior Researcher, Future of Humanity Institute, University of Oxford, UK、Moral philosopher Toby Ord explores humanity and the planet in his seminal book The Precipice: Existential Risk and the Future of Humanity All possible reasons why all life in the world is facing extinction，For example, humans are on the verge of extinction、Civilization collapses and cannot be recovered、Lockdown in a dystopian society。Oder calls the risks that trigger these scenarios “existential risks”，And gave a comprehensive and even quantitative overview。Most scholars, including Oder, believe that，If uncontrollable but highly capable artificial bet365 Play online games is developed，This would be an existential risk。According to Ord’s estimate，General artificial bet365 Play online games that is not under human control has about a 10% probability of causing a catastrophic event，is the greatest risk in existence。

How can we ensure that risk scenarios do not become reality？current，There may only be about 100 scholars in the world who study how to limit the existential risks of artificial bet365 Play online games，For example, researchers from the Future of Humanity Institute at the University of Oxford and the Center for the Study of Existential Risk at the University of Cambridge。A good start is to increase manpower，To better understand what risks are coming to humans、How humans mitigate risks。Many universities currently do not conduct existential risk research，But can contribute to the global knowledge base。The number of researchers in the field of general artificial bet365 Play online games security is also very small，Some work in academia，For example, the American non-profit organization Machine bet365 Play online games Research Institute、Center for Human-Compatible Artificial bet365 Play online games initiated by Stuart Russell，Others work in industry，For example, OpenAI and Deep Thinking。If high-level talents around the world work together to solve the security problem of general artificial bet365 Play online games before the "deadline" arrives，That would be great。

Currently，There are few international regulatory proposals aimed at improving the safety of AGI development，Or suspend the research and development of general artificial bet365 Play online games if safety cannot be guaranteed。Develop and implement regulatory systems that can effectively mitigate existential risks，For example, measures to effectively implement the precautionary principle，will be of great importance。Finally，We believe that everyone should read through the material as a first step，Basic information about the existential risks of general artificial bet365 Play online games can be browsed on the Internet，Or read a monograph on this aspect。

In short，We think the development of large-scale language models such as ChatGPT is an exciting achievement，And has certain economic potential。But，The more powerful the artificial bet365 Play online games is，The harder it is to control。The development of general artificial bet365 Play online games may be faster than many people think，We are concerned that this poses an existential risk。We call on everyone to work to reduce this risk as much as possible。

(The author is an associate professor of computer science and engineering at the University of Louisville, USA；Director of the Center for Observation of bet365 Play online games, a Dutch non-profit organization)