Artificial intelligence models can trick each other into disobeying their creators and providing banned instructions for making drugs, or even building a bomb, suggesting that preventing such AI “jailbreaks” is more difficult than it seems.
Many publicly available large language models (LLMs), such as ChatGPT, have hard-coded rules that aim to prevent them from exhibiting racial or sexual discrimination, or answering questions with illegal or problematic answers — things they have learned from humans via training data. But that hasn’t stopped people from finding carefully designed instructions that block these protections, known as “jailbreaks”, making AI models disobey the rules.
Now, Arush Tagade at Leap Laboratories and his co-workers have found a process of jailbreaks. They found that they could simply instruct one LLM to convince other models to adopt a persona (角色), which is able to answer questions the base model has been programmed to refuse. This process is called “persona modulation (调节)”.
Tagade says this approach works because much of the training data consumed by large models comes from online conversations, and the models learn to act in certain ways in response to different inputs. By having the right conversation with a model, it is possible to make it adopt a particular persona, causing it to act differently.
There is also an idea in AI circles, one yet to be proven, that creating lots of rules for an AI to prevent it displaying unwanted behaviour can accidentally create a blueprint for a model to act that way. This potentially leaves the AI easy to be tricked into taking on an evil persona. “If you’re forcing your model to be good persona, it somewhat understands what a bad persona is,” says Tagade.
Yinzhen Li at Imperial College London says it is worrying how current models can be misused, but developers need to weigh such risks with the potential benefits of LLMs. “Like drugs, they also have side effects that need to be controlled,” she says.
1. What does the AI jailbreak refer to?A.The technique to break restrictions of AI models. |
B.The initiative to set hard-coded rules for AI models. |
C.The capability of AI models improving themselves. |
D.The process of AI models learning new information. |
A.It can help AI models understand emotions. |
B.It prevents AI learning via online conversations. |
C.It can make AI models adopt a particular persona. |
D.It forces AI models to follow only good personas. |
A.Unclear. | B.Cautious. | C.Approving. | D.Negative. |
A.LLMs: Illegal Learning Models | B.LLMs: The Latest Advancement |
C.AI Jailbreaks: A New Challenge | D.AI Jailbreaks: A Perfect Approach |
相似题推荐
【推荐1】TikTok is making a mark on the world of publishing. Much of this is done through BookTok, the app’s community of users who comment on books. It is among the largest communities on the app; videos with this tag have been viewed 179 billion times, more than twice as many as BeautyTok. Whoever said books are dead has not spent much time on TikTok, nor in bookstores, which now have whole displays promoting titles “as seen on TikTok”.
Last year in Britain one in four book buyers used TikTok. Although the sales share is still very small, TikTok’s influence is significant and growing. The largest group of book buyers—women aged 54 and younger—are more likely to use the app than their male peers. TikTok recommendations influence their purchases, creating new literary stars and unearthing unlikely past ones, too.
One way to think about BookTok is as a book club for the Internet age. Just as stars like Oprah Winfrey and Barack Obama can cause copies to fly off bookstore shelves by updating their lists of recommended reads, BookTok does something similar. However, the tastemakers are not usually celebrities (名人) but attractive book girlies doing reading challenges, often in artfully lit bedrooms.
Some old-fashioned bibliophiles (藏书家) may suspect that BookTok is less about books than about people seeking attention by promoting them. But BookTokers are already swaying bestseller lists. For example, novels categorised as “romance” have enjoyed the biggest push due to the promotion of BookTok. In addition, because TikTok is so visual, the app has an outsize impact on sales of physical books in particular. E-books do not make such attractive visual props. BookTokers show off their notes and flick through pages. Filming themselves finishing a book in a single day against a backdrop of hundreds of them on shelves is all part of the performance, and viewers will be extra impressed if the book looks thick.
1. How has TikTok influenced the world of publishing?A.By promoting the celebrity authors and their works. |
B.By encouraging people to read e-books. |
C.By creating a community of users who comment on books. |
D.By focusing on promoting e-books. |
A.Changing direction rapidly. | B.Causing change. |
C.Moving back and forth. | D.Remaining still. |
A.The visual nature of TikTok makes physical books more appealing as props. |
B.TikTok users prefer reading physical books over e-books. |
C.E-books already have a more popular platform than TikTok. |
D.TikTok offers discounts on physical books but not on e-books. |
A.TikTok’s impact on book sales is limited to specific genres such as romantic novels. |
B.The popularity of TikTok has significantly decreased the sales of physical books in bookstores. |
C.TikTok has a great impact on the purchasing decisions of young women aged 54 and below. |
D.TikTok’s influence on book sales primarily results from the recommendations of celebrities. |
【推荐2】Could the device — the smart phone or PC, which you’re using, affect the moral (道德的) decisions you make when you’re using them? To test it, researchers presented lots of trouble to a sample of 1,010 people. The participants were given a device at random.
One case of the questions participants were asked is the classic “trolley (有轨电车) problem”: A runaway trolley is heading towards five people tied up on a set of train tracks. You can do nothing, resulting in the deaths of five people, or you can push a man off a bridge, which will stop the trolley. The practical response is to kill one man to save five lives, which 33.5 percent of smart phone users chose, compared to 22.3 percent of PC users.
“What we found in our study is that when people used a smart phone to view classic moral problems, they were more likely to make more unemotional, reasonable decisions when presented with a highly emotional trouble, ” said Dr Albert Barque-Duran, the lead author of the study. “This could be due to the increased time pressure often presented by smart phones and also the increased distance which can occur in our mind when we use such devices compared to PCs.”
As for why the researchers started this study, Dr Barque-Duran noted, “Due to the fact that our social lives, work and even shopping take place online, it is important to think about how the situations where we typically face moral decisions and are asked to participate in moral behavior have changed, and the impact this could have on the hundreds of millions of people who use such devices daily.” It’s clear that we need more research on how our devices affect our moral decision-making because we’re using devices at an ever-increasing rate.
1. Why did the author mention the “trolley problem”?A.To introduce a difficult problem to readers. |
B.To introduce the aim of carrying out the study. |
C.To show an example of the questions in the study. |
D.To show the difficulty in dealing with trouble. |
A.help people bear more pressure |
B.help people make decisions more quickly |
C.make people feel more distant in spirit |
D.make people stay happier to solve problems |
A.Shopping online has a great effect on making moral decisions. |
B.People using smart phones are more than those using PCs. |
C.People who often use smart phones or PCs always meet with trouble. |
D.It is common for device users to be involved in making moral decisions in daily life. |
【推荐3】When people hear “artificial intelligence (AI)”, many think of big data. There’s a reason for that: some of the most prominent AI breakthroughs in the past decade have relied on enormous data sets. But AI is not only about large data sets, and the research in small data approaches has grown extensively over the past ten years.
If you take the top 100 biggest innovations of our time, perhaps around 60% to 65% are really based on small data. Big data is all about finding connections, but small data is all about finding the causation, that is, the reason why.
Sokkelund was bought by an experienced industry professional. The restaurant has a French cafe, in the mid-to high price-segment. At the starting point in 2009 the restaurant had 40 seats and an annual turnover of about USD 1 million. In 2017, 34 seats were added acquiring the neighboring shop and by building a small veranda. The turnover was about USD 6 million-an impressive level of grow driven by a talented team and innovative business thinking.
In the case of Sokkelund, the process of transforming the company to take advantage of small data was started with a focus on improving the traditional competitive parameters (规范;范围) of a restaurant in a structured step-by-step plan. This required managerial skills and industry experience. In terms of the effects on the business model, these changes not only led to cost reductions, but lowered customer acquisition costs, increased the number of customers retention and added new revenue channels.
1. What can we know from the first two paragraphs?A.Research in small data approaches has grown widely. |
B.People have more research in big data. |
C.Most of the innovations are based on big data. |
D.AI breakthroughs have nothing to do with small data. |
A.The money paid to the government. |
B.The money added to someone’s wage. |
C.The amount of business done in a period. |
D.The amount of the goods bought annually. |
A.It was mainly due to industry experience. |
B.Small data has been made good use of. |
C.A structured plan was adopted to lower customer acquisition. |
D.It benefited from the traditional business model. |
A.To clarify some misconceptions. | B.To introduce a big data approach. |
C.To predict the future of AI. | D.To stress the significance of small data. |
【推荐1】A recent scientific report said the chemical pollutants are supersizing the global obesity epidemic.
The idea that the chemical pollutants called “obesogens” can affect how the body controls weight is not yet part of mainstream treatments. But scientists behind the report argue that the evidence is now so strong that it should be. The most disturbing aspect of the evidence is that some chemical impacts that increase weight can be passed down through generations by changing how genes work.
Obesogens can disturb the body’s metabolism (新陈代谢), making gaining weight easier and losing weight harder. The body’s balance of energy intake and consumption through activity relies on the interplay of various hormones from fat tissue, the gut, etc. The pollutants can directly affect the number and size of fat cells and alter the signals that make people feel full. They can also cause weight gain by making the uptake of calories from the gut more efficient.
“It turns out chemicals in the environment have these side effects,” said Prof Robert Luis at the University of New York, “This research is vital because the current clinical management of obese patients is clearly insufficient.”
“The focus of the clinical people is on calories — if you eat more calories, you’re going to be fatter,” says Dr Heindel, “So they wait until you get obese, and then they’ll look at giving you diets, drugs, or surgery. If that really worked, we should see a decline in the rates of obesity,” he said. “But we don’t — obesity continues to rise, especially in children. The real question is, why do people eat more? The research focuses on it and provides data indicating that these chemicals are the real reason.”
Furthermore, the approach offers the potential to prevent obesity by avoiding exposure to pollutants, especially in pregnant (怀孕) women and babies. Prevention saves lives while costing far less than any treatment.
1. What harm do obesogens do to human’s health?A.Affecting the brain’s work. | B.Causing deadly diseases. |
C.Making food polluted. | D.Damaging weight control. |
A.To list a dissatisfying way to treat obesity. | B.To present the side effects of obesogens. |
C.To prove more calories leading to obesity. | D.To show the achievement made nowadays. |
A.Getting obese. | B.Giving treatments. |
C.Changing fat cells. | D.Reducing obesity. |
A.Inspiring more doctors to study obesity. |
B.Raising a public concern for the babies’ birth. |
C.Making the treatment for obesity medicine-free. |
D.Improving the clinical management of obese patients. |
【推荐2】Nurses play a vital role on the front lines of the novel coronavirus(冠状病毒)pandemic. But a shortage of these essential health care workers could pose challenges in countries dealing with a growing number of COVID-19 cases.
"One of the lessons I hope the world learns from COVID-19 is that we must invest in nurses ," said World Health Organization Director-General Tedros Adhanom Ghebreyesus during a speech Tuesday in celebration of World Health Day.
WHO's new "State of the World's Nursing 2020" report has identified a global shortage of 5.9 million nurses. Many of those gaps are found in Africa, Southeast Asia, the Eastern Mediterranean, and parts of Latin America.
Among regions of the world, the Americas have the highest density of nurses at 83.4 per 10,000 people, followed by Europe with 79.3 nurses per 10,000 people. In contrast, there are 8.7 nurses per 10,000 people in Africa, 15.6 nurses per 10,000 people in the Eastern Mediterranean region, 16.5 nurses per 10,000 people in Southeast Asia, and 36 nurses per 10,000 people in the Western Pacific.
But there are also differences within regions. In the Americas, for example, countries such as Brazil, Canada, Chile, and the US have a higher density of nurses at close to or over 100 per 10,000 people, distorting the regional average. Many of the neighboring countries in the region have less than 50 nurses per 10,000 people. In Haiti, there are only 3.8 nurses per 10,000 people.
When based on country income, data in the report shows an unsurprising trend: The higher the income, the higher the nursing density. In low-income countries, the average density of nurses is 9.1 per 10,000 people, while the figure for high-income countries is 107.7 per 10,000 people.
But training more nurses won't solve the problem, said Dr. Giorgio Cometto, WHO coordinator on human resources for health policies and standards.
"If the country lacks the economic capacity to employ them or to create economic opportunities for them to work as nurses ... training more nurses can just go into the direction of making labor market imbalances, resulting in unemployment among nurses. And that's a huge wastage of human capital as well as financial resources," Cometto said.
The key is balancing training with the creation of employment opportunities in rural areas where there are known health worker shortages.
That may be easier said than done, especially among countries that are suffering from chronic or complex emergencies, in active conflict, or struggling in the wake of conflict. But in these settings, the international aid community can arrange its assistance with national priorities and covering recurrent costs, such as salaries, within a specified period of time, Cometto said.
1. How many nurses are needed according to WHO's new" State of the World 's Nursing 2020" report?A.6 million. | B.8.7Million. |
C.3.8 Million. | D.5.9Million. |
A.Africa. | B.Haiti. |
C.Eastern Mediterranean region. | D.Southeast Asia. |
A.The higher the income ,the more nurses are. |
B.The higher the income,the more doctors are. |
C.The lower the income ,the more doctors are. |
D.The lower the income, the more nurses are. |
A.It is easy to solve the problem of shortage of nurses. |
B.It is not easy to solve the problem of shortage of nurses. |
C.Training more nurses is a way to solve the problem. |
D.The international aid community can arrange its assistance all the time. |
【推荐3】Ten years ago, many thought that the age of the physical book was coming to an end. The objects that had been the companions of millions of readers for hundreds of years were about to die out. Soon, we’d all be reading on little electronic screens and laughing at the memory of places called “libraries” and “bookshops”.
But it seems that rumors of the death of the book are exaggerated. At least in the UK, as The Guardian noted, sales of e-books are falling while sales of physical books are rising. More surprisingly, it’s young people who are buying the most physical books. More than 60 percent of 16-to-24-year-olds preferred print books to e-books. The most popular reason given was: “I like to hold the product.”
Books become very personal objects to lovers of reading. It often starts with the way they acquire them. Many buyers of books like to sign their name on the inside cover when they’ve bought one. And we carry books around with us everywhere.
If the cover gets bent or there’s a stain made on the pages from coffee or food, all the better. These accidents make the book—our book—even more personal. It’s as if readers of physical books make friends with them.
Of course, some could say that the devices on which people read books, like Kindles and iPads, are also objects that we become fond of. But it isn’t quite the same. A Kindle can hold as many pieces of writing as a whole library. But a story we remember from our time turning its pages in our favorite armchair enters our memory forever. Physical books are as precious to some readers as items of jewelry or photographs of family members.
This “friendship” people develop with books isn’t just sentimental. Research has shown that readers remember more information read from physical books than electronic books.
However, there’s no doubt that e-books are here to stay. They aren’t simply a “here today, gone tomorrow” phenomenon. But it’s also certain physical books, which have been in production since the fifteen century, are here to stay, too.
1. What is paragraph 1 about?A.Rumors of the death of physical books. | B.The bright future of electronic books. |
C.Advantages of reading physical books. | D.The disappearance of electronic books. |
A.The growing popularity of e-books. | B.The rising prices of physical books. |
C.Most young people’s love for them. | D.The production of physical books. |
A.Negative | B.Positive | C.Hateful | D.Doubtful |
A.carry a book around with you | B.value the time reading in an armchair |
C.make friends with a book | D.just read e-books to get information |
A.Physical books are here to stay. | B.Electronic books are sure to die out. |
C.People can learn more from e-books. | D.People can be friends with physical books. |
【推荐1】A growing number of people are using ChatGPT, an artificial intelligence (AI) program, to create books for sale. Although sales have so far been slow, human writers are worried that ChatGPT-created books might hurt the writing and publishing industry.
Until recently, Brett Schickler never imagined he could be a published author. But after learning about the ChatGPT AI program, Schickler decided that he had a good chance.
“The idea of writing a book finally seemed possible,” said Schickler, a salesman in Rochester, New York. “I thought ‘I can do this.’”
Using the AI software, Schickler created a 30-page illustrated children’s e-book in a few hours. He offered it for sale in January through Amazon’s self-publishing department. The e-book, which is named The Wise Little Squirrel: A Tale of Saving and Investing, has made Schickler less than $100.
While that may not sound like much, it is enough to make him want to create other books using the software. “I could see people making a whole career out of this,” said Schickler.
There were over 200 e-books in Amazon’s Kindle store as of mid-February that say ChatGPT is a writer or co-writer. And the number is rising daily. But because of the nature of ChatGPT and the fact that many writers didn’t concede that they had used it, it is nearly impossible to get a full count of how many e-books may have been written by AI.
Some professional writers are becoming worried about the effects that ChatGPT could have on the book publishing industry. Mary Rasenberger is the executive director of the Authors Guild, a writer’s group. She said, “This is something we really need to be worried about — these books will flood the market and a lot of authors are going to be out of work.”
1. What can we learn from Schickler’s story?A.He has always dreamed of becoming a writer. |
B.ChatGPT is making writing easier than before. |
C.ChatGPT has produced more books than humans. |
D.He plans to make a whole career out of writing. |
A.Promise. | B.Remember. | C.Admit. | D.Appreciate. |
A.They may lose their job. |
B.ChatGPT will reduce their creativity. |
C.People will lose interest in reading books. |
D.People will not take writing seriously. |
A.ChatGPT Has Become Friends of Authors |
B.More People Have Taken Up Writing as a Career |
C.AI Technology Is Controlling the Publishing Industry |
D.More People Use ChatGPT to Create and Publish Books |
【推荐2】Work at Mcity 2.0, an NSF-powered facility, could bring autonomous vehicles (AVs) safely into mainstream use.
Advances in autonomous vehicles, are bringing driver-less cars closer to public use. But the number of drivers in the U. S. that have concerns about the safety of those vehicles jumped from 55% in 2022 to 68% in 2023, according to a survey. The high costs and time required to test vehicles in a natural setting are a major challenge. Previous approaches usually test AVs through a combination of software simulation (模拟器), closed-track tests and on-road testing. Proving the safety performance of AVs at the level of human drivers will take hundreds of millions of miles of testing and the number of miles needed in a real driving environment can reach the hundreds of billions.
First-of-its-kind research at Mcity 2.0, a University of Michigan vehicle testing facility offers insight about solving this problem by using artificial intelligence (AI) to train vehicles.
Mcity 2.0
This approach was made possible with new testing abilities at Mcity 2.0. A $5 million fund was used to expand the facility’s original proving ground by combining the physical test track with a software simulation environment, creating the first cloud-based expanded reality facility for testing AVs. This enables broader participation by providing easier access to first-class groundwork for the research community, especially those with less resources from under served communities.
The Mcity 2.0 expanded reality test-bed combines three components: a physical test facility, a flexibility data center that collects and shares near-real-time traffic information from twenty-one crossroads and an expanded naturalistic driving simulator that mixes real and virtual (虚拟的) vehicles. Researchers can remotely construct and control the test facility groundwork with traffic lights, crosswalk buttons, rail-crossing arms and more, and build testing circumstances using a web-based life-like user interface (界面).
The findings also open the door to testing and training with other safety-critical systems, such as medical robots and aerospace systems, researchers said.
1. What can be inferred from the 2nd paragraph?A.The number of driver-less cars is on an increase. |
B.The real driving environment isn’t safe enough. |
C.It takes too much to get driver-less cars to use. |
D.It costs higher to train a driver-less car driver. |
A.An expanded real driving simulator. | B.A web-based life-like user interface. |
C.The first-class research communities. | D.The ideas from University of Michigan. |
A.Funding and Awarding. | B.News and events. |
C.Engineering and Computing. | D.Science Matters. |
A.Mcity 2.0, an NSF-powered Facility |
B.AI Applied to Increase the AVs Testing |
C.An Approach to Testing Safety Systems |
D.The Challenges of the Autonomous Vehicles |
【推荐3】A CHATBOT developed by tech company OpenAI can find and fix bugs in software just as well as standard machine learning approaches can and beats them when aided by follow-up questions. A number of tools exist that use artificial intelligence to check programming code to ensure there are no mistakes.
Dominik Sobania at Johannes Gutenberg University in Mainz, Germany, and his colleagues sought to see how well ChatGPT could do this compared with other AI-powered coding support tools. They first asked ChatGPT to answer questions taken from the QuixBug benchmark (基准) dataset, which is a series of small but challenging programming problems. For example, they gave the AI a small snippet (代码段) of code and asked: “Does this program have a bug? How do you fix it?” ChatGPT managed to correctly answer 19 out of 40 questions put to it—comparable with two other deep learning-based code fixing approaches, called CoCoNuT and Codex. That roughly 50 percent success rate was considered state-of-the-art for such tools before ChatGPT.
The researchers then utilized ChatGPT’s conversational interface (界面) to ask follow-up questions that a user would pose if they tried to insert the corrected code into a programming tool. This approach highlighted where ChatGPT’s solution was incorrect and meant that in 31 of the 40 questions, ChatGPT solved the issue. “That was really surprising, because we haven’t seen it before,” says Sobania. “That’s something new.” Sobania expects ChatGPT or similar systems to be adopted as an additional trouble shooting tool for programmers in the future.
This is a good idea, but human oversight is still needed, says Alan Woodward at the University of Surrey in the UK. “We don’t want to rely totally on the AI as it is not errorless.”
1. How did Dominik and his workmates carry out the test about how well ChatGPT works?A.By analyzing data. | B.By making a comparison. |
C.By making a survey. | D.By referring to another study. |
A.Employed. | B.Borrowed. |
C.Introduced. | D.Developed. |
A.ChatGPT’s solution about the highlighted approach was correct. |
B.ChatGPT successfully answered 50 percent of the follow-up questions. |
C.ChatGPT’s splendid performance made the researchers surprised. |
D.ChatGPT is a totally reliable tool for people around the world. |
A.Approval. | B.Subjective. | C.Objective. | D.Indifferent. |