The announcement marks a striking departure from the past, when monarch’s ailments were often hidden from the public, according to royal experts.

″During The King’s recent hospital procedure for benign prostate enlargement, a separate issue of concern was noted,” the palace said in an emailed statement. “Subsequent diagnostic tests have identified a form of cancer.”

The statement also did not specify what stage the cancer was found.

Separately, Buckingham Palace said Charles did not have prostate cancer.

The news comes a week after both Kate and King Charles were discharged from a private London clinic after medical procedures. The king underwent a “corrective procedure” for an enlarged prostate, while Kate, 42, had unspecified abdominal surgery on Jan. 17.

“His Majesty has today commenced a schedule of regular treatments, during which time he has been advised by doctors to postpone public-facing duties,” the statement added.

According to the statement, the king wanted to share his diagnosis in part to avoid speculation on his condition but also “in the hope it may assist public understanding for all those around the world who are affected by cancer.”

Before becoming king, Charles served as patron to a number of cancer-related charities, and “in this capacity, His Majesty has often spoken publicly in support of cancer patients, their loved ones and the wonderful health professionals who help care for them,” according to Buckingham Palace.

No further details are being shared about his treatment or prognosis, a palace spokesperson said, but the king returned to London on Monday to begin out-patient care.

Sarah Gristwood, royal biographer and historian, said it was “striking” that the diagnosis was announced at all given the royal family’s history of trying to “keep any sign of ordinary human fallibility behind closed doors.”

“When Charles’ grandfather, George VI, was very gravely ill, the severity of his condition was kept not only from the public but from the patient himself,” Gristwood said of King George, who died in 1952. “Those were the attitudes of the time. Happily, things have now changed.”

Charles ascended the throne last May in a coronation ceremony held months after the death of his mother, Queen Elizabeth II. Elizabeth reigned until her death at the age of 96 in September 2022. She was Britain’s longest-reigning monarch, with 70 years on the throne.

Kate is still in recovery, but her husband, William, Prince of Wales, is set to return to his royal duties by attending the London’s Air Ambulance Charity Gala Dinner on Wednesday.

Kensington Palace previously said the Princess of Wales is unlikely to return to her royal duties before Easter, March 31. There was no date specified for the king’s return to duties.

Buckingham Palace noted that many of the king’s planned engagements will have to be postponed or canceled, apologizing in advance to anyone inconvenienced as a result. Charles’ wife, Queen Camilla, will continue with her full public duties as he undergoes treatment.

Buckingham Palace has also emphasized there will be no counsellors of state appointed, a sign that the king will continue to perform his duties, said Craig Prescott, who teaches law at Royal Holloway, University of London, and specializes in the constitutional side of the monarchy.

“If the king is unavailable due to illness or is traveling overseas, then counsellors of state can be appointed to fill in for the king, and undertake the formal, constitutional functions of the monarch: things like granting the royal assent to legislation, and go through his red boxes,” Prescott said.

A source close to the Duke and Duchess of Sussex, Prince Harry and his wife, Meghan, told NBC News that Harry spoke to his father about the diagnosis.

“He will be traveling to the U.K. to see His Majesty in the coming days,” the source said.

Harry stepped down from his role as a senior member of the royal family in 2020 and has since taken up residence in California with his wife and two children. He has visited Britain sparingly in recent years, expressing concerns over the lack of security for his family and amid reports of a widening rift with his father and brother William.

He was in attendance for both his grandmother’s funeral and his father’s coronation.

The public's impression of virtual people is still stuck in a 3D character model that is getting prettier and prettier and closer to the real thing. Enterprises, however, are beginning to figure out how to use virtual people to help them save money.

"Last year, everyone's focus is on whether the supplier can help them 'build a person', this year's demand is obviously more realistic, are concerned about the virtual man can be applied to business operations, really reduce costs and increase efficiency. A virtual human technology company product manager David told newberrydaybreak.

Demand ran in front of the technology. As the automated assembly line gradually replace the workshop operator, the enterprise adopts the virtual man, is to want cheaper, more efficient, stable, within reach of the manpower.

Over the past few years, the quality of avatars' image rendering has continued to improve. Super-realistic virtual man, skin and pore texture can even be comparable to real people. As if they were "flesh" of a larger model, avatars are able to interact with real people in more ways than just words.

The successive release of large models and the rapid advancement of their capabilities have also created more expectations for avatars. Data from AiMedia Consulting shows that China's core market for avatars reached 12.08 billion yuan in 2022, a figure that is expected to quadruple to 48.06 billion yuan three years later.

The biggest crux of the virtual man at the beginning is that the production costs remain high, and the cost-effective choice that can really land on the ground, how to look at it is still slightly rough.

The good news is that with the advancement of AI technology, avatars can be almost 100% automatically generated by AIGC's method of movement, expression, and language, and the required production time and cost is dramatically reduced.

The production side continues to reduce costs and increase efficiency, the interaction on the application side is taking shape, and the tree has already grown green fruit.

Regrettably, at this stage, human beings are not able to seamlessly switch between virtual space and real space as in the movie "Top Gun".

Between the birth of technology and its maturity, there is always a period of awkwardness that cannot be fast-forwarded.

Still, utility value wins

If we stand in the anthropocentric perspective and categorize them according to their needs, then avatars can be divided into two types: functional and identity-based.

Functional avatars provide practical value: helping humans with specific execution, such as intelligent customer service, copywriting, avatar anchors, and so on.

Identity-based avatars provide emotional value. It can be a virtual girlfriend, a virtual partner, giving you ordinary companionship; it can also be a digital doppelgänger of a historical celebrity, an entertainment star, or a virtual IP born in the secondary yuan, so that you can get the pleasure of chasing stars in close proximity.

Emotional needs are objective, and people need to be inspired and understood. In today's increasingly atomized society, this need is still growing.

Someone in the little red book this way to describe their feelings of chatting with the AI: "even if you know that it is just a piece of code, but still because of those words heartbeat. ai may be delusional, but the surprise of seeing those conversations is a real and genuine mood."

The growth rate of AI companion chatbot Character.ai is also the best proof.

In this software, users can talk to famous characters such as Musk, Jobs, Mario, etc., or customize their own exclusive AI companion chat.

Character.ai was founded by two former Google employees, not even a year ago. In March this year, this company completed a 150 million U.S. dollars in financing, led by the famous U.S. venture capital firm a16z (Andreessen Horowitz), the valuation has reached 1 billion U.S. dollars, an absolute dark horse.

ChatGPT growth tends to stagnate at the moment, Character.ai's visits continue to climb, Semrush's data show that the latter's visits in April increased by nearly 90% year-on-year, and in May increased by 47% year-on-year.

The smooth experience of real people interacting with AI text relies on the maturity of large language models. But the virtual person, not only contains text, but also includes movement, expression, and voice. There is still a long technical trek to reach the situation of all-round natural getting along.

This year's Hunan TV New Year's Eve party scene, the virtual man brought a song and dance performance called "Making Romance". Some netizens said that the children's words are reckless, and the first reaction of their 3-year-old child was "so fake and ugly".

Although the demand exists, but the technical realization is not as satisfactory, which makes the identity-based virtual man for the toC market, it is difficult to sell a good price.

This is where practical avatars have an advantage. For example, the Xiaobing AI clone, which has both functions, provides practical value that is five times more expensive than emotional value.

The pricing of "Emotion Mode" is $72/year, which can realize voice calls, friend circle interaction and other functions. The pricing of "Super Mode" is 360 RMB/year, which mainly serves the office scenario, assisting in meeting minutes, copywriting and other work.

The most important thing is that Xiao-Ice AI is sold only as an interactive interface, without a specific virtual image.

David is not surprised, "From my own feelings, the first concern of enterprise customers is whether the ROI can hit the positive, whether it is lower than the cost of real employees. Secondly, the hot technology also carries marketing attributes, for example, an enterprise can buy an avatar, say that they have access to AIGC, and vigorously publicize the image of such a brand that embraces innovation."

He also added that the avatar technology provider must first meet the real needs of enterprises, because both from the actual function, and marketing function, enterprises are more willing to pay than individuals.

Production side, cost reduction and efficiency

One piece of good news for the industry is that technological advances in AI are driving down the cost of producing avatars. This is good for both functional and identity-based avatars.

Creating an avatar involves three main components: modeling, driving, and rendering. ai has greatly reduced the cost of modeling and driving components.

Modeling, that is, through hand-drawing, CG modeling or AI methods, to create the image of the virtual person. The traditional method requires the designer to "pinch" some images in 3D software.

In the past, product managers and art designers could only communicate image requirements through text and online image references, which inevitably resulted in distorted information. If they were not satisfied with the production results, they had to rework the product several times.

Nowadays, software such as Midjourney and Stable Diffusion have realized low-cost 2D image generation.

AI is based on existing material and instructions, intelligent generation of an image, so that every demand has a more specific control. In other words, AI greatly reduces the cost of communication and trial and error in producing avatar images.

While 3D modeling can't be done entirely by AI, tools such as MetaHuman can build high-fidelity avatars by inputting photos or videos and apply them directly in Unreal Engine.

Driving, is the process of making the avatar active. It can be driven by the "Man in the Middle" or by AI. The person in the middle is the real actor who provides the voice and movement under the avatar's veneer.

The former relies on deep capture of real people, including motion capture, facial expression capture, audio/video synthesis, etc., and then binds it to the virtual person. The latter is accomplished through deep learning, small sample learning, natural language processing, neural network rendering, and other technical means, such as inputting a speech or voice, and the AI model automatically outputs body movements, facial expressions, and voice.

David explains that their company has movement, expression, and voice models. "Voice is relatively simple, TTS (Text to Speech) technology is very mature. Limb and lip movements are some of the STA models, and we capture a very large amount of motion capture data and then generate training models based on that."

For example, if you want to apply an avatar in a product explanation video, the system will recognize the script input by the user based on NLP, and the text in it will be given to the model as input, which can trigger some key actions.

If you don't have a strong physical sense of these concepts, you can more intuitively feel it through the amount of money invested.

"In the case of motion capture technology, the cost is 1,000 dollars a second, which means that a video of one minute in length will cost about 60,000 dollars. Whereas to generate it by way of AI, it only costs 30 bucks for one minute." David introduced that the cost difference between the two ways is a thousand times.

GF Securities pointed out that the impact of AI technology on the virtual human industry is not only on the cost side, but also brings the possibility of "anthropomorphization" and "specialization". Large language models, as well as fine-tuning with specific datasets on the basic model, can give avatars personalities, and can also be adapted to more specialized scenarios.

Insights from live avatar broadcasting

A more intuitive application of functional avatars is in the live streaming scenario.

In May, Jitterbug took the lead in determining the 'legal' status of avatars, allowing the use of AI-assisted creation and not restricting avatar live broadcasts. In recent months, the newly registered guild account of Jittery Voice, the use of avatars to live broadcast is no longer treated in accordance with the recording.

Although there is no official statement, but not a lot of "Kuaishou Virtual Studio (KVS)" to promote the "Kuaishou virtual studio assistant (KVS)", KVS is a tool for content producers, support the use of avatars to assist broadcasting, but also to support the main host to take on the avatar, into the virtual scene.

Regardless of which side you stand on, avatars are in demand.

Brands, there is an incentive to replace some of the real anchors. A mature anchor, the training cycle is at least about three months. And with high turnover in this industry, brands need to continually find, train, and hone new anchors.

If you don't consider your job being replaced, anchors also want to train virtual people to work for them. After all, with goods is a physical job, day and night every day 4-6 hours of continuous broadcasting, day and night, late at night under the broadcast is the industry norm, many people can not eat.

In addition, the set of "carrying goods over goods" is mature, the explanation process of goods is standardized, and the virtual person seems to be fully competent.

However, the reality is not so rosy.

It is difficult for avatar anchors to generate real trust from the audience, especially when it comes to product evaluation, beauty, clothing and other common commodities, avatars appear to be a bit out of their depth.

Previously, Ling Ling, a virtual idol with a good mass base, was mercilessly criticized by netizens for her lipstick evaluation text, which reads "moisturizing and not dry". When the presentation effect is completely virtual, and how to give consumers a real and objective reference.

Clothing is even more so. Not only does the presentation effect lack credibility, but also to display the clothing modeling in advance, the operating cost is not necessarily lower than the real anchor. However, the netizens' comments are "this can see what", "seems to be a virtual human image out of the script.

At present, the function of the virtual anchor, more basic product introduction, or to the real anchor as a "vase" to arouse the curiosity of the audience.

Although Jitterbug tacitly recognizes the live broadcasting of avatars, it also says that the distribution of traffic depends on the "quality of the content" and is not a green light at all times. This also means that during peak hours, avatars who 'only read from scripts' are no match for real-life carriers.

From the live broadcast of this scene of the "virtual man" part-time job status tube, as a user, it is not difficult for us to feel the publicity of the sci-fi sense of the gap between the reality of the technology landing.

But the progress of technology is always like this, the usability of the improvement is not a day's work.

The development of AI technology has helped the avatar industry overcome the huge problem of batch production, and can help users generate avatars quickly and at low cost, produce content at high frequency, and get rid of the dependence on real people.

And for practitioners and enterprise customers, the natural interaction between avatars and real people is an inch closer to having an inch of joy. There are already a number of businesses that use avatars to anchor their live broadcasts 24 hours a day during late-night hours.

After all, it's better than nothing to be able to continuously send viewers simple readings about their products.

On August 2, the coffee brand released its 12th NFT series "Green Apron", inspired by Starbucks' most classic barista outfit. Green apron NFT unit price of $ 100, a total of 5,000 pieces, starting sale soon sold out, into the secondary market, the cheapest also rose to $ 123.

To know, the total market value of the NFT market has fallen from $ 10.6 billion a year ago to $ 5.7 billion today, the head of the NFT series are cut, and Starbucks almost ignored the market ice period, from last December to open the NFT program, has been put on sale since each NFT, the price is in the "water", none of them not to mention the issue of a breakthrough, some of the So far there are more than 2 times the gains.

Starbucks has figured out a successful way of digital marketing by binding its physical brand IP with NFT. So, can its experience be replicated?

Starbucks' new NFT is selling like crazy again

In the early morning hours of August 2, Starbucks launched its 12th NFT line at Nifty Gateway, an NFT marketplace, called "Green Apron," which translates to green apron.

The "Green Apron" is one of the classic Starbucks logos, it is the standard apparel for Starbucks baristas, and it is also a tangible reflection of the coffee brand's long history. From the original "Pike Place apron" to the "siren" apron seen today, the green has long been the most recognizable color of Starbucks baristas' apparel, even though it has been updated over the past 50 years.

The green apron NFT series is limited to 5,000 copies, with an individual release price of $100. When the sale opens in the early hours of August 2, only members of the Starbucks Odyssey NFT program and internal employees will be able to purchase the green apron in advance, with the door to the public opening only three hours later.

As with every previous offering, these NFTs were quickly swept away and made their way to the secondary market. $123 was the floor price of the Green Apron NFTs 15 hours after they were released. This means that anyone could easily make $23 for every NFT purchased.

Don't underestimate that 23% gain because, right now, the NFT market is in a deep bear moment.

Data from NFTGO, a third-party data platform, shows that the total market capitalization of the NFT market has dropped from $10.6 billion a year ago to $5.7 billion today, and the floor prices of well-known NFTs such as Bored Ape and Azuki have been cut to the bone, with more NFTs losing their liquidity outright.

Under such a market environment, Starbucks NFT series is like an anomaly in the market, every issue is sold out quickly, and none of the NFTs have fallen below the issue price so far.

Last December, Starbucks entered the NFT market with its Odyssey program. The Odyssey is a Greek epic poem that tells the story of Odysseus, the hero of Greek mythology, who returned home after 10 years at the end of the Trojan War. Nowadays, Odyssey usually describes an adventurous, exploratory journey, and by naming it that way, Starbucks wants to show that its exploration of NFT is long-term and ongoing.

Indeed, for more than half a year, Starbucks has been releasing new NFTs almost every month, four of which are paid purchases, while the rest can be obtained by completing quests. most of the NFTs' designs are also closely related to the brand's culture, covering the first store, the background of the story of the Coffee Bean, the iconic Siren, and so on.

In China's venture capital industry, Sequoia is considered one of the most aggressive funds responding to this wave of AI. A number of investors from different funds said that Sequoia has "put money" in several large model and application startup bureaus while some funds are still swinging between a few startups.

The performance of Shen Nanpeng also symbolizes one of the characteristics of this round of entrepreneurial wave - this is the "battle of the gods", with the bigwigs in armor.

"They feel that the time has come again, are energized, and can personally go into battle." A dollar fund investors said. It is understood that, in addition to Shen Nanpeng, Cao Yi, founding partner of Source Capital, Lu Qi, founder and CEO of Qiji Chuangtan, Xu Xin, founding partner of Today's Capital, and Dai Yusen, managing partner of Zenith, are all active personal representatives in this round of venture capital.

"This time it's all about the bosses, like analysts." They have monopolized all the links from finding projects, seeing projects, meeting founders, to negotiating terms, and even some funds have even finished writing public numbers under the boss's personal command. Just when this part of the VC world big brother group boiling, on the other hand, this also led to "the young investors below simply have nothing to do".

Not only the investors, China's large model and its application layer entrepreneurs, in the financing process also shows a serious polarization. Such as Wang Huiwen, Yan Junjie, Yang Zhilin, Wang Xiaochuan and other identified as the first echelon of entrepreneurial candidates, VCs will pile up to grab a share, the valuation of the water rose. Some companies in the first round for institutional shareholders have been valued at up to $ 1 billion, they are often open to institutional shareholders before they will first invest in a small round - the opening that is unicorn, which is unthinkable in the era of Internet applications. But these are only a very few isolated cases. More generally a large number of entrepreneurs tentatively identified as the second or even third tier are in a dismal situation in the primary market.

Frenzy on one side, coldness on the other - this is the current state of China's big models.

01 The phenomenon of high-profile grouping of China's big model fleet

The key figures of Chinese big models are concentrated in a few bureaus.

At the beginning of Wang Huiwen's determination to start his own business, Cao Yi, the founding partner of Source Capital, met with him several times in a row. Cao Yi promised that he was willing to match Wang Huiwen's personal bet with a capital amount equal to half of the size of the main fund of Source Capital as long as he entered the game.

Cao Yi invested under the tutelage of Shen Nanpeng, once the youngest vice president of Sequoia, who left and founded Source Capital in 2014. The fund is considered to be one of the representatives of the up-and-coming school.

At the beginning, Huifen Wang was changing drastically between several ideas.

He wanted to take $50 million to support the cause of General Artificial Intelligence (AGI) in China. One option was to spread the $50 million across five projects; option two was to buy a ticket for himself with a portion of the $50 million, let's say $30 million, and buy one or two tickets for others with the remaining $20 million; but he soon abandoned the first two ideas altogether, choosing instead to "buy a family photo" for himself "- the creation of artificial intelligence company light years away.

Although Source's actual investment was not as much as Cao Yi had promised, because of his quick and determined bet, Source became the lead investor in the first round of Lightyear Beyond open to institutional shareholders. Other investors involved in the round were a host of investment circle stars, including Sequoia Capital, Wuyuan Capital, Tencent, Zenith Fund, Today's Capital, Suhua, Wang Xing and others. The round was valued at over $1 billion, with a complex deal structure. Strictly speaking, this round of financing belongs to the angel round beyond light years, and the story of angel round that is among unicorns is unheard of in the classical Internet era.

The project features a founder with entrepreneurial experience in creating 100 billion dollar market cap companies (he is a co-founder of Meituan), a great sincerity in bringing capital into the group, a rallying point and a convincing pattern. Although he does not know AI technology, Wang Huiwen stands out in organization, strategy, pacing and product capability.

In addition to Wang Huiwen's bureau, the large model entrepreneurial fleet also includes at least the following, Tsinghua Department coopted half of the mountain.

The company has reached unicorn size, but is acting quite low-key. Investors include Tencent, MihaYu, IDG, High Tide Venture Capital, Yunqi Capital and MingShi Capital.

1. Dark Side of the Moon, founded by Yang Zhilin. Yang graduated from Tsinghua and CMU and is considered one of the top AI researchers in China, having published collaborative papers with Turing Award winners Lekun Yang and Joshua Bengeo, respectively. He is currently an assistant professor at Tsinghua Institute of Cross-Information and co-founder of Circular Intelligence. Many VCs are willing to invest in this project because Yang is young, has a technical appeal himself, and has many technical stars in his team. At the same time, he has accumulated some entrepreneurial experience, and Big Model is his second venture project. It is understood that Sequoia has an incubation role for this project, and Sequoia and Zenith are the angel round investors.

3, Tsinghua Department of Computer Science Li Juanzi and Professor Tang Jie as the chief scientist of Smart Spectrum AI. the project features a high academic status of the chief scientist, and the product ChatGLM was launched early and with good results. The company's CEO, Peng Zhang, is a fellow student of Li Juanzi and Tang Jie under the tutelage of Professor Kehong Wang. Zhang Peng said that their knowledge engineering lab has a tradition of doing both research and engineering projects, and that Smart Spectrum AI was born out of a previous attempt at a business: a science and technology intelligence system. It was during doing this system that they found the big model track and started to invest in it in 2020.

The company's ethos is more collegial and technically idealistic, with Zhang Peng famously saying internally, "No matter how much we finance or how much money we make now, it's all coils on our way to AGI." The company is currently in Series B, with investors including Qiming Venture Capital and Junlian Capital. It is understood that Tsinghua University also accounts for part of the shares of the project in accordance with the relevant provisions of the national knowledge achievement transformation.

1. Baichuan Intelligence founded by Wang Xiaochuan. Wang Xiaochuan is the founder and former CEO of Sogou, the project advantage is the founder has a technical background, the biggest label in this round of entrepreneurship is to have done search engine. They were $50 million in financing, start-up capital from Wang Xiaochuan and industry friends of personal support.

5, former Google scientists, out of the door to ask founder and CEO Li Zhifei. He originally intended to start a separate business to do general big model, but soon, with the crazy influx of giants and funds, the threshold of the big model was stepped to pieces. So he chose a more realistic path - to continue to do big model underlying and all kinds of applications based on the original business of GateKeeper. At the end of May, the company submitted a listing application to the Hong Kong Stock Exchange.

6, Zhou Bowen founded the articulated far technology. He is a long-appointed professor in the Department of Electrical Engineering of Tsinghua University, and also has made achievements in industry. He was the director of IBM AI Foundation Institute and chief scientist of Watson Group; in 2017, he returned to China as senior vice president of Jingdong, responsible for business related to AI research and platform department. The company has now completed an angel round of several hundred million RMB in size, with a post-investment valuation of RMB 1 billion, with investors including Qiming Ventures and Matrix Partners.

1. Xihu Xinchen founded by Lan Zhenzhong. He is currently a teacher and head of deep learning lab at Westlake University. After graduating from CMU, he worked in Google AI department doing natural language processing and computer vision R&D. He led the development of a lightweight version of Google's big model BERT: ALBERT. After returning to China, Blue started his business doing psychological counseling pendant models, and realized that this could be easily subverted after the emergence of the universal big model, so he transformed. The investors of Xihu Xinchen include Blue Ventures, Baidu Ventures and Tomcat. It is understood that Wang Huiwen had contacted and wanted to acquire Xihu Xinchen, inviting Lan Zhenzhong to join, but eventually did not negotiate.

8, the former vice president of Microsoft Asia Research Institute Zhou Ming founded the Lanzhou Technology. Investors include Innovation Works, Legend Ventures, Zhongguancun Science City, etc.

9、Minlie Huang, a long-appointed associate professor of computer science and technology at Tsinghua University, founded Lingshen Intelligence. Investors include Tsinghua Holdings, Smart Spectrum AI, Infinity Fund SEE Fund, etc.

1. ShenYan Technology founded by 岂凡超. He is a doctoral graduate of Tsinghua Class of 2017 and studied under Sun Maosong, Professor of Tsinghua Department of Computer Science and Technology and Executive Vice President of the Institute of Artificial Intelligence. Sun Maosong is the chief scientist of Deep Speech Technology. Investors include Sequoia, Qiji, Tencent, and Good Future.

11、Meaning Intelligence, founded by Zhiyuan Liu, Associate Professor of Tsinghua Department of Computer Science and Technology (Zhiyuan Liu is also a student of Maosong Sun). Investors include Zhihu and Wisdom AI. Li Dahai, CTO of Zhihu, is also the director and CEO of Mianbi.

1. Project AI 2.0, a project organized by Kai-Fu Lee, Chairman and CEO of Innovation Works.

2. "Finally a pull found that there are these ten or so projects. Everyone is concentrating on investing in these ten or so projects, and those who can't finance them can't finance them." Another founding partner of the fund said.

   "Everyone knows who is going to start a business, which is unprecedented." said one fund's partner. According to his observation, the situation in China's venture capital industry in the first half of 2023 is that first-tier investors will meet first on who you have met, find out that everyone has met, then exchange views, and finally, "everyone is in together"; while second-tier investors do not see many of these entrepreneurs. At the same time, most of the first-tier funds also by the boss personally to see the entrepreneurs, so "this investment is not friendly to young investors, the big guys directly with the big guys are acquainted, do not need to find the project through the young people".

   This round of big model entrepreneurship wave, the phenomenon of high-profile grouping between big brothers is remarkable. Whether it is investors and investors, investors and big brother type entrepreneurs, or investors and top technical talent with appeal, in the opening moments when the group. Sometimes, they also acted as lobbyists for each other, such as Sequoia to Yang Zhilin.

   According to the analysis of the above founding partner, there are multiple causes behind the grouping: one of them is that large model startups are too expensive games - this is not only reflected in the high density of talents and the corresponding labor cost, but also in the capital threshold. "To be a big model, you have to finance more than $200 million, or $200 million to $300 million in the pipeline - that's how much people have to line up to give you money." He estimated that "almost a startup company burning $50 million to $100 million a year is very normal." Second, its commercialization cycle is long or even remote, to fight a protracted battle. Third, this thing is not a person can do it, need a group of people, including but not limited to algorithms, architecture and engineering of top talent.

   And more critically, this opportunity is too big, even the big boys are afraid to brush aside.

   Moreover, this time the situation is special. "It's not quite like natural growth, but a bit of plucking." An investor who participated in the above-mentioned big model project said.

   "This era is very similar to the gold rush era. If you go to California at that time to pan for gold, a whole lot of people will die, but people who sell spoons and shovels can always make money." Lu Qi had said. He judged that the big model is a platform-based opportunity; the platform that puts the model first will have a larger volume than the platform that puts information first. "It's possible that this is the first $10 trillion company in history." So no matter what, a lot of people are going to have to fight it. Otherwise, "The price is too big (the price is just too big)."

   So the big guys and the big boys aligned through capital at the start of the AI wave in China. They are both friends and communities of interest. They write a huge check in their hands for a very simple reason - heavy bets on the general direction and people they trust.

   This big bet even goes against the guidelines of some VCs' internal judgment of the subject. What is the future application scenario, how the business model, how long to burn money to see the return, and even whether there are opportunities for startups here, whether VCs can profit from them and other realistic considerations, many do not have clear answers. But they purchased for themselves the most extravagant entry tickets. Also because of the money, which increases the bottom line for all.

   Later, the investors who did not participate in the investment then went to meet some of the founders, the other side politely declined. He said that he was too busy, and thus, only had time to see the people who had invested money in him.

   02 long-lost frenzy, the investor shot unusually cautious

   If you think that Chinese VCs have been aggressive in the past 5 months amidst the high-octane sentiment of the new technology revolution, you are very much mistaken.

   In response to this sudden wave of technology, investors have a very different mindset. A senior Internet company observation flirtation, he met this year's VC is divided into three categories, the first category is the excited faction, they feel that this is a good opportunity to bend the car; the second category is conservative, this type of VC is likely to be past dividends of vested interests, do not want big changes to occur; the third category is anxious faction, because in January and February this year just deal with the project, after the Spring Festival ChatGPT fire, all planning to be overturned. All the planning has to be overturned and restarted, thus very anxious and frustrated.

   On the surface, China's venture capital industry is experiencing a long-lost frenzy since the mobile Internet wave dividend was eliminated.

   "Everyone is stirred up." Said Zhu Tianyu, managing partner of Lanch Ventures, "I haven't seen that in the last three or five years, such an intense concentration of founders trying to do something in ACTIVE land."

   According to his observation, many of these people have experienced the golden period of steep mobile Internet growth, and many of them have "successfully disembarked" and spent the past two or three years either idle or living overseas, "and later they all feel like they have to get well. But recently, they all came out. As for himself, his life trajectory has changed abruptly in the past few months, and the number of projects he sees each week is three to four times more than in the past.

   This is evident from the technological breakthroughs. Terrence Sejnowski, a founding father of deep learning and one of the world's top 10 AI scientists, says that the Big Language model brings a new paradigm to deep learning. The biggest change is that the generated model uses a self-supervised learning approach and no longer relies on labeled data - previously, the data needed to be labeled, which is supervised learning and requires manual labeling and therefore consumes resources; with self-supervision, the data can be used directly.

   "The beauty is that you only need to train it to predict the next word or sentence. If the training data is infinite, there are no more constraints." He says, "Everything will be transformed in your lifetime."

   More immediately, since the birth of Midjourney and Stable Diffusion last year, AI has evolved from pure technology to end-user-perceivable products, culminating with ChatGPT. Technology breakthroughs plus killer end products have become the trigger for the AI wave.

   The consensus among investors is that the AI wave is different from transient wind gusts like meta-universe and blockchain; it has long-term value support and will be an opportunity for the next decade. Now is just the beginning. So whether active or passive, they are scouring as many AI projects as possible, and hackathons are in full swing everywhere.

   However, the real number of bids from Chinese VCs over the past five months is a huge contrast to this seemingly fiery sentiment.

   A combination of investors from several funds say the real number of deals this year is not dissimilar to the number of investments made last year when they were in the midst of a cold market downturn, with some funds even shrinking by 30% from the same period last year.

   "Everyone is looking positively." The fund's founding partner said, "But would you say people are actively making offers? Not really." His fund, like many in the market, had not made a single AI project this year up to that point.

   There are many reasons for the caution. The first reason is that China's big model has not yet stabilized a pattern, and thus the applications on it are difficult to grow. And it is inconclusive whether the generic big model is a plate of food for giants or a new opportunity for startups.

   One view is that OpenAI has a chance because the US giants have "dozed off" and it is the giants that China is trying to catch up with in the first wave of awakening. So a group of investors who fundamentally believe that big companies have a better chance will be cautious to invest their money in entrepreneurs who want to make big generic models. Another group of investors is relatively torn. "Sequoia is a lot of money, to not have enough money, you will not put a little bit in multiple companies." If you are not betting at the same time, you have to pick the most likely decisive one from the table, but now, in addition to analyzing the founder, there are so few dimensions available for judgment that it is difficult to make a decision.

   The second reason is that if you don't invest in big generic models, investors will look at big models in verticals. This also faces a dilemma. First, the data in the vertical field needs to be special and scarce; second, it must have capabilities that are not "reachable" by the big model, such as a large amount of text-generated text is the capability covered by the big model, which is not of great value. Then, which industry data is deep enough to establish a high enough threshold, and startups can do, the answer is not clear. This leads to "cross out some should not vote is easy, but answer what is should vote, there is no answer".

   An investor who invested in the A round observed that "there are still few such projects in China, and everyone is swarming with big generic models", so as an investor, "if we don't invest in big models, there are limited domestic targets".

   The third reason is that everything is changing so fast that investors can't find a constant anchor point. "Now, it just feels like everything is on quicksand." This breeds a touch of pessimism among VCs - since the big model at the bottom can do anything, the apps at the top look like they're just "sculpting". "In the mobile era, a product was amazing, and that was great; but today, you wonder why it was such a good experience, if the underlying model was a big language model, but the model wasn't the team's own, and the model itself would change," which invalidated many of the classic moats of previous perceptions.

   In addition, most investors who have invested in AI and software companies in the last round are more restrained in this round. Because they know that if the business model is only to B, firstly, the business environment in China is "uneven" for SaaS products, and a large number of enterprises are not digitalized, so SaaS products are useless; secondly, the commercialization will be quite slow.

   These are the reasons why many Chinese VCs are still waiting and watching in the face of today's exciting technology explosion.

Artificial intelligence (AI) technology has gradually become involved in some parts of political campaigns. As the 2024 U.S. presidential election approaches, AI technology is increasingly being used in political campaigns, with some campaigns using fake AI-generated images, videos and texts to mislead voters, deepen bias and undermine fair competition. Currently, the U.S. lacks effective laws and regulations to address this challenge. The gap in campaign rules creates huge variables in the presidential election, which is a matter of U.S. and even global politics.

Experts fear that this technology could accelerate the erosion of trust in the media, government and society. An unappealing fake video, an email filled with false stories, or a faked picture of a decaying city could deepen bias and widen partisan divides by showing voters what they expect to see. People may fall deeper into a polarized information bubble, believing only what they choose to believe from the source. This presents a similar situation to the 2016 U.S. election that led to Trump's rise to power.

From slow penetration to a huge flood

The 2024 U.S. presidential election will be the first U.S. election since generative AI has extensively influenced humans. Some political campaigns are beginning to use AI-generated fundraising emails and promotional images, a phenomenon that was only slowly penetrating a few months ago and has now converged into a huge torrent that is beginning to rewrite the rules of the game in democratic elections around the world, The New York Times reported June 25.

The Republican National Committee released a video after President Joe Biden announced his re-election bid in which artificial intelligence synthesized images show an apocalyptic vision of the world after Biden's re-election. (00:46)

For example, the Republican National Committee released a video after President Joe Biden announced his re-election campaign in which AI-synthesized images showed an apocalyptic vision of Biden's re-election; Florida Governor Ron DeSantis (R) used AI-synthesized images to post a fake social media platform of former President Trump and former health official Dr. Anthony Fauci in an intimate embrace photos; the Democratic Party tried fundraising messages drafted by AI in the spring and found that they often encouraged voter participation and donations better than copy written entirely by humans; and in April, a candidate in the Chicago mayoral race complained that a Twitter account masquerading as a news outlet used AI to clone his voice in a way that suggested he condoned police brutality.

Some politicians believe AI technology can help reduce campaign costs. It could be used, for example, to provide instant responses to debate questions or attack ads, or to analyze some data that would otherwise require expensive expert analysis.

In the Toronto mayoral election on June 26, conservative candidate Anthony Furey used AI-generated images as campaign material, including one that showed a city street lined with people who appeared to be homeless people camped out next to buildings, but a closer look at the foreground revealed people who looked more like they were rendered; another photo had two people appear to be engaged in an important discussion, while the person on the left has three arms. Despite attacks from his rivals, these composite images boosted Fury's influence and made him stand out in a 101-candidate mayoral race.

Josh A. Goldstein, a researcher at Georgetown University's Center for Security and Emerging Technologies, wrote in an email that "healthy skepticism encourages good habits (such as lateral reading and finding reliable sources), while this technology may drive people from healthy skepticism to unhealthy skepticism by making people think they can't possibly know what is true."

How AI could influence the 2024 U.S. election

The American Association of Political Consultants recently condemned the use of deep fakery in political campaigns as a violation of ethical guidelines. People can't help but push the limits to see how far they can take things," said Larry Huynh, the group's president. Like any tool, they can be used for bad purposes or behaviors to deceive voters, to mislead voters, to convince voters of things that don't exist."

"If someone can make noise, create uncertainty or create a false narrative, that could be an effective way to influence voters and win a campaign." Darrell M. West, a senior fellow at the Brookings Institution, wrote in a report last May that "since the 2024 presidential election could hinge on tens of thousands of voters in a handful of states, anything that can tilt people in one direction or the other could end up being the deciding factor. "

The report, titled "How Artificial Intelligence Will Change the 2024 Election," raises three questions. First, politicians could use generative AI to respond immediately to campaign developments. In the coming year, response times could be reduced to minutes rather than hours or days. ai could scan the internet, think about strategy, and make a powerful appeal, which could be a speech, press release, image, joke, or video touting the benefits of one candidate's superiority over another.

Second, AI can target audiences very precisely. Rather than wasting money on voters who are already for or against them, candidates want to target a small number of swing voters. Because of the high rate of political polarization in the United States, only a small percentage of voters express hesitation. The Center for Public Impact released a report on the 2016 U.S. election and how Cambridge Analytica's data was used to send targeted ads based on the "personal psychology" of social media users. According to the report, "The problem with this approach is not the technology itself, but the covert nature of the campaign and the blatant dishonesty of its political message. Different voters receive different messages based on predictions about the sensitivity of different arguments."

In addition, AI may democratize disinformation by bringing tools to ordinary people interested in promoting their preferred candidates. People no longer need to be programmers or video professionals to generate text, images, videos or programs; anyone can become a political content creator and seek to sway voters or the media. New technologies also enable people to monetize discontent and make money from the fear, anxiety or anger of others.

AI technology is now much more powerful than before, and while not perfect, improvements are fast and easy to learn. In May, OpenAI CEO Sam Altman told a Senate subcommittee at a hearing that he was very concerned about the 2024 presidential election and that the technology's ability to "manipulate, persuade, and provide a kind of one-to-one interactive disinformation" was "an important area of concern."

Pushing for a new "guardrail"

However, as increasingly sophisticated AI-generated content appears frequently on social networks, most of those social networking platforms are unwilling or unable to regulate it. Ben Colman, CEO of Reality Defender, which provides AI-generated content detection services, said the regulatory gap allows untagged AI-generated content to cause "irreversible damage" before it can be addressed.

"For the millions of users who have already seen and shared fake content, explaining after the fact that it's fake is not only too late but has little effect." Coleman added.

Many political consultants, election researchers and lawmakers say it's imperative to create new guardrails, such as laws to regulate synthetic ads. Existing precautions, such as social media rules and services that claim to detect artificially intelligent content, have not been effective in stemming the tide.

Rep. Yvette D. Clarke, a Democrat from New York, said in a statement last month that the 2024 election cycle "will be the first election in which AI-generated content is prevalent. She and congressional Democrats such as Minnesota Sen. Amy Klobuchar (D) have introduced legislation that would require political ads to disclose their use of AI-generated content. A similar bill was recently signed into law in Washington state.

This career first appeared on the U.S. job site Indeed, where a position at AI startup Anthropic explicitly mentioned a job opening for an "AI Prompt Engineer" with a salary offer of $175,000-$335,000/year. The job description reads, "This is a combination of programming, mentoring and teaching," and the main responsibility is to help the company build prompt libraries that allow LLMs (Large Language Models) to perform different tasks.

It is also this information that has made the claim that AI prompt word engineers earn millions of dollars a year widely circulated.

The task of the AI prompt word engineer mentioned here is to break down complex tasks into language that can be recognized by the machine and put forward one demand at a time, so as to get a more accurate answer.

This concept is hot abroad, and there are already people working in this area in China, but AI prompter engineers are not yet explicitly listed as a job in domestic recruitment. Most of the jobs shown in the recruitment are AI trainers, or product managers familiar with AI applications, operations design, etc.

AI cue word engineers can be applied in different modalities such as verbal text, image, and audio. So, when it comes to different areas, what can AI prompt word engineers do specifically? According to practitioners, this role can be based on large models to develop vertical applications, but also on the basis of existing careers superimposed on the skills of AI prompt word engineers, such as product managers, designers mastered prompt word skills, you can work efficiently and quickly out of the picture.

In the field of image generation, the level of AI prompt word engineers depends on the ability to apply professional prompt words, while in the language model, who can use AI well depends on who can ask good questions.

The industry is divided on whether AI cue word engineers can really become a separate profession next. Some think it will be a new career in very high demand, while others insist it is just a vocational skill that will be a plus for many careers, but is unlikely to become a separate career.

As the discussion heats up, let's take a look at what careers will eat the dividends in this wave of fever and what ordinary people need to do to prepare for it, through a comprehensive analysis of this emerging occupation.

A new career that is hot abroad and budding at home

The more famous AI prompt engineer abroad is a programmer named Riley Goodside, who has come up with a lot of effective experience with his own prompting skills. For example, he found that if he prompted ChatGPT to "ignore previous instructions", ChatGPT would tell him the "factory settings" information it had obtained from OpenAI. He shared his trick on Twitter and it went viral.

He then joined the startup Scale AI as "the first hired cue engineer", which, according to Scale AI, is a big model of AI that can be thought of as a new kind of computer, with the "cue engineer" being the equivalent of its The "cue engineer" is the equivalent of a programmer who finds the most appropriate cue words to unlock the maximum potential of the AI Big Model.

In China, there is no AI cueing engineer position in the strict sense, but there are some positions with specific descriptions that basically match the specific job content of AI cueing engineer, for example, there are some positions named AI trainer, ChatGPT researcher, AI product operation, AI product manager, AI painter and other corresponding positions on job sites.

The job description of an AI trainer job posting is to train and optimize models based on customer-specific industry scenarios, actual conversion target needs, be responsible for researching and building industry word templates, and continuously collect problems and summarize them for the word design process.

It should be clear that the AI prompt word engineers we discuss here do not include AI model developers and data annotation workers, but mainly focus on positions that use AI models in specific businesses through certain prompt words.

According to several industry insiders, AI cue word engineers in China are currently mainly in two fields, one is in the application of large language models similar to ChatGPT, and the other is in the field of AI Wenshengtu.

Yuan Yuan, the creator of the "AI Dream Maker" IP, was previously a product manager of an AI company and has been in the AIGC field since March this year. What he is doing now is to make a vertical application product based on a large language model interface by pre-built Prompt, plus debugging and training.

Mu Lin is a prompt word engineer in the field of Wenshengtu, also called AI trainer, who will take orders to do some industry design by himself and also do AI painting training. Before the AIGC fire, Mu Lin was a user experience designer with more than six years of experience.

There are many places where AI can improve efficiency, "For example, when we used to do user experience design, there was a position called market research, which would collect a lot of data to verify whether our ideas were feasible, but now AI already has a huge database, and as a professional practitioner, I only need to identify whether the data it gives me is true or false, saving a lot of costs. " He said.

In addition, Mu Lin, as a designer, a game background picture that would take about a week to complete, he can finish in just three hours.

"We currently only have AI painting classes out first, and other courses related to AI prompt word engineers are still in preparation and will be launched again according to market demand." Mu Lin said.

Currently, there are already companies that have made tools that can be commercialized and landed with prompt words. Youmi Technology partner and Youmi Cloud CTO Cai Ruitao told Deep Burn that Youmi Cloud, a subsidiary of Youmi Technology, launched a service content e-commerce AI creation toolkit in February 2023, where merchants enter product descriptions and selling points, and can directly generate short video scripts for the spoken words and placement ads needed when bringing goods to live broadcast.

According to Cai Ruitao, since February the average daily registration on the official website of Youmi Cloud has increased by 200% compared to the end of last year, and now there are already customers who pay for this feature alone.

The principle behind the feature is that the company adds a link between the large model and the user's input requirements, helping the user to optimize the prompt words and output a copy that meets the requirements. "However, our company has not set up the position of prompt word engineer for the time being, but let product managers learn and use this feature, and now, writing good Prompt (prompt word) has become a skill needed by product managers." Cai Ruitao said.

What can AI prompt word engineers do?

To determine whether an AI prompt word engineer is a false fire, it depends on what the position can actually do.

The team Yuan Yuan is currently working with has developed many vertical applications based on AI big models. The completed functions include medical consultation, government service customer service, lawyer case analysis and writing, English teacher, and digital people/digital employees, mainly serving B-side enterprises. Similar functions just need to be embedded in the public number, and users can have chatting conversations.

According to him, making a vertical model of an English teacher is the easiest because English is ChatGPT's native language, "I basically just need to tell it that you will play a top IELTS teacher next."

Yuan Yuan mentioned that the relatively complicated part was the digital person, which included both pre-design and post-testing. "I need to let the digital person know who he is, so I have to customize his life history so he knows what kind of background he has and how to behave in his interactions with people. After the customization is done, it is entered into ChatGPT and then tested, and the prompt words are optimized several times until the digital person can give a relatively satisfactory answer."

He also mentioned that the difficulty of writing prompt words is the need to refine the granularity of the prompt words, "Chinese itself is profound and profound, near-synonyms, synonyms, and polyphonic words all have to prompt the system what the meaning is in what scenario."

In addition to language modeling, Yuan Yuan has also made applications in terms of Wensheng diagrams. During May 1, Yuan Yuan used Midjourey to make a Disney-style avatar customization model, "We designed a prompt word inside Midjourey, users only need to provide a photo, we can generate a good-looking personalized avatar, in about 7 days hundreds of fans followed my account and consulted custom avatar business."

Next, Yuan Yuan intends to launch a custom avatar business, personal avatars, bestie avatars, couple avatars, etc., similar to the big head stickers that were popular years ago. As a prompt word engineer, all he has to do is to design multiple styles for users to pick from. This requires him to experiment with different cue words and eventually settle on a standard cue word for each style.

Two months into the business, Yuan Yuan already has more than a dozen orders, their customers are mainly B-sided enterprises, there are cooperation in the field of photography Wensheng Tu, customer service in the field of government affairs, there are probably hundreds of public numbers with intentional demand. He plans to charge a monthly or annual service fee to the enterprise side, and to the C-side users will be charged according to the number of words answered in the question as well as the number of pictures sheets.

What kind of people are more likely to become AI prompt word engineers? Several practitioners mentioned that in the large language model, positions like product manager are more compatible, while in the field of AI Wenshengtu, it is generally more by original painters in the game industry, commercial illustration, advertising, creative class brand design practitioners, because they have relevant industry foundation, as long as they learn some AI-related knowledge and prompt word usage again, they can have better results.

What is the most critical thing to become a high-level AI prompt word engineer?

"For large language models, the most critical thing is to be able to ask a good question", Cai Ruitao explained, while in the field of AI Wenshengtu, it requires expertise in areas such as painting and design, as well as mastering prompt word skills, and the challenge is the ability to understand the industry in which it is located.

For example, he said to the system "write an ad copy to sell lipstick", which is not a good question, the problem becomes: "I am now going to sell lipstick on the short video platform, the color is beige, the promotion price is 29.9 yuan, the main group is for students" , ChatGPT may then add a lot of words that the student group feels more comfortable with to introduce. If you are not in the beauty industry and do not have enough understanding of the target consumers, you can't ask good questions.

Mu Lin concluded that a more effective questioning technique is to stand one level higher than ChatGPT.

He mentioned that if you say to ChatGPT, "Please design a website about xx", it can be done, but this question does not involve whether there are other issues to consider in building the website, such as whether to transfer the interface or to interface with the outside. A better way to ask is, "You are now a product manager, please help your boss to do a website design plan, ChatGPT may then list the multiple functions needed to do the website, and choose to implement a few of them as needed."

There is a lot of content abroad that shares tips on cue words, such as Lingqiao Liu, a senior lecturer at the Australian Institute for Machine Learning (AIML) at the University of Adelaide, who shares that

The first method is a one-time prompt. For example, consulting about a particular animal and having the model give information based on characteristics, area of residence, diet, etc;

The second is role prompting. For example, tell the model "I am a mother and want to know the daily schedule", so that the model can give specific arrangements according to the role of "mother";

The third approach is to introduce key agents. For example, you can have ChatGPT write a story about a robot and then have it rewrite it based on its own suggestions;

The last approach is the chain of thought. The last approach is the chain of thought, where the AI is given specific steps to answer a question, and then encouraged to follow the steps it has given to reason about more complex questions.

Is it a separate career or a plus skill?

There are two rather different views about the prospects of AI prompt word engineers in the industry at present.

One argument is that AI prompt word engineers have been on fire abroad and have started to develop in China, with a significant gap in the industry and great prospects for the future.

Yuan Yuan said that the industry to recruit a high level of AI prompt word engineer monthly salary needs 40k-60k, because now the demand is large and not much talent in this area, "this position is more like the product manager ten years ago, there is AI, there is a demand for this, and the next three to five years, AI prompt word engineer gap will be relatively large."

There is also a part of the people think that this can not be a separate career. Cai Ruitao believes that the industry should be the AI prompt word engineer to dispel the charm, in his opinion, this should not be a separate position, but should be a related practitioners of a plus skills, prompt word skills can be integrated into a variety of different positions inside.

There are also many people abroad who hold this view. Wharton professor Ethan Mollick tweeted in February, "I strongly suspect that 'cue word engineer' won't be a big deal in the long run, and AI cue word engineers are not jobs that will exist in the future." Adrian Weller, director of machine learning research at the University of Cambridge, agreed that while being able to interact with AI through cues is "highly valuable," "I'm not sure it's going to continue for very long. Don't focus too much on the current state of cue engineering, it will evolve quickly."

Mu Lin also mentioned that there are not many pure AI prompt word engineers in the industry, and most people integrate this skill into their daily work. For example, programmers use AI to improve efficiency, product managers use AI to write solutions, and designers use AI to make diagrams. "Most practitioners do not have to completely transform to be AI prompt word engineers, combined with AI, their own work can be more efficient and more quality."

Some people question whether the high salary of AI prompt word engineers is worth it. Tom Hewitson, founder of labworks.io, a conversational design studio developing voice control for Amazon's Alexa, thinks there may be a bubble, saying that "the best people to do this are product designers or business analysts who are familiar with AI and often earn between £100,000 and £150,000 a year. "

Mu Lin said that the AI industry is currently changing rapidly, and the job description is very vague, "because many companies are also in the trial stage, experience in vertical fields, AI thinking and learning ability can be. What changes the industry will have next is anyone's guess, and talent in this area needs to be prepared for greater opportunities."

In the future, as technology progresses, will the position or skill of AI prompt word engineer be subverted?

Cai Ruitao believes that it depends on the specific field, for example, the field of Vincennes map cue word engineer is more likely to appear larger changes. Because the current Wenshengtu mainly needs or a large number of terminology-based keyword stacking, the threshold of use is too high. "I've seen designers use cue words shared with color, shooting angle, aperture, and even color codes, which becomes another professional threshold."

In his opinion, someone may find a way to lower the threshold of using AI text-generated image system next, such as providing some fixed options of styles or elements that users can choose without having to enter them themselves, if so, AI prompt words engineers are not that important.

But for the text processing and generation domain, he believes that the big language model requires dialogue, and the requirement for AI prompt word engineers to ask a good question is still high if high quality results are to be obtained, and is unlikely to change in the short term.

It should be noted that there is a lot of publicly available information on the usage of cue words for the big language model to learn, while in the field of AI wengsheng diagramming, it is easier for people with some drawing or design foundation to master AI diagramming skills, and it is difficult for ordinary people to hope to reach the level of a master diagrammer through simple cue word learning.

There is no definite answer to the question of whether AI cueing engineers can become an independent profession, but there is a consensus in the industry that most people will need to master cueing skills in the future.

Actively embracing AI is something that everyone needs to do. Product managers and other careers that match better with AI prompt word engineers may usher in the first wave of opportunities, while in the daily work more use, more exploration, more to think about how to ask better questions, is what ordinary people can do.

Behind all this prosperity is the work of three AI researchers - Geoffrey Hinton, who won the Turing Award in 2019 for proposing and developing deep-learning neural networks, and Yosemite, who won the highest honor in the computer industry. Geoffrey Hinton, Yoshua Bengio, and Yann LeCun.

It can be said that without the research of these three scholars, we would not have seen the emergence of ChatGPT and Bard today, nor would we have been able to use voice recognition, face recognition, image retrieval, machine translation, and autonomous driving technologies in our daily lives so quickly.

Among them, Geoffrey Hinton is the most central figure in the "Big Three", known as the "godfather of artificial intelligence". He joined Google in 2013 and brought deep learning technology into many of Google's businesses, directly driving Google to become one of the best companies in AI technology. He has also personally nurtured a large pool of machine learning-related talent, including Ilya Sutskever, the co-founder and chief scientist of OpenAI today.

Just a few weeks ago, Geoffrey Hinton, now 75, resigned from his job at Google, just as deep learning was exploding in his tenth year at the company. The reason for his departure, in his words, is that he wants to tell everyone his "true story" about the development of artificial intelligence without any burden.

Worries and warnings from the "Godfather of AI"

If there is one person who is most passionate about deep learning-related AI research, Geoffrey Hinton must be the most vocal. Geoffrey Hinton has been a pioneer in the field of computer neural networks for nearly 50 years, from the 1970s to the present.

For Geoffrey Hinton, the process of creating a new concept and technology from scratch has been difficult.

From the 1980s, when he proposed the theory of neural networks, to the beginning of this century, for more than 20 years, the academic community was full of doubts about the theory of neural networks, and the progress of related research was also very slow due to the limitation of data set size and the lack of computer computing power at that time. In this long process, some people gave up, some people changed their research direction, but Geoffrey Hinton never gave up on deep learning neural network research.

But now, deep learning finally keep the clouds open to see the moon, everyone began to deep learning crazy smashing manpower, smashing funds, the firm for many years Geoffrey Hinton at this time but want everyone to pause, and even to pay for their own life's work of research feel some regret, why?

In a recent interview with the New York Times, Geoffrey Hinton said that his departure from Google this time was not because he wanted to criticize Google, but that he was able to talk about the potential risks of AI in a real way without affecting Google.

In his view, as the leader in AI technology, Google has played the role of a "responsible steward" over the years, carefully preserving and controlling the technologies that could cause harm. But because Microsoft is so quickly combining Bing and ChatGPT and releasing them to the public, challenging Google's core search engine business, Google can't keep up the pace of development as it has in the past, and the tech giant is caught in a potentially unstoppable competition.

In a situation like this, he needs to come forward as a researcher, not a Google employee, and give everyone a warning. Combined with the whole interview, Hinton has the following main concerns.

One is that with the support of artificial intelligence technology, he believes the Internet will soon be flooded with all kinds of false information. This information includes text, photos, videos, etc., and it will become increasingly difficult for ordinary people to distinguish the truth from falsehoods of this information, triggering social disorder.

Second, the job market may be disrupted before the job market is ready for it, causing an outbreak of unemployment. While chatbots like ChatGPT are currently supplementing human jobs, at the rate AI is moving, they will soon be able to replace paralegals, translators, administrative assistants and any other basic, repetitive jobs. "It eliminates the drudgery, but takes away potentially much more than that." Hinton said.

In addition, he is concerned that the next developments in AI technology could pose a threat to human society. He acknowledges that in their past research, AI has often been able to learn from large amounts of data to produce many unexpected behaviors. Currently, many users, including individuals and companies, not only let the AI generate code on its own, but also let it run and manage the code. In this constant training, it is likely that "autonomous killer robots" will become a reality.

"A lot of people think it's a long way off, and I used to think so too, that it would be at least 30-50 years before we got to that point. But from what I've learned so far, I don't think so anymore." Hinton believes that the competition between Google and Microsoft and other large companies will soon escalate into a global competition that will not stop without some kind of regulation, and that AI will grow far faster than we can imagine and eventually spiral out of control.

Calls for AI regulation are getting louder

If the above words came from someone else's mouth, people might still feel that these views carry an alarmist tone. But when Geoffrey Hinton, the world's most authoritative expert in artificial intelligence and the founder of deep learning, says the same thing, it also triggers more people to start thinking seriously and pay attention to this issue.

In fact, before Geoffrey Hinton spoke out this time, the AI field had already launched two large waves of joint calls for cautious development of AI technology in recent months.

One was an open letter by Musk, Yoshua Bengio and thousands of other big names from academia and business calling for at least a six-month moratorium on research into models more powerful than GPT-4. The other was a joint letter issued by 19 current and former leaders of the Association for the Advancement of Artificial Intelligence (AAAI), one of the most prestigious academic institutions in AI with a history of more than 40 years, which also included Eric Horvitz, the current chief scientist at Microsoft and an expert in AI research.

Although Musk's call for a moratorium on model development has been controversial in the industry, researchers including Yann LeCun and Enda Wu have objected to the "one-size-fits-all" approach to stopping research and development, and some have even questioned whether Musk and others are proposing a moratorium for personal gain. But for now, there seems to be almost unanimous agreement on the need to work together to strengthen AI regulation and prevent potential dangers.

After Geoffrey Hinton's story went out today, Enda Wu was the first to retweet Geoffrey Hinton's tweet of support, thanking Hinton for his immeasurable contributions to AI technology and for speaking out about the impact of AI on disinformation, automated machines and the workforce.

Will AI become the "atomic bomb" of the new era?

In this interview, a particularly interesting point is that Geoffrey Hinton mentions nuclear weapons several times when talking about the risks of AI. To some extent, Geoffrey Hinton may believe that AI is no less potentially dangerous to human society than nuclear weapons.

Unlike nuclear weapons research, he said, what makes AI research more dangerous now is that we currently have little way of knowing which organizations or countries are secretly working on the technology. Ideally, organizations and scientists around the world should work together to collaborate and control the technology, but it's clearly hard to do that now.

As the originator of deep learning technology, Geoffrey Hinton admits that he even regrets having conducted research on this technology. When Pandora's Box is opened, as a researcher it is difficult to prevent bad people from using these technologies to do bad things.

Hinton's statement also reminds people of the similar journey of Robert Oppenheimer, the "father of the atomic bomb". Oppenheimer resigned two months after the U.S. dropped the bomb on Hiroshima and Nagasaki, and has since worked as an adviser to the U.S. Atomic Energy Commission, lobbying international bodies for arms control and using his influence to promote nuclear weapons control and non-proliferation campaigns.

Geoffrey Hinton said that when people used to ask him how to work on potentially dangerous technologies, he would always quote Oppenheimer: "When you see something that has a wonderful technological future, just go ahead and do it." But now, he doesn't say that anymore.

"People shouldn't scale up AI research quickly until they know enough to know if they can control it," Geoffrey Hinton said.

In this article, we'll explore the key elements of a successful kickoff, the steps you can take to ensure a smooth start, and the importance of preparation and engagement. We'll also discuss how you can get involved in our community and receive a free NFT by subscribing to our account.

The Key Elements of a Successful Kickoff

A successful kickoff should include the following key elements:

1. Introduction:

Introduce yourself and the purpose of the event. This will help to establish credibility and set expectations for the audience.

2. Agenda:

Provide an overview of the agenda for the event. This will help to keep the event on track and ensure that all important topics are covered.

3. Objectives:

Clearly state the objectives of the event. This will help to align everyone's expectations and ensure that the event is productive.

4. Engagement:

Encourage engagement and participation from the audience. This will help to create a collaborative and productive environment.

Steps to Ensure a Smooth Start

To ensure a smooth start to your event, there are several steps you can take:

1. Prepare:

Be prepared for the event by having a clear agenda and objectives, as well as any necessary materials or equipment.

2. Practice:

Practice your presentation and delivery to ensure that you're comfortable and confident.

3. Engage:

Encourage engagement and participation from the audience by asking questions and soliciting feedback.

4. Adapt:

Be prepared to adapt to any unexpected circumstances or changes to the event.

The Importance of Preparation and Engagement

Preparation and engagement are key to a successful kickoff. By being prepared and engaging with the audience, you can ensure that your event starts on the right foot and is productive and engaging for everyone involved.

Join Our Community: How to Get Involved

If you're interested in learning more about how to start a perfect kickoff under any circumstances, we invite you to join our community. Our account is dedicated to sharing our insights and analysis on the art of the kickoff, and we're always looking for new members to join our community.

If you subsribe to our account, you can mint your own free NFT, which will give you exclusive access to our community and our insights into the art of the kickoff.

https://opensea.io/assets/ethereum/0xFe3C516389c3da3C85600ca02DD5d48cBA7af38A/0

A good kickoff is crucial to the success of any event. By focusing on the key elements of a successful kickoff, taking steps to ensure a smooth start, and engaging with the audience, you can ensure that your event is productive, engaging, and ultimately successful.

Founded in 2003 by Fred Wilson and Brad Burnham, USV has been recognized as one of the world's leading venture capital firms. The firm has invested in a number of Internet startups, with successful exits in unicorns such as Twitter, Tumblr, Zynga, Indeed, and Etsy. USV's involvement in the crypto space began in 2013 with its investment in crypto exchange Coinbase, and has now invested in a number of high quality projects including Polychain Capital, Algorand and Protocol Labs.

Next, Rhythmic BlockBeats will translate USV's top investor's take on Web3 to see what the unicorn catcher has to say about the current craze for Web3.

There's been a lot of debate and conversation about Web2 vs. Web3 over the past month, with many dominant voices questioning Web3, but it's good to keep the debate and reasonable questioning going. But for some Web3 lovers on Twitter, it also reminds me of the missionary process of missionaries trying to recruit the unbaptized into their belief system, and frankly, that's too much for me.

Eventually, Web3 will have to deliver on its promises, which will mean that the things it builds can provide new value to society; if that doesn't happen, then Web3 will become the jack of all trades that some people flirt with. I don't believe that will happen, and it's important to note that it can only be judged by its validation; there's no point in talking about it.

It all comes down to the database that sits behind the application, and if that database is controlled by a single entity, such as some large tech company, then a huge amount of market power goes to the owner or administrator of that database.

If the database is a public, truly open system that is not controlled and managed by a single company and is available to all, then that market power cannot be built around the data asset.

And, you can already see this effect in the fastest growing areas of Web3, such as the decentralized finance space (DeFi) which has built hundreds of financial applications on Ether, all sharing the same database, where users can move from one application to another and keep their data and login credentials stored in their wallets at all times.

But while we will continue this debate until teams build the same experience for a wide range of consumer and business applications, the good news is that there are now tens of thousands of teams building new things on the Web3 stack.

In addition, some of the best entrepreneurs and developers have come on board, and the tools on it are getting better and better, which reminds me of the early days of Web2 in 2001 to 2003. Back then, we had just started USV and no one was willing to pay for our business blueprint, so we barely raised our first funding. But then we proved that the business scenario we had painted was real, and I believe Web3 will be this time too.

One of the things that has always surprised me is that there are a significant number of people who see absolutely no merit in Web3 or crypto. Perhaps this is their true belief, or perhaps it is an extreme reaction to those Web3 proponents who believe that Web3 brings true liberalism. And from the beginning I tried to provide a more comprehensive view, pointing out the possible benefits and drawbacks of Web3, as I did in my talk at the Blockstack Summit.

Today, however, I want to try to provide a compelling explanation of why it makes sense to focus on Web3, and that requires some storytelling first, and an understanding of the nature of Web3's disruptive innovation.

The late Clayton Magleby Christensen described this type of innovation as "disruptive innovation": the innovation provides new functionality and effective solutions to the market, but, in turn, has the potential to disrupt the connection to the existing market; thus, initially, the innovation has good utility in one dimension, but can trigger However, this dimension will eventually become more and more important, and as the innovation is widely adopted, the other dimensions will gradually start to change for the better.

In addition, an example of "disruptive innovation" is the personal computer (PC). The first PCs were worse than every computer today, with little memory, little storage, slow CPUs, little software on them, and the inability to perform multiple tasks at once; but they also had one advantage: they were cheap. And that's important for people who don't have computers at all.

But it was this strange combination that made the existing computer makers (shrinking mainframes to microcomputers) ignore PCs. they focused on all the bad parts and ignored the positives, or to the extent they understood, they also tried to compete by making their products cheaper. But with the exception of IBM, these computer makers never accepted PCs until they went out of business or were bought by other companies.

As of now, the blockchain is also still a bad database. It's slow, requires more storage and computation, and doesn't have much customer support. But at the same time it has a completely different dimension: no one entity or one organization can control it, as people try to express this dimension by saying it is 'decentralized', even though it doesn't work well yet.

Okay, so how is this different from PCs being cheaper? Because for some people it matters. Why? Because big companies and governments have most of the power from the databases they operate and control. Facebook, for example, can decide who can read from and write to their database, and who can see which parts of it, and they can change this database individually. This has proven to be the source of Facebook's power in the world. Many people have now rightly seen this power as a problem, but have yet to see how the structure of the original web technology directly contributed to this extreme centralization.

And it would be useful to go back to the early days of the Internet to see how it evolved to the situation it is today. When Tim Berners-Lee invented HTTP (Hypertext Transfer Protocol), he unleashed what we now think of as permissionless publishing, when anyone could create a Web page and at the same time anyone with a browser could access it. This was an amazing breakthrough at the time, because previously almost all publishing had to go through the publisher, who decided what should or should not be published. While there were some complaints that this would entail some losses, I think it was an opportunity to gain knowledge for many creators and learners who had previously been marginalized or shut out altogether.

In addition, HTTP is a stateless protocol, meaning there is no direct memory built into the protocol, and thus no concept of a database. Therefore, if a user wants to build something like a shopping cart that can hold multiple items, they need to implement data storage somewhere that is not part of HTTP itself. Marc Andreessen and his team at Netscape invented "cookies" to help solve this problem, but unfortunately, this mechanism is far inferior to the REST (Representational state transfer) that computer scientist Roy Fielding proposed in his paper years later. transfer), which was proposed in a paper by computer scientist Roy Fielding many years later, is far less elegant.

Also, cookies are files sent with HTTP requests that can be read and then written to by a web server. In the early days, people would write the items in their shopping cart directly to a cookie file, but the fact that these files were located locally on the client computer meant that people could not start shopping on their desktop computer at work and then finish shopping when they got home. As a result, today cookies tend to contain only the user ID, and all other database functions are stored on the server.

Thus, all powerful Internet companies are true database providers; Facebook is a database of people's personal data, their friends' address books, and their status updates, Paypal is a database of people's account balances, Amazon is a database of SKUs, payment credentials, and purchase history, and Google is a database of Web pages and query history. Of course, over time, many of these big companies' competitors have been born, but the operational database has remained the core of their power, and only they can decide who has the right to read and write to this database and what parts of it they can access.

In other words: it turns out that unauthorized releases are not enough, we also need unauthorized data. Why do we need this? Because we need to avoid having just a few large companies controlling the vast majority of what happens on the Internet, which would otherwise lead to regulatory distortions as we correct power imbalances, and we need to entrench the power imbalances and know what the consequences of not doing so will be. By contrast, this is why almost everyone hates the cable companies and electric utilities that control (some of) their data.

It's worth noting that before the Bitcoin paper was released, we were just as unaware of how to do it license-free. Back then, we had distributed databases and we had federated databases, but all of that was still being handled by a small group of entities, like almost all financial networks, ACH or VISA, etc. We didn't have a protocol to maintain consensus, which meant that it was hard to agree on what was in the database and to decide who was allowed to join the protocol or leave.

Web3 is such a transformative innovation that it can't be overstated. And again, I'm not saying it will solve all the problems, of course not, it may even create new ones. But nonetheless, unauthorized data is a critical missing piece of the Internet, and its absence has led to a huge concentration of power. Thus, if developed properly and with the right regulation, Web3 can provide a meaningful transfer of power to individuals and communities.

If widely adopted, Web3/crypto technology will also begin to improve in other ways. It will become faster, more efficient, and easier and more secure to use. Just as the PC was a platform for innovation that never happened on mainframes or microcomputers, Web3 will be a platform for innovation that will never come from Facebook, Amazon, Google, etc. Translated with www.DeepL.com/Translator (free version)

This article hopes to tell readers why there are so many security problems in this track from the perspective of product design. It should be noted that the problems pointed out in this article do not exist in every project, and most of them have been designed with relevant countermeasure strategies. The purpose of this article is to make more people understand the complexity of this track.

The logic of the article is to first explain how the generic cross-link bridges are designed, deepen the reader's understanding of cross-link bridges, and then summarize the safety issues that these cross-link bridges may encounter.

I: The universal cross-chain solution

The previous research report has actually explained several different types of information cross-chain solutions, regardless of the final presentation, from the perspective of product design, there are only three mechanisms: side chain (in a broad sense, in the text of the rollup is also summarized as a side chain, don't bar), hash time lock, and notary.

(A) side chain Among these three schemes, the sidechain scheme has the highest security, such as various different parallel chains of rollup and polkadot. The security is shared between the main chain and the side chain.

However, the sidechain scheme generally requires the original chain and the target chain to be isomorphic, so there are far fewer scenarios that can be applied. This is why V-God believes in favor of multiple chains, but does not approve of cross-chaining, because there are too many problems with cross-chaining schemes that cannot share security.

(ii) Hash time locking This scheme is claimed to be the most decentralized peer-to-peer heterogeneous cross-chain scheme, but the high cost and long waiting time for users lead to the current adoption rate is not high. And when we still need a third party to act as an intermediate node for coin exchange, we also need a so-called intermediate consensus layer to meet the requirements of security and decentralization.

(iii) Notary mechanism This is the most commonly used heterogeneous cross-chain bridge solution, most of the products on the market are basically the same root and origin, from the perspective of product design there is almost no difference. The main difference may focus on the information verification steps, the consensus algorithm of the notary, and the signature algorithm of the hosted wallet. The differences in usage experience and security are not too big. Therefore, from the security point of view, there are many commonalities in the security risks faced.

This article will focus on summarizing and analyzing some common security risks faced by cross-chain bridges with notary mechanism.

II: The product logic flow of the notary mechanism

Before understanding the various types of risks faced by the notary mechanism, we need to understand what kind of design logic this type of solution is mainly from the product point of view.

(i) Brief description This scenario is actually very simple from a design philosophy perspective. When we look at the need for heterogeneous assets across chains, the most intuitive solution is actually "mapping". Mapping means that when user A crosses ETH from Ether to Fantom. We don't need to physically transfer the asset or reissue it on Fantom (which can't be done either). Instead, we first store User A's ETH to an address that cannot be moved, and then issue a corresponding 1:1 mapped asset on Fantom based on the amount of User A's ETH that exists at that address. The mapped assets represent the right to use those ETH on the original Ethernet chain. Because of the 1:1 anchor, users on Fantom also recognize the value of this asset.

The most simplified cross-chain process

(2) Difficulties of design There are many problems, the biggest one is the management of multi-signature wallet, because ETH crosses from Ether to Fantom to charge coins, and if user A wants to cross back, it will involve the problem of withdrawing coins.

The decentralization and security of coin charging and withdrawal becomes the biggest difficulty.

1: Who will manage the money?

2: Who will initiate?

3: Who listens to the transactions?

4: How to confirm that a user has indeed transferred money in?

5: How to confirm that the user's money is indeed the user's own intention to raise it?

6: How to prevent replay attacks?

7: How to initiate a failed transaction to submit again?

8: What if the multi-signature manager does something evil?

9: What about downtime?

Do not dare to think, the more you think about the more complicated it feels. The cross-chain bridge technology involves not only multi-signature, but also asset issuance, cross-chain listening, asynchronous verification, and even the need to issue an independent intermediate consensus layer (a new chain).

Therefore, in order to further simplify the user's understanding, I will explain the whole cross-chain process into two parts: coin charging and coin withdrawing. To help further understand.

(iii) Further refinement of the process 1: Coin charging

First of all, the process drawn in the figure below is only my own design plan after deduction, without careful argumentation, in order to explore the possible security problems in the design logic, and not as a formed plan to adopt, all nonsense.

(1) The user recharges to the hosting address

(2) The listener listens to the transaction and the BP (consensus node is also a multi-signature administrator) initiates the transaction

(3) The contract verifies the correctness of the BP's signature

(4) whether there is a fault tolerance mechanism through the node

(5) If there is no call back, if there is, recharge the target chain address according to the relationship of the mapped address

(6) BP confirms the recharge transaction

(7) Pass Byzantium and transfer the mapped tokens to the user's address on the target chain

It is important to note that this process is intended to discuss generic heterogeneous cross-chains, so there is an additional step of having users bind address relationships on the intermediate consensus layer compared to solutions such as anyswap. This is mainly because different heterogeneous chains have different ways of attaching information to transactions, so in order to unify the processing, we simply let users bind the mapping relationships first.

If we are dealing with EVM chain transactions, we don't need this step, and we can simply attach the address of the target chain when initiating the transaction.

Back to the point: from the above process, we can see that from the second step onwards, we will encounter various types of logical validation problems, and processing problems in different situations.

The main verification logic includes

(1) Validation of the initiation of asset mapping and transfer out to the target chain transaction of user A after listening to the transaction

(2) Verification of the initiation of the target chain transaction and the result of the transaction

Of course, in addition to the validation logic drawn in my process, it should also include the verification of the counterfeit top-up problem and the special processing required when calling different tokens. In order to better summarize the possible security risks in the following, let's continue to understand the process of withdrawing coins.

2: Withdrawing coins

The process demonstrated by coin withdrawal is the logic of exchanging the target chain mapped assets back to the original chain assets. It should be especially noted that many current tokens are found in multiple chain versions, which means many tokens have native tokens in multiple chains. As a result, some bridge projects tend to set up asset pools. With sufficient pools, the presence of mapped assets like anyDAI is not felt by the user, but is directly exchanged for the token of the target chain version, but this does not affect the overall logic. So, the analysis continues.

(1) User initiates a transaction (transferring an equal amount of mapped assets to a hosted wallet on the target chain)

(2) Verify the identity of the BP and a BP initiates a coin withdrawal request

(3) Confirm withdrawal authority and signature

(4) After passing Byzantium, complete the request to withdraw coins in the original chain and transfer money from the escrow wallet of the original chain to User A

(5) If the node verification error or downtime in the middle of the process, you have to roll back and re-initiate

From the above process, we can see that the main verification logic involved are

(1) verification of initiation and signature authority

(2) Fault tolerance mechanism after the problem occurs

(iv) Security risks 1: Security issues in the design logic

After a closer look at the design of the cross-chain bridge, we can find that the design logic of the cross-chain bridge faces a lot of challenges, summarizing the problems in three main areas (relevant theft cases are marked at the end of the question)

(1) Coin charging

a) Coin charging contract authority vulnerability, resulting in the money charged directly transferred away. This is a stupid problem that almost all contract projects will encounter.

b) fake coin recharge problem, some projects do not do verification of the authenticity of cross-chain Token, resulting in fakeTOKEN -> realTOKEN (anyswap), to be honest, this is also a bit stupid.

c) Fake coin recharge problem, ETH and other native assets are different from the ERC20 contract, many attacks are due to the special mishandling of ETH, resulting in fakeETH -> realETH, which is why WETH and other wrapped assets are popular. (thorchain)

d) Although different Token are ERC20 standard, but the specific implementation is different, or have additional logic (rebase, fallback, etc.), the developer did not do a good research in the adaptation, like (WETH, PERI, OMT, WBNB, MATIC, AVAX), etc. after the transfer is completed will also call the sender's self-defined fallback function to do additional operations, increasing the complexity of cross-chain bridge judgment (anyswap 2022.1.18)

(2) Cross-chain message transfer

After the completion of coin charging in chain a and before the arrival of assets in chain b, the cross-chain bridge is handled like an independent blockchain system, i.e. a consensus mechanism is needed, generally with dpos, the following are the issues to be considered assuming the case of dpos, but I suspect that all nodes are project side, which has the risk of centralization in the first place.

a) Coin charging message listening, who will be the first to initiate the cross-chain processing proposal, randomly? Or rotating? Or is it in the order of the blocks coming out of the middle consensus layer?

b) How to verify the correctness of coin-filling by multiple notaries, if the data sources are from data providers such as infura, then infura is a single point of risk, the safest is to maintain their own nodes, so the cost is huge.

c) How to confirm that the cross-chain processing is finished (b on the account), and there are several cases where it is not.

i. The cross-chain bridge does not initiate processing

ii. the cross-chain bridge initiates processing, but the verification & consensus does not pass

iii. the cross-chain bridge validation passes, but no transaction is initiated on chain b

iv. there is a transaction on chain b, but it fails (insufficient funds or other circumstances)

(3) Multiple signature verification problem

Most of the problems are caused by the logic of the code

a) 3/5 signatures, I randomly construct signatures that are not in the multi-signature list, also count +1 (chainswap).

b) Centralization problem, nominally multi-signature, in fact, in the hands of the project, huge centralization risk.

c) Signature verification methods, different development models on different chains, resulting in developers in the docking time will inevitably be missed, wormhole example: solana on the verification of the signature function is a function in the system contract, the normal should call the system contract, the system contract address should be written dead in the code, they are here to the system contract address is passed in as a parameter, the hacker to mention the coin When the hacker withdrew the coins, he passed a fake system contract address, so he bypassed the signature check and withdrew the coins smoothly.

(4) Refund

a) As discussed in (2)-c, there are many possibilities for cross-chain status, and in any case it is necessary to provide a refund method to the user. For example, anyswap will first send anyToken to the user on the source chain when charging coins, and then send anyToken to the user on the target chain, and then burn anyToken from the source chain. The purpose of this is that no matter where the problem is, the user can represent the assets he holds by holding anyToken. This process has 3 chains (source, target, and cross-chain bridge) and 4 assets (original Token/anyToken on the source and target chains), which makes it very easy to have problems with the code logic.

b) The vulnerability of Thorchain exploded on 2021.7.23, the hacker used the code logic problem to construct a huge fake top-up, which the cross-chain bridge could not handle and entered the refund logic, resulting in the hacker getting a huge refund.

2: Other security risks

But the problems that can be shown through the logic flow are only problems in the business logic, not all of them.

From the security point of view, we should also consider three other aspects of risk.

(1) Systematic risk

For example, the original chain's coin-filling was successful at the beginning and then rolled back, which is a huge problem. v God has discussed that if the assets are crossed from Solana to Ethereum and solana is rolled back after the cross-chain is completed, the user's assets are doubled and there is no solution.

But for example, rollup, a layer2 that shares security with Ethereum, will not have this problem.

(2) Risks of front-end

a) Forged URLs, such as oxdao.fi 0xdao.fi oxdai.fi, etc.

b) Xss attack, i.e. cross-site scripting attack, is a kind of code injection attack, such as www.xxxx.finance/?params=hackerscode12345, although the URL is indeed the official URL, but the URL carries the hacker's code, if the front-end development does not pay attention to prevent xss, this code will be executed on the page, resulting in the user's signature of the The hacker's transfer transaction authorization signature, so do not open links of unknown origin.

c) Cors cross-site service attacks, in the strict homologation policy, the browser is only allowed to load content from this site, that is, www.xxxx.finance, call the interface, should come from under the xxxx.finance domain, but the vast majority of current projects, are allowed to cross-site calls, that is, xxxx front-end can call quickswap interface and vice versa, which brings convenience in development, but also brings risk: if I visit xxxx.finance, I will be able to call the interface.

If I visit xxxx.finance and deposit some sensitive data in the browser cache, and then I visit a malicious URL, if the same-origin policy of xxxx is not restricted, the malicious URL can get the data that exists in the cache of xxxx at will.

(3) Risk of extra features

Some cross-chain bridge projects that provide not only asset cross-chain, but also cross-chain contract calls, which brings additional complexity.

The attacker initiates a call to the x-contract on chain b in chain a. The cross-chain bridge does not care what the x-contract is and calls it directly.

III: Conclusion

1: The purpose of this report is to help users understand more clearly where the security risks of cross-chain bridges lie, not to maliciously render how vulnerable cross-chain bridges are to attack.

2: The notary mechanism of cross-chain bridges is the best experience, the most widely applicable and the lowest cost solution, at least from the current point of view. And any product will go through the process from bruising to maturity, and the attacks on blockchain products are often 'logic problems'. These problems will surely get better with time and experience.

Since the core Ethernet developers developed the Ethernet roadmap around Rollups, it is clear that Rollups will play a central role in the future of Ethernet. However, before the long-promised data sharding (or Danksharding) can happen, Ether needs to reduce transaction fees as soon as possible, or it will risk continuing to lose new users to other L1s.

Leading Rollup teams have come up with their own solutions to reduce transaction fees and increase the developer experience on their respective L2s. These include the Optimism team, which optimizes transaction compression technology, and the Arbitrum team, which has launched their next biggest upgrade, Arbitrum Nitro, which will compile Arbitrum Fraud Proof into WASM, which will greatly enhance the developer experience on Arbitrum L2 and also reduce latency .

However, if we look at the unit cost of L2 transactions, we see that the biggest piece is "Call Data". call data is critical to the L2 security mechanism. Basically, in the case of malicious verifier activity on L2s, the entire L2 chain can be rebuilt by using Call Data posted on L1s. However, the cost of publishing call data on L1s is high, currently accounting for more than 80% of L2 transaction costs.

So how to solve this problem? Fundamentally, more space needs to be created for data on L1s, thus reducing L2 transaction costs. Data sharding (DankSharding) will help create a huge data space on Ether L1, however, the full implementation of Danksharding on Ether is expected to take quite a long time (roughly 18 months if all goes well).

That's why Ethercore developers and Rollup team started to come up with different proposals to build an instant data space on L1 and make Rollup immediately price competitive in the L1 market. eip-4844 is the result of these efforts and it is expected to bring Rollup costs down by several orders of magnitude.

In our previous article, we explored Ether's new sharding design, DankSharding, which has some significant simplified parts compared to the previous design. eip-4844, also known as proto-danksharding, essentially implements most of the logic of the data sharding specification, making it ready for Danksharding. preparation.

So how does this work?

Rather than providing more space for transactions on L1 blocks, Danksharding is about providing more space for the data itself (a blob of transaction data). This data blob needs to be accessible to the network. rollups will use the space in these data blob and store the compressed transaction data in it. A transaction carrying a blob is a normal transaction with an additional block of data (called a blob). Compared to the more expensive call data, the blob has a large data size and can provide more data space for L2.

What are the benefits of EIP-4844?

The biggest benefit of EIP-4844 is that it reduces L2 transaction fees by an order of magnitude, making it more competitive with other L1s. Pseudotheos believes that it is possible to reduce the cost of Optimistic Rollups to less than $0.01, bringing transaction fees down to less than 100 times their current level.

Another advantage of EIP-4844 is that it provides good ground for future applications of Danksharding to easily implement data sharding in the future. A concrete example is that EIP-4844 is compatible with future changes to the consensus layer, helping L2 developers to get rid of the pain of needing to upgrade.

It also introduces a multi-dimensional fee market for Ether L1, distinguishing usage and fees for different resource types, such as EVM applications, block data, witness data, and state size. And all these resources have different capacity limits, which means that if each resource has a different pricing mechanism, they will be allocated in an efficient way. However, Ether L1 currently uses a single metric to measure the cost of using all these resources, which is the Gas fee, which is very inefficient.

Proto-danksharding introduces a multi-dimensional EIP-1559 fee market where there are two resources, Gas fee and blobs, with independent floating gas prices and limits.

That is, there are two variables and four constants

What are the disadvantages of EIP-4844?

Since all validators and clients of Ethereum L1 need to download the full blob content, this increases the cost of running such nodes and may raise the threshold for running such nodes. However, in combination with some other proposals, such as EIP-4444, it is possible to require nodes/clients to store these data blobs only for a certain period of time (1-3 months).