Applications, challenges and prospects for generative AI

Machine-assisted translation of the archived article. Read the Chinese original.

Are you really ready for the Aigc era, aigc artificial intelligence generation of content, and relatively earlier content production patterns, which are made by experts, pgc, user-produced content gc,

And with the emergence of ai j c and an exponential increase in content productivity, this video tells us from the following sections about what the development of ani j c means to ordinary people.

In July 2015, Google launched dj to create a first-name model, and in January 2021, an OpenAI launched a humane model based on the text.

And it's like, "What's going on here?"


In July 2022, Midjourney, an independent lab, launched the same image generation model, Midjourney, designer Jason allen,

Using the images that were generated, the Space Opera, winning the first prize in the art competition at the Colorado Expo, and expressing their outrage that one or four, humans lost, caused global debate and discussion.

In August 2022, the open-source image generation model of the stability ai, stable diffusion, could be operated by a single computer, as of October of that year.

More than 200,000 developers have downloaded this series of models, detonating the ai painting field, marking the penetration of artificial intelligence into the art world, while in December 2012, the country’s most recent development project was launched on the Internet.

An OpenAI's large-scale language generation model trigb t sets off the entire web, and it allows us to communicate with humans as they react, and it enables us to use short-text poetry codes.

The different types of content output, such as computing logic, make one wonder if the ChatGPT is really human intelligence, just a few months later, but it's not enough.

An upgraded version of GPT4 was born out of nowhere, and not only did it achieve a digital multi-modular input, but it also showed that in the field of expertise it was more human than human beings, even beyond humans.



Redwood Capital also wrote that 2022 is the year of aigc, representing the beginning of a new paradigm shift, and that the ai model has had a surprising effect on the production of text pictures.

High-level and complex areas like video and 3d are in the process of being explored, but taking them down is just a matter of time for the speed of learning.

The reason for the outbreak in 22 years is to have at least four conditions, and you might ask, as a normal person, what's the use of this, to sell a little bit of it.

The first is to talk about in-depth learning technological innovation, unlike other industries, where artificial intelligence is open science, constantly innovative algorithms, and high-quality model open source that accelerates technological change.

In this case, an OpenAI multi-mode pre-training model that allows both natural language understanding and computer visual analysis so that model training can be directly used.

Images on the Internet, with their own textual descriptions, have significantly lowered the data acquisition threshold and are a technical basis for image generation models.


Two-synthetic data-driven training, the biggest constraint of early model training is that data face, data labeling is time-consuming and with low accuracy, sample numbers do not cover marginal cases, high-risk scenes, data are difficult to collect privacy, data face problems such as legal and regulatory constraints, but with the development of i, synthetic data from model outputs and anti-nursing models, the process of training leads to accelerated flying wheels, it can be said that synthetic data affect the future of ai, three gpu chips support us to feel ana through the network, but real model training requires huge investments, driving giant servers. It also directly makes the gpu chip and server manufacturer nvidia, the world's most valuable company, and the four industrial ecology is gradually taking shape, starting with an upstream pre-training model, where over 80 per cent of artificial intelligence research is concentrated on basic model training, a level that has high thresholds, such as the GPT3 training costs for language processing models, which are roughly estimated at $12 million, so those who are able to make pre-training models, mainly head-tech firms and unicorn companies.


The second is midstream, which is based on good pre-training models, developing and selling intermediate models in vertical subdivisions, models and services, model as a service becomes a reality, a level that is more suitable for start-up companies and individual entrepreneurs, based on gpt models, analysis of qualitative data from supporting firms, direct research based on academic papers, etc., and so forth, and lastly, downstream providers or individuals who use the ai service, people come to ChatGPT, people combine multiple ai tools to make videos, people take advantage of the aai effect map to make money.


Some people find that the hints are a lot of science, a specialized study, and even some forget it by breaking the law, but in short, it's time to open up the door to the new world, and know what's going on, knowing that when someone invites you to invest a million dollars in their garages to train the Ai model, make a big profit, you can give him an elegant fork out, which is good, not knowing that, in retrospect of our country's autonomous aai advantage, which may be a huge population size, without the possibility of professional and training data, but with no regard for the disadvantages, the high training costs of one of them, making ordinary businesses look forward to it. Since 2018, the United States has restricted not only China's imports of high-end chips, but also China's access to the tools to produce them. Perhaps when the United States made a decision, it was possible to calculate the risks by using an ai model, and this set of Chinese research and development is a long way to go. In the business sector, a mass-production design program can be produced and its value pre-judged. For example, the Japanese Society Plug provides a packageage design ai, which requires only the uploading of material by the user, which automatically produces thousands of packaging designs within an hour, and calculates their popularity in 10 seconds, when Japan's snack carder Caraby used to work with them to optimize the performance of new packagings, an increase of one hundred and thirty-two percent over old packagings, the restoration of valuable information, and the 100% use of file models to regenerate the map of the Fucheng Mountains. The use of ai to help repair the mural painting has greatly improved the efficiency of the protection of cultural objects.


On April 1st, 2022, the emotional singing bond of the Hong Kong stock financed by Ai technology, which makes thousands of civilized tears with 4k high-level quality, perhaps in the near future, will be able to use vr to take you through the scene of a concert in the early centuries, where she sings the triple digital twinning technique, makes full use of the ai model, finishes modelling of real scenes in virtual space and develops the life cycle of the entity, in product design, medical analysis, high-risk jobs, intelligent city construction, where Tesla is building a huge model that is more secure and efficient. To achieve its training goals for auto-driving cars, the content base of the four-dollar universe, and the neural universe not only has users, but also has many intelligence bodies like npc that provide interactive services to users, which are very much in need of ai-model production and drive, digital extension of five human lives, frequent thunderstorms in recent years, collapse of houses leading to brand damage, sending virtual people as spokespersons, risk and efficiency can be much more manageable, and for sick people, ai can generate therapy, even voice, help teachers and students talk, and the world's first Saborge Peter Scott Morgan, I'm not sure what I'm talking about.


In addition to living people, ai technology can learn about human videos, audio recordings, and digitally recreate the lost scene, as early as September 2013, when the Hollywood independent visual effects studio dj-dj-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-d-e-d-d-d-d-d-d-d-d-e-d-d-e-d-e-d-e-d-e-e-d-e-d-e-d-e-e-e-d-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-


To takechatgpt, it can randomly generate a lot of what appears to be right, but looking at the wrong answers, and if these errors are carried to the web platform without any brain, they seriously interfere with the accuracy of Internet knowledge, although many serious technology exchange sites have banned users from sharing their ChatGPT responses, but the status quo is difficult to distinguish between the creators of this content, who are people or humans, even if OpenAI produces its own content identification program, with less than 30% accuracy, people are excited about using ChatGPT learning, but the danger is that unless you already know the answer, you cannot judge when it's producing the wrong result. Many outlaws use open-source ai models to create malicious content with lower thresholds and more efficient ones, such as stable diffusion, which is soon used as a coloring platform, claiming that more than 4 million coloured images have been produced so far, including many American politicians who are being misled.


At the same time, with the help of ai learning and falsifying audio and video content, using personal identities for extortion, in 2021 there were fraud syndicates who used an ai to change their face, elon musk, six months defrauding digital currency worth more than 200 million yuan, personal privacy debris, as we surfed the Internet, the ai model, the powerful ability to reason that complete personal data could be assembled by means of multiple cross-checks of information, which would be inexcusable if abused by outlaws, and under the industrial structure of models and services, there would inevitably be some internal security problems with sophisticated ai models, which would be subject to attack or data poisoning. The question of how to remove toxic data and ensure the security of user data is a huge one, as well as an ethical challenge, training models are based on human world data, so they naturally perpetuate and amplify human society, existing social prejudices and stereotypes, OpenAI's ChatGPT and dall-e-2 models, with clear religious prejudice and gender discrimination in text and image generation, and with a model that can easily be taught bad, Microsoft's chat robots that are taught less than 24 hours, with a very anti-Jewish, homogenous discrimination, gender discrimination against girls, intellectual property rights issues. The ai model learns a lot of human work data, and inevitably imitates the personal style of an excellent creator, which is likely to be a copyright violation, and also works produced by an ai model, which belongs to the user training model company, which creates a designer or ai model, which will itself trigger controversy, the human aspect, the ai model is used to revive dead actors, create the possibility for people to speak with the messenger, but the deceased does not, nor can they exercise their right to consent or reject it, so some scholars believe that this is detrimental to the privacy and agency rights of those who lived, if the ai model is self-aware, not only per day, but also to read the tragic world. In 22 years, Googleai engineers claimed that their ai chat robot lmbda was self-conscious and once again pulled the issue on the table.


The three environmental challenge ai models, whether training or running, will consume a lot of resources, and the ceo sam Ottman of OpenAI also says that, while ChatGPT chat costs only a few cents per time, with the proliferation of users, it will have to suspend the free-of-charge model because of costing, for example, the drawing model Midjourney of people, there are hundreds of thousands of users who are taking on very high numbers, and if the number of users reaches 10 million, the world does not have enough to support it, gpt3 model training room, CO2 emissions are 552 tons, immediately after the model is commercialized. The more people use CO2 emissions, the more businesses around the world are now looking for zero carbon emissions, and we're burning a hole in the Earth by talking to one of them.


The challenge is to train an excellent ai model, which is far beyond the reach of individuals, and even a specially funded tertiary institution like Stanford, unable to build supercomputers based on immedia chips, economists are concerned that the efficiency obsession with the ai will drive the whole society into the Turing Trap, that more and more people will lose their jobs, that their ability to share the benefits of automation will be hampered, that the wealth and rights will be in the hands of fewer and fewer people or companies, and that it will be difficult for ordinary people to cross class, and another problem is the threat to national security, since aai can serve both businesses and individuals. It can also be a way for countries, such as creating super-power computer viruses or creating false information, with low-cost scale awareness campaigns, just as military powers insist on having their own satellite nuclear weapons, they want to stick to their own brains, and have powerful ones, in other countries, it creates a fallback, and finally the threat to the survival of all humanity.


And I don't think it gets direction, you think there has been an impact to people, how to alert one has become, and there are much more affected views for that they are coming out, and the danger of one is so much better than the future, wahblawarned company of poker for people, matters of godlike digital, supercharges, they take off the world, what I would ai is far more dangerous at least when, there's an evil spectator that you respect to die, but for an ai, there would be no dai could look forever, you have a predictable for, which we can never escape


Imagine, if a lab model were to build a model that could make a better model, and this better model was to build a better model, and they all grew up very quickly, which made me wonder about the woman's house Mary Shelley, 1818 novel Frankenstein, scientist Frankenstein, putting together an autonomous human form of life with a dead person's organ, initially being kind and learning fast, but without understanding in society, eventually becoming a social order breaker, Frankenstein, who had at first acquired the capacity to create and flourish, but he found guilt in the problem. To track human beings until the Arctic is far and wide, who knows exactly where things are going, will they play?


The moment he really came, perhaps only a very short moment, if no opportunity is seized, humans will be left behind by one of them, and then they will only pray for their own creation, God be merciful enough, when humanity is on the eve of a great change, and the era of one is coming, and change will inevitably be painful, like any great innovation in history, where humans will have to find a way forward, regardless of the outcome, and no one can predict the future, but there is one thing that can be sure, that there is no going back!