In recent months, people have begun to fear that workers in professional fields will be replaced while being swept by the practical application of AI technology. How to calmly face the opportunities and discomforts in the AI era? I hope that the author's explanation in this article can bring you some inspiration.
In 2013, the iPhone equipped with multi-touch screen technology created a new smartphone interface interaction mode, which gave birth to the mobile Internet.
And the AI painting at the end of 2022 to recent months, ChatGPT-generated articles, Copilot-generated code, Midjourney, Stable Diffusion and DALL-E can quickly generate pictures with just some keywords, and people are beginning to worry about whether most of the work of human beings will be replaced by artificial intelligence in the near future.
At the same time, various industry giants have also come down, the former Google image generation model Imagen, Meta's text-to-video model Make-A-Video, Microsoft is another way, DALL-E 2 built into the new service Microsoft Designer to meet the daily design needs of ordinary users such as invitations and postcards.
The convenience of technology comes with the fear that professional workers will be replaced.
This time, let's talk about the opportunities and discomforts in the AI era.
The emergence of AI super applications such as ChatGPT, Mid-Journey, and Stable Diffusion marks the maturity of accelerated computing and AI technology has entered the maturity, looking at a series of unfamiliar words, we who have done Internet workers, although we don't quite understand what it means, but also really feel that our lives may change and fear like the movie "I, Robot".
AI is infiltrating all walks of life at an unprecedented speed, driving a new technological revolution.
After the continuous development of AI technology in recent years, powerful computing power and advanced large-language models provide a suitable application platform for AI, prompting various head manufacturers to reconstruct their products and business models.
On the other hand, the pursuit of safety and stability in human nature has led more people to wait and see the AI tools and the content they produce or resist new changes.
, with the resources invested by various head companies, such as NVIDIA, Microsoft, Google, Adobe and domestic Internet manufacturers have withdrawn from their AI services in a similar time, in today's era of involution [if you don't want to be subverted, you must subvert others first] Perhaps we should first put aside doubts and resist and understand the new technology this time.
01 The history of AI technology
People are saying that this year is an era of AIGC (AI generated content) outbreak, and the most discussed tools ChatGPT, Stable Diffusion, Mid-Journey, etc. can be called AI, so what is AI?
AI refers to the use of artificial intelligence technology to produce content, that is to say, the information content we browse on the Internet is produced by various types of artificial intelligence tools from people [UGC (User-generated Content) user-generated content, mainly characterized by advocating personalization] or institutions [PGC (Professional Generated Content) specializes in producing content, content settings and product editing are very professional].
1. From deep model concept GAN to Transformer deep learning model architecture
The concept of AI was conceived in 2014 by artificial intelligence expert Ian Goodfellow after a drink, which can be based on the deep model concept of CNN (deep convolutional neural network, starting to learn an unfamiliar thing through a problem and extracting features) GAN [Generative Adverserial (Adverserial) Nets], through GAN to pit two neural networks, that is, generator and discriminator.
The generator generates images through the input is used to generate "fake data", the discriminator is used to judge the authenticity of the data, in the training process, the two alternately, so that the image generated by the generator is more and more realistic, and the discriminator is more and more discriminant, if you want to know more about the GAN algorithm model can click to view a more detailed explanation, involving the model algorithm and other professional content here do not explain too much.
GAN algorithm as the basis of the AI technology revolution at that time, and later widely used in image generation (AI face change), high-definition reconstruction, black and white film coloring, video generation, speech synthesis, image style conversion and other fields, known as one of the most powerful algorithm models in the 21st century, Ian Goodfellow has also become one of the most well-known experts in the field of AI.
Since 2015, GAN has begun to be put into practical use, and related papers have also exploded, and have also become the most common algorithm model in AI to generate images and process images.
In the same year, OpenAI, a company invested by Silicon Valley bigwigs such as "Iron Man" Elon Musk, Y Combinator President Altman, angel investor Peter Thiel, and so on, was established.
At that time, the original intention of OpenAI was to prevent the catastrophic impact of artificial intelligence and promote the positive role of artificial intelligence.
In 2016, OpenAI launched its own AI training architecture, Universe, which uses games, web pages and other applications around the world to measure and train the ordinary intelligence of AI through the network, so that AI agents can solve any problem by using computers like humans to learn themselves in any dimension through virtual keyboards and mice.
Although AI technology at that time was powerful in a specific field, it could not work once it exceeded this specific field, that is, AI technology could not escape the category of "narrow AI", such as AlphaGo can easily win anyone in Go, but there is no way to play other board games.
It wasn't until 2018 that the development of the Transformer (Deep Learning Model) architecture changed the development of NLP (Natural Language Processing) technology.
In the field of NLP, there are three main feature processors - CNN, RNN and Transformer, Transformer abandons the traditional CNN and RNN neural network, the entire network structure is completely composed of the Attention mechanism and the feedforward neural network, making the Transformer different from the CNN one-way processing data, can process all input data in parallel faster and more efficient.
On the other hand, Transformer's self-attention mechanism improves performance in language translation and sentiment analysis tasks by capturing the relationship between words or marks in sentences, and better handles long text or speech sequences in natural language tasks. To a certain extent, it solves the problems of low computational efficiency of long sequences and disappearance of gradients in traditional RNN and other sequence models.
At this time, OpenAI also launched the model GPT series created in the field of NLP (natural language processing), and the first product, GPT-1, was officially launched in 2018.
GPT-1 differs from other AI models in that it is "semi-supervised", in the previous NLP model, AI needs to learn large-scale data based on specific tasks, and these data need human "supervision" and data annotation.
GPT-1 can start with unsupervised learning pre-training, enhance language ability through data learning, and finally perform partially supervised fine-tuning. Simply put, GPT-1 can learn more efficiently with fewer resources and data, but GPT-1 at that time was not good on the one hand due to limited training data, and did not reach the ability to dialogue.
In 2020, OpenAI launched GPT-3, which has more than 175 billion parameters (GPT-2 has about 1.5 billion parameters) compared to the previous two generations, which is equivalent to one-tenth of the neural connections of the human brain. On the other hand, GPT-3 uses the human feedback optimization language model RLHF (Reformer Language Model with Hybrid Flow) to fine-tune ChatGPT through a combination of supervised learning and reinforcement learning, combining reinforcement learning and human feedback into NLP to minimize the output of unhelpful, distorted or biased.
When the language model is pre-trained by RLHF, it can generate different responses to a conversation while allowing people to rank the results. The key 3 steps of RLHF are as follows:
- Pre-trained a language model (LM) + fine-tuning with labeled data
- Collect data and train a reward model
- Use reinforcement learning to optimize policies for reward models
The use of pre-trained language models improves GPT-3 performance, enabling them to recognize deeper text meanings, conduct dialogue simulation through continuous supervised learning, artificial error correction, and reinforcement learning, and naturally dialogue with humans and provide feedback, and finally form a model that is closer and closer to human language.
On the other hand, in the field of image generation, although GAN can already produce high-quality images and content, it is inefficient, and on the other hand, the generated images are always unsatisfactory. The emergence of the Transformer architecture has made the field of image synthesis bid farewell to the GAN era, ushered in the combination of NLP (Natural Language Processing) and computer vision technology to produce pictures that are more in line with user needs.
In the above process of AI evolution, in addition to GAN language, Transformer architecture, and the training of language models, what is NLP (natural language process)? Why did companies suddenly start building their own big language models at the beginning of the year?
2. From Natural Language Models (NLP) to Large Language Models (LLM)
In the movie series "Rise of the Planet of the Apes", due to the widespread spread of the virus, orangutans will become intelligent due to the virus, while infected humans will lose their ability to speak and their IQ will be greatly reduced, until they are called orangutan slaves.
One of the villains, the colonel, said, "The virus will not kill us, but it will take away what makes us human, our language, our thoughts, it will turn us into beasts." "This shows the importance of language for the existence of human civilization.
As a unique human tool for expressing emotions and communicating ideas, language is a special social phenomenon, composed of phonetics, vocabulary and grammar. Speech and writing are the two basic properties that make up language, speech is the material shell of language, and writing is the written symbol system that records language.
In the long-term evolution of human beings, a common set of symbols, expressions, and processing rules are used to communicate, that is, to exchange ideas, opinions, and ideas. Among them, symbols are transmitted in the form of vision, sound and touch, modern human beings have the current high civilization, not the independent creation of contemporary human beings, inseparable from the recording and retention of their own invention achievements by predecessors, so that we can iterate and create new on the basis of predecessors.
Of course, human beings are not born with language ability, and they need to be learned to acquire it.
Language model (LM) is a model that calculates the probability of a sentence (sequence of words) or the probability of the next word in a sequence based on objective facts. The natural language process (NLP) is a language problem from the perspective of humans, whether the sentence is normal and reasonable.
When the words of a sentence always appear sequentially, each word calculates the probability by all the previous words, multiplies the probability of all these words, the larger the total probability value, the more like human language, the language model helps solve whether the statement that appears in the AI is reasonable. Language models also go through the stage from expert grammar regularity models to statistical language models and then to neural network language models.
Expert Grammar Regularity Model - (to the 80s)
In the early days of computer programming languages, models were established by inducting grammar rules for natural language, and many applications were used to improve the performance of speech recognition and machine and machine translation.
The whole process requires people to first obtain knowledge from the data, summarize the rules, write them out and give them to the machine, and then the machine will execute the set of rules to complete a specific task.
However, due to the diversity and colloquialism of natural language itself, the iteration of different grammars or buzzwords with the development of time, as well as the local words of spatial grammar of different national and regional languages, and the strong error correction ability of people themselves, resulting in a sharp expansion of grammar rules, there is no way to continue iterative use.
Traditional natural language processing systems mainly rely on manual writing of various rules based on the above components, and from the practical results, this method is time-consuming and labor-intensive, and the effect is not ideal.
The same exists in the field of image recognition. For example, in the early days of image recognition, if you want to identify a cat, you must first extract and formulate various feature rules for the cat.
Due to the variety of cat forms, and when there are obscuration, distortion, etc., it will become more difficult to manually extract features.
Statistical Language Models - (to the 00s)
A model that calculates the probability of a sentence (word sequence) or the probability of the next word in a sequence, predicts a small word of the sentence by giving the above text, if the predicted word and the next word coincide, then the probability of the above + the word appearing is greater than the probability of the above + other words, then the system judges the above + the word is more reasonable, that is, the statement with a high probability is more like human language than the statement with a low probability.
Different from the previous stage, the statistical language model has changed from the previous need to relay knowledge through people to the machine automatically learning knowledge from data, plus a large amount of corpus data.
Neural Network Language Models - (So Far)
On the basis of the statistical language model, through the superposition of the network and the layer-by-layer extraction of features, it is possible to represent similarity, grammar, semantics and other aspects in addition to lexicon.
Compared with traditional networks, neural network models can handle long-term dependencies between words, capturing the context of words and the relationships between other words in the sentence, while neural network language models can use more data to learn and iterate independently over time, while traditional language models need to be manually updated to improve their accuracy.
ChatGPT's current natural language model is the "neural network language model" stage.
In 2023, many companies based on deep learning architecture, through a large amount of text data training, so as to have a large number of parameters containing billions of parameters Large Language Model LLM (Large Language Model), used to process a variety of natural language tasks, the purpose is to allow machines to understand human commands and follow human values, which means that the use of AI is not limited to specialized fields, but general tasks, that is, linking people and machines through natural language, to meet the needs of machines to be independent, Accurately understand and complete the corresponding instructions while completing independent learning, such as text summary classification, question and answer, dialogue, etc. The role of people will gradually shift from instructors to supervisors, and even from human-computer collaboration, machine to human learning, to human-to-machine learning, and even from machines to humans.
At present, common large language models are:
GPT-3 (OpenAI): Generative Pre-trained Transformer 3 (GPT-3) is one of the most famous LLMs (Large Language Model) with 175 billion parameters pre-trained using a one-way language model. The model demonstrated remarkable performance in text generation, translation, and other tasks, generating enthusiastic responses worldwide.
BERT (Google): Bidirectional Encoder Representations from Transformers (BERT), the model is based on Google's large language model LaMDA driven, using a bidirectional method from a word to the left and right to capture context, so that the performance of various tasks is improved, suitable for understanding classes, doing understanding classes, a specific task of a karmoni, such as sentiment analysis and named entity recognition.
T5 (Google): The Text to Text Converter (T5) is an LLM that limits all NLP tasks to text-to-text problems, simplifying the process by which the model adapts to different tasks. T5 demonstrates strong performance in tasks such as summarizing, translating, and answering questions. Many large language models in China adopt the T5 model.
ERNIE 3.0 Wenxin Big Model (Baidu): The large language model ERNIE 3.0 launched by Baidu introduces large-scale knowledge graphs for the first time in tens of billions and hundreds of billions of pre-training models, and proposes a parallel pre-training method of massive unsupervised text and large-scale knowledge graphs.
Today's large-scale language models are also used to improve AI's autoresponder capabilities, intent recognition capabilities, optimize human-computer interaction experiences, and many other practical use scenarios.
3.AI Painting - Text generates images
At the Colorado State Fair Art Competition, contestant Jason Allen's work "Théâtre D'opéra Spatial" generated by the AI painting tool MidJourney won first place in the Digital Art Award. At the time, it caused great controversy, and some people even made the talk of "art dying", but then two judges said that even if they learned about it in advance, they would still award the first prize to Allen.
In the actual creation process, this painting was not completed at one time, and the entire painting process was modified and perfected thousands of times, and it took nearly 80 hours to complete.
Around the same time, a number of high-precision and efficient AI painting platforms such as Stable Diffusion and Disco diffusion began to attract attention around the world.
Many people understand that AI painting should generate images that are exactly in line with the user's mind with one click, and in the actual operation process, it is necessary to generate images by constantly entering key information, and its operation logic is very different from the artist's creation.
In fact, AI painting is similar to large language models such as ChatGPT, which requires the operator to first abstract the text understanding of the desired image, such as the composition, exposure, setting, angle, etc. of the picture need to express the abstract picture through which precise and concrete language, the entire creative process also needs multiple human intervention optimization, multiple input adjustments, and the system interacts according to the semantic understanding ability, sufficient data annotation, detail processing, and the user's Prompt prompt. In order to get a picture that meets the operator's wishes.
We see a variety of Prompt phrases for different styles on social media, so if you use MidJourney to produce images that are not satisfactory, you may want to find the corresponding style of descriptors (such as theme, medium, background, lighting, color, atmosphere, perspective, composition, art style, etc.).
Therefore, the generation of AI images is not a simple result, but more a process of expressing the operator. Since the semantic expression in natural language is larger than the phrase space, AI's understanding of semantics and humans themselves will inevitably deviate, so the essence of AI painting is collaboration and expression, so the more detailed the description of words in the process of using Midjourney, the more accurate the resulting picture will be.
When you need to modify a certain detail of the AI-generated image, you need to modify the Prompt again, and we can't know whether the AI will disassemble the Prompt you input into the modification on the corresponding picture, and there is no guarantee that the input modification is valid, so after Midjourney generates the image, it still needs to be processed by PS, AI and other tools for secondary processing, perhaps to get the final effect we want.
AI painting did not start to be researched in the past two years, nor did it take the method of text to generate images at the beginning, under the continuous technical iteration of computing power and models, various companies and related personnel continue to try to conceive and realize product landing and business models, so the speed of AI tools such as Chat GPT and Midjourney out of the circle and update iteration is not achieved overnight:
- In 2012, Ng and Jef Dean used 16,000 CPUs and 10 million cat face images from YouTube to train the largest deep learning network at the time, which took 3 days to instruct the computer to draw a cat face, and finally got a model, as well as a very blurry cat face. Officially opened the "new" research direction of deep learning models to support AI painting.
- In 2015, Google's open-source project Deep Dream, based on AI instructions, completed psychedelic surreal pictures. That same year, intelligent image recognition, using algorithms to identify and label objects in images, began to think about reverse operations to generate images with text.
- In 2016, Diffusion Models proposed to generate images using a random diffusion process.
- In January 2021, OpenAI announced DALL-E, the underlying technology is Diffusion Models, laying the foundation for the importance of diffusion models in this wave of technology development.
- In February 2022, open source community engineers such as somnai began training their own AI generator, Disco D infusion, and quite a few products based on this have since appeared.
- In March 2022, Midjouney, an AI generator built by the core development of Disco diffusion, was officially released.
- In April 2022, OpenAI's artificial intelligence online drawing application DALL· E 2 Common Test.
- In July 2022, stability.ai open-sourced stable-diffusion, the most available open source model available today, and many commercial products are based on it, such as NovelAI. On October 18, Stability.ai announced the completion of a $1.01 seed round with a valuation of $1 billion.
We can see that there are various AI drawing tools on the market, such as Midjouney, Stable Diffusion, Disco Diffusion, and so on. Compared with other AI painting tools of the same type, Midjourney can enter the Midjourney channel after registering a Discord account through the form of a community, and join the beta server to start using it.
The way to use is also very simple, users only need to enter the command prompt, about 1min to get the corresponding high-quality pictures.
Midjouney continues to iterate through the Discord community, with an influx of new free trial users, sometimes even causing servers for paid users to go down. The low user threshold, simple way to use, and quick feedback make it impossible to resist the fascination of a large number of users with Midjourney AI painting, even if it cancels the free trial and requires users to spend $30 per month.
Midjouney also relies on subscription services to achieve annual revenue of $100 million without financing.
Before the release of Stable Diffusion, the best open-source tool for AI painting was Disco Diffusion, but Disco Diffusion had problems such as slow generation speed, high cost, chaotic picture structure caused by poor image generation logic, and could not generate people and objects. On the other hand, Stable Diffusion as a free and open source tool, users can make local configuration, can ensure information security, after formulating a suitable database, AI directional learning drawing style, complete the mass production of directional style pictures.
Although Stable Diffusion has high control over Midjourney, to control Stable Diffusion requires a powerful computing environment on the server or local side for it to run.
In other words, even if you have a strong imagination, without the support of powerful natural language learning, processing power, and AI computing power, you still cannot use Stable Diffusion.
Therefore, if you are a complete novice, you can try Midjourney Experience AI drawing tools first, but if you have a large business need for work, you can choose Stable Diffusion to deploy and customize your own AI painting database.
02 Different voices about AI
At the moment when AI discussions are in full swing, many people have experienced AI-generated text or AI painting, and different voices have begun to appear in various industries about the use of AI.
1. Academic thesis script creation
At first, ChatGPT was noticed that in addition to ChatGPT's accessible conversation mode, many students began to use ChatGPT to complete papers and even get a high score of A+. Teachers had to carefully discern whether students' work was done using ChatGPT.
In the public school system in New York and Seattle, ChatGPT has been banned across the board's WiFi networks and devices. HKU also prohibits the use of chatGPT or other AI tools to take classes, do homework or take exams.
If it must be used, the prior written permission of the relevant course instructor is required, and the violation of the above temporary measures is regarded as [potential plagiarism]; If teachers suspect that students are using hatGPT, they can ask students to discuss relevant papers or works, set up additional supplementary oral examinations, add classroom exams, etc.
Students use ChatGPT to complete assignments, papers, etc., and can easily solve some problems that need to be queried and thought about.
However, the school believes that this technology is unfair to students who do not use ChatGPT, and on the other hand, the use of artificial intelligence to help complete homework and papers, which is not clearly defined at the time, is plagiarism.
For students, over-reliance on artificial intelligence to complete the course will prevent them from developing their logical reasoning, critical thinking and language skills throughout the learning process, and will also lose their independent learning process and knowledge verification process.
And 28 UK universities have explicitly banned the use of Chatgpt in essays and coursework, otherwise it will be considered academic misconduct. Many schools are already experimenting with AI-related courses or other assessment methods, such as class assignments, handwritten essays, group assignments, and oral exams.
In addition to completing a coursework paper, ChatGPT can also be used to write novels, poems, or screenwriting.
On May 2, local time, negotiations between the Screenwriters Guild of America (WGA) and Hollywood and other film and television giants broke down. 11,500 association members armed with uniformly made placards poured into the streets of New York and Los Angeles to march on strike. And what they are protesting against is not AI, but the film companies that use and train AI behind the scenes.
Since the generation of AI is not based on independent creation, but on the basis of feeding the machine related writing, story synopsis, and picture style [imitation creation], all creations are based on existing data, and for many creators, it is equivalent to using their own works to ruin their livelihood. This brings us to copyright issues, which will be discussed in a special explanation later.
2. Automated programming leads to programmers being replaced
CSDN has proposed 5 levels of automated programming:
- The first level (C1): Autocomplete based on the current line of code.
- Level 2 (C2): AI can predict the next line of code as it is written.
- Level 3 (C3): Generate code based on natural language; Programming language translation based on natural language.
- Fourth level (C4): Highly automatic programming. Projects and comments can be generated based on natural language, modules and comments can be generated based on natural language, functions and comments can be generated based on natural language, and functions, modules, and project granularity can be automatically tested and generated; Correct translation between mainstream programming languages; Generate the next line of code based on the current line of code; Code debugging (bug location and correct correction suggestions); Autocomplete based on the current line of code; Code inspection (natural language cue issues).
- Fifth level (C5): Fully automatic programming. Can generate systems and annotations based on natural language; Generate projects and annotations based on natural language; Generate modules and annotations based on natural language; Generate functions and annotations based on natural language; Function, module, project, system granularity automatic test generation; Best translation to and from all programming languages; Generate the next line of code based on the current line of code; Code debugging (bug location and automatic fixing); Autocomplete based on the current line of code; Code inspection (natural language precise prompting problems); Code automatically corrects for optimal error.
At present, ChatGPT is more like the strongest auxiliary for programmers, using ChatGPT to solve the problems of coded code completion, compilation errors, syntax errors and other problems in a few seconds, and provide information on how to use specific languages, APIs and frameworks in different languages and frameworks (that is, part of the work of C1-C3).
However, ChatGPT can only be used to quickly and accurately call up factual answers to help improve programmer productivity, and cannot be applied to tasks that require high-precision requirements such as "logical reasoning". That is, in the end, humans still need to confirm and test the correctness of the code and make changes.
The situation that junior programmers will be replaced has been replaced for more than a decade, and many enterprises have long begun to save time with low-code development, and the emergence of GPT has made this trend more obvious, and less creative development activities are easily replaced. But for small white users who have not learned programming systematically, the threshold of programming is also relatively easy.
AI has lowered the threshold distance between us and some professional occupations, he can accurately and quickly find professional information, structure information data, replace basic information collection and popularization, but still need higher professional personnel to supplement and correct information to provide real, different scenarios of solutions.
3. Design industry
Due to the explosion of AI painting tools such as Midjourney and Stable Diffusion, designers such as design, original art, and illustration feel precarious, and with the blessing of AI painting tools, it seems that everyone can create high-quality creations.
Some painters adopt a resistant and repulsive attitude towards AI painting, while some begin to try to use AI painting as a productivity and try to break through from another direction.
Major companies have a lot of obvious attitudes, and are trying to use AI painting tools such as Midjourney and Stable Diffusion to reduce costs and increase efficiency, and reduce the proportion of investment in low-end repetitive work content.
But in the actual implementation process, there will still be a variety of problems, because AI can not understand the relationship between the elements of the picture, the visual unity of the picture cannot be guaranteed, the generated content is uncontrollable, it is impossible to modify specific parts, etc., such as wrinkles on clothes, logos, specific patterns, patterns, etc., need to have special people to modify, so that AI painting does not seem to be as easy to apply to the commercial field as the legend.
This is because the underlying logic of AI painting is actually a natural language model, and AI's understanding of semantics will inevitably deviate from humans themselves.
The advancement of technology, accompanied by the disappearance of old forms of work, but also with the emergence of work content, above we roughly understand the impact of some AI on some occupations. AI and humans are also evaluating current jobs, citing the new standard of exposure to assess which forms of work will be replaced by AI.
Next, let's talk about the impact of AI on existing occupations and the emergence of new occupations.
03 AI brings a new direction of work
When the steam engine was first born, some textile workers were angry that the machine had brought human unemployment to mankind and smashed the machine.
In the beginning, people will indeed lose a lot of traditional jobs because of the birth of new technologies, but new advances will eventually create better jobs, new economic growth and creativity.
At that time, the industrial revolution created a large number of highly automated machines to replace manual labor, so that humans began to engage in a lot of mental work, and the emergence of AI also caused many work occupations to be affected.
1. Whether AIOE exposure assessment will be affected by AI tools
On March 20, the results of the Human Assessment and the GPT-4 Joint Survey showed that the professions of interpreters and translators, survey researchers, mathematicians, news analysts, journalists, and journalists were the most affected by GPT technology, followed by writers, tax officers, scribes, blockchain engineers, legal secretaries, and administrative assistants.
Using AIOE (AI Occupational Exposure), a new standard of "exposure" was introduced to measure the "exposure" of identified jobs and industries to AI advances, defined as whether the use of GPT models and their related technologies will reduce the time required for humans to perform a particular task by 50%.
- E0: No exposure.
- E1: Direct exposure, using only large voice models (such as GPT-4 chat interfaces), can reduce time by at least 50%.
- E2: Indirect exposure, using large speech models alone will not achieve the effect, but developing additional software (such as graphics generation) on top of it can reduce the time by 50%.
Level E0, which mainly includes manual labor, such as:
Ultimately, human assessors marked 100 percent of the "exposure" for 15 occupations, meaning that using GPT technology would reduce the time it takes humans to perform specific tasks in that profession by 50 percent, including writers, math pluses, bonders, financial quantitative analysts, web and digital interface designers, and others.
GPT-4 marks 86 occupations with "exposures" of 100%, including mathematicians, accountants and auditors, journalists, clinical data assistants, legal secretaries and administrative assistants, climate change policy analysts, and more, nearly six times the human assessment.
In general, if the work is highly dependent on the scientific method and critical thinking, then there will be less exposure to GPT technology, and if programming and writing skills are involved, GPT technology will be more easily exposed or affected.
From the industry level, the five industries most affected by GPT technology are securities commodity contracts and other financial investment industries, insurance industries, data processing and hosting industries, information service industries, and publishing industries.
The five least affected industries are those known for manual labor, such as agriculture and forestry, wood products manufacturing, logging, food manufacturing, and mining (excluding oil and gas). This means that "white-collar workers" with higher incomes are more likely to be affected, because this group of people is more likely to have access to and need to use ChatGPT and related tools.
2.AI New opportunities that come with it
Previously, people speculated that with the advancement of science and technology, a large number of manual labor jobs will be replaced, followed by cognitive labor, and with the arrival of AI, we have entered a new era in advance, and the least affected is physical labor.
On the other hand, AI has also spawned new professions, due to the shortcomings of AI, such as making up facts, unable to carry out logical reasoning, etc., in real work scenarios, AI can only make mistakes, it cannot be allowed to complete its work independently.
Therefore, the popularity of AI has also brought new hot high-paying occupations, such as prompt word engineers, AI trainers and other positions, which have also become popular recruitment in recent times:
- Prompt engineers – engineers responsible for developing and optimizing artificial intelligence prompt algorithms to train large models. A deep understanding of technologies such as artificial intelligence, natural language processing, machine learning, and proficiency in human language expression is required. Daily work is designed for various applications and platforms, optimizing prompt words to improve user experience and efficiency.
- AI trainer – designs and implements training programs for machine learning models so that the bot understands what the user is saying. Use a range of methods to improve the performance of machine learning models, including data cleansing, hyperparameter tuning, regularization, and more. It is also responsible for extracting, filtering, and writing meaningful structures from the available data in order to train machine learning models and solve problems that arise during training. On the other hand, it is necessary to collect the latest advances in machine learning technology and continuously improve training methods to improve the accuracy and reliability of models.
- AI code organizer - organize and correct the code generated by AI, and finally obtain a complete and logically run-through code document. (Part-time for university students)
- AI Grapher - Through the AI mapping tool "midjourney", refine keywords according to demand, and let the tool generate a picture through prompt words or keywords. Splitting keywords is at the heart of this work, and the more refined the keywords, the more you want the drawing to be. At the same time, the generated image needs to be modified in detail.
- AI narrator - uses ChatGPT-based AI tools to write web articles and stories. You need to imagine the background of a story first, and then break up the story in detail, which must be split in enough detail so that the text feedback from the AI can be more detailed and realistic.
With the explosion of AI, as a new tool, the impact of AI has naturally begun to spread to all levels and corners, and on the other hand, we also need to see the risks and limitations of AI.
04 Limitations and potential risks of AI
1. Limitations of .AI
Whether it is a literate text or a literate graph, AI is based on natural language models and a large amount of data computing power, and it still has problems in understanding and informing data. ChatGPT can collect, organize and feedback information quickly, accurately and continuously, but it also makes some obvious mistakes, including fabricating information, so in the real work environment, there must be relevant personnel to supervise its work, and cannot complete the work independently.
In terms of AI painting, the randomness of the generated images, the inability to adjust details, the cognitive dependence on input prompts and input information users, and poor directivity (inability to generate numbers or accurate to pixels) all need to be secondary processed on the generated images or used as creative concept inspiration, and cannot be directly used as commercial results.
2.AI potential risks
At the end of March, a number of prominent people, including 2018 Turing Award winners Yoshua Bengio, Musk, Steve Wozniak, Skype co-founder, Pinterest co-founder, Stability AI CEO and many others signed supporting the call for all AI labs to immediately suspend for at least 6 months and not train AI systems more powerful than GPT-4.
This pause requires AI labs and independent experts to jointly develop and implement a shared set of advanced AI design and development security protocols that should be rigorously reviewed and overseen by independent external experts.
At a time when AI is advancing so rapidly, relevant supervision and auditing methods are still slow to keep up, which means that no one can guarantee the safety of AI tools and the process of using AI tools. The appeal letter raises questions:
- Should we flood our information channels with machines, spreading propaganda and lies?
- Should we automate everything, including those that are satisfying?
- Should we develop non-human minds that may eventually surpass and replace us?
- Should we risk losing control of civilization?
It is important to note that this advocacy letter does not mean to suspend the development of AI, but rather hopes to focus research and development on improving the accuracy, safety, explainability, transparency, stability, consistency, trustworthiness, and loyalty of existing robust, advanced systems. In the letter, it argues that AI developers need to work with the government, at a minimum:
- new agencies dedicated to AI capabilities;
- Oversee and track high-performance AI systems and large pools of computing power;
- Used to help distinguish between real and synthetic provenance and watermarking systems, and to track model leaks;
- Strong audit and certification ecosystem;
- liability for injuries caused by artificial intelligence;
- Strong public funding for technical AI safety research;
- Well-resourced institutions to deal with the enormous economic and political disruption that AI will cause (especially to democracy).
Attached is a link to the original letter, and students who wish to learn more can check it out for themselves, "Pause Giant AI Experiments: An Open Letter".
On the other hand, the completion and improvement of AIGC models rely on a large amount of data training, and the data used for training often contains copyrighted content (such as pictures in the image copyright holder's gallery, works of well-known artists, etc.), on the other hand, whether the artist has the willingness to be imitated by AI.
Therefore, there has been a lot of controversy over the copyright of AI-generated products for commercial use. Whether the copyright of the image produced by AI belongs to the user, the platform, or needs to be owned after copyright registration or released under the CC0 (Creative Commons license) license, different platforms or users have their own set of ideas.
It was not until March 16, 2023, that the U.S. Copyright Office (USCO) published Part 202 of the U.S. Regulations stating that works automatically generated by AI are not protected by copyright law.
According to the USCO, the author's creation of the picture work through Photoshop is protected, and there is human participation in the entire process from the initial idea to the completion of the creation. The works automatically generated by AI tools are automatically completed by robots, and the training data needs to be based on works created by humans, so they are not protected by copyright law.
At present, there is no clear law in China on the copyright of AI products.
In other words, no matter what kind of AI tool is used to create scripts, novels, paintings, music, etc., no one can own the copyright, and anyone can use it.
05 Epilogue
You haven't tried any AI tools yet, you can contact ChatGPT, Midjourney, or other AI tools first, but currently ChatGPT needs to go to the OpenAI official website and follow the prompts to enter the ChatGPT page.
At present, domestic mobile phone numbers are not supported, and third-party platforms are required to assist in code connection, if you want to experience quickly, you can try other platforms with low thresholds such as Notion or Writesonic. The same situation with Midjourney currently has a large number of tutorials on how to register and use and keyword explanations, which are not explained too much here.
The waves of innovation in history have never eliminated humanity, nor have they led to mass unemployment, but have led to an unprecedented increase in the demand for labor.
For example, although machines have partially replaced manual farming, they have also spawned and connected upstream and downstream industries such as production and maintenance around agricultural machinery, each of which contains a large number of jobs. After the popularization of computers, the digital economy and platform economy are still expanding the boundaries of our life and work today.
The exploration of AI is never a sudden outbreak, technology and new brings more possibilities, perhaps more often we need to remain optimistic and sensitive to technology, and find our own direction in the constant changes in the situation.
Source of resources:
- "PenAI CEO Latest Interview, 30,000 Words Details Technology, Competition, Fear, and the Future of Humanity and AI"
- The LLMs Big Language Model is an abstraction of the physical world
- What is the LLM Big Language Model? Large Language Model, from quantitative change to qualitative change"
- "Transformer Popular Understanding"
- What is GAN (Generative Adversarial Network)? 》
- Don't let GPT-4 evolve again! Musk took the lead in signing a letter of thousands of people, urgently calling on AI labs to immediately suspend research
- When AI impacts automation programming, who will benefit? 》
- "Your AI Violated My Copyright": A Brief Discussion on the Copyright Protection Issues Behind AIGC
Author: No such person; Public number: No. 9 study room; Zhihu column: No. 9 study room.
This article was originally published by @9号自习室 on Everyone is a Product Manager. Reproduction without permission is prohibited
The title image is from Unsplash and is based on the CC0 protocol