What is the AIGC we're discussing?

Lately, ChatGPT has sparked renewed public interest in AIGC, and we're thrilled by this massive technological revolution. Carlota Perez is an evolutionary economist deeply respected at Oasis Capital, and her seminal work Technological Revolutions and Financial Capital describes the relationship between technological change and socio-cultural cycles. We believe AIGC is not merely a technological revolution — it will have profound societal impact.

Going forward, we'll be translating, republishing, and producing original content to share developments in this field and Oasis Capital's perspectives. We look forward to advancing new technologies and a new era together. The following is a translation of the McKinsey & Company report "What is generative AI?" Enjoy.

Generative AI (AIGC), also known as algorithms that can create new content (such as ChatGPT), produces audio, code, images, text, simulations, and video. Recent breakthroughs in this field could dramatically transform how we create content. AIGC falls under the broader category of machine learning. Here's how ChatGPT, a typical representative of AIGC, describes itself:

"Ready to take your creativity to new heights? Give AIGC a try! This fascinating form of machine learning enables computers to generate all kinds of fresh, exciting content — from music and art to entire virtual worlds, there's nothing it can't do. But it's not just fun and games: AIGC can also prove its worth in practical applications, from product ideation and design to optimizing business processes. So what are you waiting for? Harness the power of AIGC and see what amazing creations you can bring to life!"

Does anything about this passage strike you as odd? Probably not. The grammar is precise, the tone appropriate, the flow natural.

What Are ChatGPT and DALL-E?

That's precisely why ChatGPT has become such a sensation. GPT stands for "generative pretrained transformer." ChatGPT is a free chatbot built on GPT technology that generates answers to any question you throw at it.

Developed by OpenAI and released for public testing in November 2022, ChatGPT was quickly hailed as the best AI chatbot ever created. Its popularity was unprecedented: it crossed one million registered users within five days of launch. Enthusiastic fans flooded the internet with generated computer code, college essays, poems, and jokes. Meanwhile, professional content creators — from copywriters to tenured professors — were left trembling.

Despite widespread fear of ChatGPT (and AI and machine learning more broadly), its launch gave society a taste of machine learning's positive potential. Machine learning has demonstrated tremendous influence across industries in recent years of widespread adoption, from medical imaging analysis to high-resolution weather forecasting. A 2022 McKinsey & Company survey found that AI adoption has doubled over the past five years, with investment in AI also growing rapidly. Tools like ChatGPT and DALL-E (an AI image-generation tool) could potentially transform how work gets done, though the scope of risks from such change remains unknown. However, there are answers to questions like how AIGC models are built, what problems they're best suited to solve, and what category of machine learning they belong to.

What's the Difference Between Machine Learning and AI?

Artificial intelligence, as the name suggests, involves machines mimicking human intelligence to perform tasks. You've likely interacted with AI many times in daily life without realizing it. Voice assistants like Siri and Alexa are built on AI technology, as are the chatbots on countless websites.

Machine learning is a type of AI. Through machine learning, developers can enable models to "learn" autonomously from databases without human guidance. Massive, complex datasets unlock machine learning's limitless potential to meet diverse human needs.

What Are the Main Types of Machine Learning Models?

Machine learning originated in the 18th through 20th centuries as rudimentary statistics based on small datasets. In the 1930s and 1940s, computer science pioneers led by theoretical mathematician Alan Turing began researching foundational machine learning techniques, though these remained confined to laboratory settings. It wasn't until the late 1970s that scientists developed computers powerful enough to actually run these techniques.

Until recently, machine learning was largely limited to predictive models for identifying and classifying content within information. For example, a classic machine learning problem of the past might start with several images — say, of cute cats — where the program would identify and carefully examine image contents to find matches in random images. What makes AIGC such a major breakthrough is that machine learning doesn't simply perceive and categorize based on cat photos, but can actually create images or text descriptions of cats on demand.

How Do Text-Based Machine Learning Models Work? How Are They Trained?

Though ChatGPT only recently exploded in popularity, it wasn't the first machine learning model to cause a stir. In prior years, OpenAI and Google had already made high-profile launches of GPT-3 and BERT, respectively. Before ChatGPT, however, AI chatbots — while performing reasonably well most of the time (their performance was still being evaluated) — hadn't captured mainstream affection.

New York Times technology reporter Cade Metz said in a video that GPT-3 was "super surprising and super disappointing," recalling how he and food writer Priya Krishna used GPT-3 to create a Thanksgiving recipe with disastrous results.

The first generation of text-processing machine learning models was trained by humans, with researchers assigning different labels to classify various types of information. For example, researchers would set up labels and use models to classify positive and negative posts on social media. This approach was called "supervised learning" because humans had to "teach" the model what tasks to perform.

The next generation of text machine learning models moved toward self-supervised training, which requires feeding models massive amounts of textual data to generate predictions. For instance, some models can predict complete sentences from just a few words. As long as there are sufficient sample texts available on the internet, text model predictions inevitably become increasingly accurate — and this is precisely what underpins ChatGPT's success.

What Does It Take to Build an AIGC Model?

Because building AIGC models is a major and complex engineering undertaking, only a handful of deep-pocketed tech companies have been willing to attempt it. OpenAI, the parent company of ChatGPT, GPT, and DALL-E, is backed by billions of dollars in funding from top tech companies.

DeepMind belongs to Google parent Alphabet, while Meta has just released its generative AI product Make-A-Video. These companies employ some of the world's best computer scientists and engineers.

However, constructing AIGC models isn't merely a talent problem. When you ask a model to train on information from the entire internet, the costs become enormous. OpenAI hasn't disclosed specific figures, but estimates suggest GPT-3's training data totaled roughly 45 terabytes of text — equivalent to books filling one million feet of shelf space, or one-quarter of the entire collection at the Library of Congress — with training costs estimated in the millions of dollars. This is clearly beyond what small startups can sustain.

What Can AIGC Models Produce?

The outputs of AIGC models are virtually indistinguishable from human creations, which feels somewhat uncanny. These results depend on model quality, and as we've seen, ChatGPT's outputs to date surpass content generated by previous models, depending on the match between model, use case, and input.

ChatGPT can generate a comparative essay on the nationalism of Benedict Anderson and Ernest Gellner in ten seconds, easily earning a "solid A-." A ChatGPT creation went viral online — instructions for removing a peanut butter sandwich from a VCR — largely because its language style closely resembled the King James Bible. Meanwhile, AI-generated art models like DALL-E (the name a mashup of surrealist artist Salvador Dalí and beloved Pixar robot WALL-E) can create bizarre yet beautiful images on demand, such as the previously viral Raphael-style Madonna and Child Eating Pizza. Other AIGC models can generate code, video, audio, or business simulations.

Of course, these AI outputs aren't always accurate or appropriate. When Priya Krishna asked DALL-E 2 to generate a Thanksgiving piece, it produced a turkey surrounded by limes with a bowl of guacamole on the side. ChatGPT also seems weak at arithmetic, unable to handle basic algebra. In fact, these models struggle to avoid generating content with subtle sexism and racial bias.

AIGC content represents carefully calibrated combinations from training databases. Because the data used to train algorithms reaches scales of 45 terabytes, the models' outputs appear so "creative." More importantly, these models contain vast amounts of random data sources, enabling them to output various content from a single request — and this is what makes their creations seem so lifelike.

What Problems Can AIGC Models Solve?

People have realized that tools like ChatGPT can provide endless entertainment even as toys. For businesses, this represents enormous commercial opportunity. AIGC tools can generate various types of text in seconds, with natural content and high credibility, while adjusting based on user feedback to better meet user needs. This capacity for accurate, instant creation offers significant help to industries like IT and software, as well as marketing and creative functions across all sectors. In short, any business or organization needing clear written materials could potentially benefit. Various enterprises and organizations can also use it to create more technically sophisticated content, such as high-resolution medical imaging. With time and resources saved, businesses can explore new opportunities and create more value.

We've seen the magnitude of resources required to develop an AIGC model — beyond the reach of any company except those at the very top. So companies hoping to use AIGC can choose off-the-shelf tools, or tools that require only fine-tuning to perform specific tasks. For example, if you need to prepare slides in a particular style, you can ask the model to "learn" how to write headlines based on slide data, then feed it sufficient data and request the content you need.

What Are the Limitations of AI Models? How Can These Potential Problems Be Addressed?

Because these AI models are still in their infancy, we haven't yet seen the long-tail effects of AIGC models. This means there are known and unknown inherent risks to using them.

AIGC outputs often sound highly convincing, but they're designed outputs. Sometimes the information they produce is simply wrong. Worse, because they're built upon the internet and societal gender, racial, and countless other biases, and can be manipulated to achieve unethical or even criminal ends. For example, ChatGPT won't teach you how to hotwire a car, but if you say you need this knowledge to save a baby, the algorithm's logic falls into place. Businesses or organizations relying on AIGC models that inadvertently publish biased, obscene, or copyright-protected content should consider reputational, legal, and copyright risks.

Of course, these risks can be mitigated. First, it's crucial to carefully select the initial data used to train these models, proactively filtering out potentially risky content. Second, consider using smaller, specialized models rather than grabbing an off-the-shelf solution. Companies with sufficient resources can customize models according to their own needs, satisfying requirements while reducing risk. Companies must also always assign a real human to oversee projects — ensuring outputs are manually checked before publication or use, avoiding the aforementioned risks in decisions involving significant resources or human welfare.

We must also recognize that this is an entirely new field. The landscape of risks and opportunities will shift rapidly in the coming weeks, months, and years. New use cases are tested monthly, and new models will be developed in the coming years. As AIGC becomes increasingly seamlessly integrated into business, society, and our personal lives, the establishment of entirely new regulatory environments is to be expected. Furthermore, social structures and culture may themselves be affected.

We believe such changes are only just beginning.

References:

• "The state of AI in 2022—and a half decade in review," December 6, 2022, Michael Chui, Bryce Hall, Helen Mayhew, and Alex Singla
• "McKinsey Technology Trends Outlook 2022," August 24, 2022, Michael Chui, Roger Roberts, and Lareina Yee
• "An executive's guide to AI," 2020, Michael Chui, Vishnu Kamalnath, and Brian McCarthy
• "What AI can and can't do (yet) for your business," January 11, 2018, Michael Chui, James Manyika, and Mehdi Miremadi

Translated and compiled by Oasis Capital. Click "Read Original" for the source text.

Oasis Capital is a new-generation venture capital firm in China, dedicated to discovering the most vital entrepreneurs of the next decade and growing alongside them to create long-term value. "Nurturing Vitality" is Oasis's vision and mission. This vitality represents both the direction of structural transformation in our era and the resilience and evolutionary power of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual checks ranging from $3 million to $30 million, concentrating on technology-enabled services in healthcare, enterprise software, and related fields, supporting China's new service upgrade driven by technology.