Generative AI : The New Disruptive Paradigm
Camille Beyrouthy

2023-02-28 16:36:26+00:00

A revolution is occurring within the media industry, spurred by ChatGPT and large language models, with Stable Diffusion being a key player. A new emerging pattern divides Generative AI players into a foundational layer group and another application layer one. This delimits the playing field going forward, as was the case in the past with “operating systems” being developed as environments containing programs and applications. Furthermore, in Section 2, Open-source versus Closed-source Gen-AI is discussed in detail and the advantages of the former over the latter are made evident, including the cost effectiveness of Open-source, resulting in granting enterprises the ability to start small and scale up when possible. Also, since Open-source code is freely available, it allows for public collaboration to address issues and provide support. More importantly, this report estimates the Total Addressable Market (TAM) within the media industry for Stable Diffusion (total potential revenue that can be captured by Stable Diffusion due to the economic disruption caused by GenAI). Spending in the media sector rests on three major pillars: Customer Spending, advertising, and internet/online access, with revenues split almost evenly currently and each pillar generating around 30% of revenues. The total size of the media market is around USD 2.5 trillion. It is split into multiple segments, with films and gaming being the dominant ones

 We have recorded the impact of Stable Diffusion in the films sub-segment through its recent ability to capture frame-by frame animation, and the impact on gaming through the new excellent “ideation” capability of Generative AI, which is likely to radically transform the production of games, expediting the process and massively reducing costs. Furthermore, using a top-down approach to quantify the TAM, the media and entertainment sector has been divided into subsegments and further niches within the subsegments to recognize, the total potential revenue that can be captured by Stable Diffusion. We found that while the TAM stands at USD 2.5 Trillion in 2023, the potential for economic disruption by GenAI amounts to nearly USD 1.07 trillion at present and is expected to grow to USD 1.4 trillion in 2025 (see table on page 17). Finally, through our review of the major players in Generative AI, we detail the Venture Capital investments in the field as a barometer of the future potential of the field. We identified fund flows per segment, category, funding rounds and geography (see Section 9). As per our knowledge and to date, this is the most exhaustive look into the purely Generative AI ecosystem

Generative AI, An intro

The core of Generative AI (Gen-AI) is the Large Language and Image models (LLM), also known as foundation models. These models grant users the ability to do the following: Generative Tech is now being considered, worldwide, as the next big thing in software. It’s a new level of human-machine partnership. It turns deep learning engines into collaborators to generate new content and ideas almost as a human would. Some have called it “Foundational AI”. AI models are the enabling base layers of the stack, with thousands of applications being built upon these. Generate Content Those models can automatically generate content, such as articles, blog posts, images, videos, etc., acting as an invaluable time-saving tool for businesses who are pressured to create content under tight deadlines. 01 Content Quality Improvement The Gen-AI-generated content can be of higher quality than that created by humans since AI models are able to synthesize large amounts of data and can identify patterns that are difficult for humans to recognize. 02 Personalize Content to One’s Specific Needs AI models can generate content based on personal preferences and individual choices. This can help businesses create client-tailored content for a target audience, hence being more useful. 03 The Generative Tech sector is developing at a very rapid pace, as reflected in real revenues and high valuations, even though a term for it had not been coined till as recently as a month ago. Eighteen months after its launch, Jasper is reported to have recorded nearly $100M in revenue and reached a $1.5B valuation. Open AI, which powers GPT-3 and other AI models, has been raising capital at valuations of tens of billions. Anthropic, another large model builder, has raised sizable amounts. The recent availability of open-source alternatives to proprietary Gen AI models, proved to be a tipping point in the last six months. In short, EleutherAI, GPT-NeoX-20B, launched in Feb 2022, is the open-source alternative to OpenAI’s GPT-3 for text generation. Stability AI’s Stable Diffusion, launched in August 2022, is the open-source alternative to OpenAI’s DALL-E 2 for images and videos. Both have been game changers on price, quality, and ease of access. The cost to generate images has dropped 100X ever since. The ability to generate output from these models through web and mobile has become “about 10 times easier” in the last six months. This will be discussed in detail in the next few pages. of outbound marketing messages from large organizations will be synthetically generated, a significant increase over the less than 2% today. of the film generated by AI (from text and video). a major blockbuster film will be released with By 2025 By 2030 30% 90% 1 https://www.gartner.com/en/articles/beyond-chatgpt-the-future- of-generative-ai-for-enterprises?source=BLD-200123&utm_medium=social&ut m_source=bambu&utm_campaign=SM_GB_YOY_GTR_SOC_BU1_SM-BA-SWG

Foundation models
Foundation models (like GPT-3 and Stable Diffusion) are extremely large models trained on broad datasets that can be adapted to a wide range of downstream tasks. Furthermore, a Generative AI model is specifically a foundation model, where the “training” involves modeling the “probability distribution” of the underlying data, for example, predicting the probability of a character being the next one in a given text sequence. We will discuss this in detail in the next few sections. The low cost and ease-of-use of these models is helping to accelerate the development of AI apps as more engineers push themselves into the field of AI. Today, foundation models are frequently adapted to build generative applications with the “wow” capability. But they can also be applied to more traditional ML use cases such as classification and entity extraction, and more importantly, they minimize (but not completely obviate) the need for startups to gather proprietary training data, label it, architect complex data transformations, tune hyperparameters, and select the right model.
Generative Applications
These are companies utilizing generative AI for its namesake purpose: the creation of net new output in various media types. This is by far the most prolific category and thus, comprises the majority of companies on our index. We are seeing startups here that are both building directly on top of existing foundation models, as well as those that have chosen the route of building their own models from scratch, particularly in domains where foundation models don’t exist (e.g., speech). The bottom layer is an AI model, which can generate novel output based on inputs that are unique to the user, such as OpenAI’s DALL-E or Stability’s Stable diffusion model. The Large Language Models (LLMs) were first developed at Google in 2016, and were used as the backbone for their translation engine, trying to preserve meaning and context. Since then, large language and text-to-image models have proliferated internally, at major tech firms, including: Google (BERT and LaMDA); Facebook (OPT175B, Blende-Bot) and OpenAI. Then in late 2021 and 2022, the following players emerged in the foundational space: Stable Diffusion; MidJourney; and Crayion (Dall-E-Mini). These models have largely been confined to major tech companies because training them requires prohibitively massive amounts of data and computing power. GPT-3, for example, is trained on a 40 terabytes training set and employs 175 billion Neural Network coefficients to generate predictions. Hence, a training round for GPT-3 costs upward of $10 million, using thousands of NVIDIA GPUs - the norm in the AI computing industry. Starting from scratch in the AI model layer is very hard. Most GenAI companies don’t possess the data center capabilities or the large computing budgets required to design their own models despite the public availability of code. Nevertheless, many application layer companies are trying to establish a foothold in the foundation layer. These include: Character AI (ex-Google employees); CohereAI (ex-Google employees); and Anthropic. The three companies above are in the nascent stage, while the only two companies far ahead of others in the field are OpenAI and StabilityAI. The funding pattern of most startups has been such that they have followed a revenue model adapted to the application layer, which is mostly B2C. There are some aspiring ones that have tried to move downwards in the model stack, but that is turning out to be challenging for them. High costs of training, associated with developing models using cloud computing resources, is a significant challenge, with one round of training requiring at least thousands of GPUs and costing millions of dollars. Additionally, algorithmically, Open AI and Stability have been very successful in developing the technology and have allocated funds and resources to this that are hard for new companies to match.

Closed Source - OPENAI
Founded by Elon Musk and the Y Combinator president Sam Altman, OpenAI rose to quick international fame when they launched ChatGPT in November 2022. Within a week, the application saw a spike in usage of over a million users. Being able to code and interact in a way that mimics human intelligence, ChatGPT has surpassed previous standards of AI capabilities and has introduced a new chapter in AI technologies and machine learning systems. OpenAI rushed to reveal their model, but similar and probably as powerful models already existed. Lemoine, a software engineer at Google, who had been working on the development of LaMDA, shared his interactions with the program, in a Washington Post article, causing a stir. Lemoine recounted many dialogues he had with LaMDA in which the two discussed topics that ranged from technical to philosophical issues. These led him to ask if the software program is sentient. Lemoine was eventually fired. Launched in 2015, before the introduction of any Generative AI concept, OpenAI witnessed the collaboration between, Musk and Altman on one side and other players in Silicon Valley the likes of Peter Thiel and LinkedIn founder Reid Hoffman who pledged close to a billion Dollars for OpenAI that year. Two major projects formed the cornerstone of Open AI. These were:

Report page (PDF, 6Mb) here