Mindset Ventures’ overview of generative AI
2024
Artificial intelligence on steroids
Human beings have always used technology, even rudimentary, to push the boundaries of imagination. In 1965, Gordon Moore proposed a theory stating that the number of transistors in an integrated circuit doubles every two years or so (and, intuitively, so would do other technology enhancements), which has proven quite accurate since then. Although we shouldn’t be surprised to experience what an exponential technological evolution feels like as we have been living in one for so long, the recent tech-related advancements have been unfolding in front of our eyes at such a fast pace that it becomes almost humanly impossible not to be impressed, if not frightened. The speed of technological progress has reached a point where even the term “artificial intelligence” feels too limited for certain applications. Today, many see generative AI as the true future of technology, offering capabilities far beyond what we previously imagined.
The truth is that generative AI refers merely to one of the uncountable ways to use AI. Simply and figuratively put, generative AI is AI on steroids, capable of creating content such as audio, video, image, text, and what we refer to as “deep fake,” which would be otherwise impossible to achieve without the “generative” part of AI.
Despite the apparent common sense that generative AI is the greatest technological creation of recent years, its origins can be traced back over six decades. The first few signs of what may resemble generative AI came up in the 60s through the creation of the “Eliza” chatbot by Joseph Weizenbaum, an MIT professor. Although the chatbot could simulate conversations successfully, it faced a limited vocabulary. It could not reach a fraction of the capacity of currently widely used chatbots.
Even so, given their broad utility, the significant leap in generative AI occurred only a few years ago with the launch of ChatGPT and Dall-E, two currently popular platforms. While generative AI has been dividing opinions around the globe, it seems it has changed the way we do our daily activities, especially professional tasks.
Rocket science made simple
As generative AI is entirely structured on artificial intelligence principles, understanding how it works requires at least some basic technical knowledge of artificial intelligence. AI consists of algorithms (or software) capable of identifying certain patterns in collected data, using them to make decisions or predictions, and solving problems whose solution can be achieved by correctly processing data, somehow resembling the decision process followed by humans on certain occasions. In other words, AI is sometimes very complex, but it’s still just math. Although AI has been extremely useful for tasks that require intense data analysis, it has always faced an important limitation: the ability to create.
That’s where generative AI comes in. It works like AI but can create new information based on previously collected data. If artificial intelligence is math, then maybe generative AI is rocket science. The first few trials of Gen AI use cases for solutions that could potentially solve entire business problems have already emerged, such as software capable of designing marketing campaigns step by step or programs to write entire corporate letters from scratch.
Generative AI involves a complex and interconnected infrastructure, which can explain the vast market it has become. This technology offers diverse entry points, creating countless business opportunities due to the various layers that must seamlessly operate for Gen AI to function. Today, the infrastructure is categorized into four distinct groups, each comprised of huge markets with potential to explore. Some were specifically established focusing on AI, such as OpenAI and Anthropic, while other pre-existing companies, such as AWS and Nvidia, decided to explore the Generative AI infrastructure. This convergence of entities reflects the dynamic and evolving landscape of Generative AI, creating collaboration and innovation across the industry.
Infrastructure Layer
The foundational models require a vast amount of processing to be trained by an enormous quantity of data and to be operated as well. Cloud-service and Hardware — GPU chips (Graphics Processing Unit, unlike the CPU, can run multiple tasks in parallel much more efficiently by breaking them down into smaller components) — providers are fundamental for this segment.
Data infrastructure Layer
In this layer, the quality of a model’s outcomes heavily relies on the data used for training. Therefore, solutions such as a Data Lake and clean data are crucial in this layer. They play a crucial role in ensuring the quality of trained models. This layer encompasses data collection, processing, storage, management, and organization, all of which are vital aspects in preparing the data used for training.
Foundational Layer
Companies in this layer train Generative AI models from the ground up. Currently, the market is categorized into two segments: the first one, closed source, comprises the largest models such as Open AI, Anthropic, Google, and many other AI models. These models are private and typically require payment for access. They come pre-trained with their own code and datasets, which are not publicly shared or modifiable. The second, open source, is publicly accessible, allowing anyone to use, modify, and train them with their own datasets. The code and the underlying model architecture are openly available, promoting flexibility and transparency. This openness encourages collaboration among developers and researchers, enabling them to contribute improvements and innovations.
Application Layer
The application layer is the practical representation and usage of existent generative models and data to perform specific tasks or applications. It includes end-to-end apps or third-party APIs that integrate generative AI models into everyday use cases across various domains, enhancing productivity, creativity, and ensuring smoother workflows.
The revolution — Use cases
Megacycles in technology are high-tech disruptions that happen occasionally, episodes of extremely fast evolution taking place in short periods, capable of revolutionizing an entire industry or generating a structural impact on the economy. Although we have barely started to explore the unimaginable number of possibilities that generative AI opens to us, the short time frame we have had contact with it is already enough for us to confidently assume we are living the first years of an important technological megacycle. If ChatGPT alone has already leveled up AI technology into new eras, who knows how many more revolutions lie ahead of us in the next few years?
ChatGPT is estimated to have over 200 million monthly active users. The graph below shows the days required for certain platforms to reach 1 million users to provide a sense of proportion. Until July 2023, ChatGPT had been the fastest-growing application in history until Threads took its place.
To dive into the specifics of this technology and understand how ChatGPT works, it is essential to uncover its underlying architecture. A Generative Pre-Trained Transformer (GPT) represents a sophisticated neural network architecture, a series of interconnected layers transforming input into content. For instance, when one asks a question to ChatGPT, the answer is the outcome of various layers collaboratively working together, identifying patterns to provide the most accurate response. This architecture trains Large Language Models (LLMs) and deep learning algorithms utilizing advanced techniques and large datasets. LLMs are responsible for understanding, summarizing, and generating new content.
One of the most significant recent events in the AI ecosystem was DevDay, OpenAI’s first developer conference. In this event, Sam Altman, its CEO, announced the launch of ChatGPT4, marking a substantial advancement in the market. A key distinction in this upgraded version lies in the number of neural network parameters (or layers) used to construct it. The quantity of parameters directly influences the model’s capacity to learn complex patterns and representations from data. ChatGPT3, with its 175 billion parameters, is adept at processing various tasks such as language translation, summarization, and question-answering. In contrast, ChatGPT4 boasts a staggering 1.76 trillion parameters, an unimaginable capability to generate content and perform vastly more complex tasks than its previous version.
We’ve been massively automating work for nearly 200 years, naturally creating new jobs while others disappear. Generative AI comes as the new wave of automation given its flexibility and usefulness of being introduced into almost any business to enhance processes while still being able to boost creative tasks for companies. Use cases related to the implementation of this technology are countless, ranging from finance to marketing, the latter being one of the main targets for generative AI as marketing campaigns such as videos, photos, and even strategies are already easily designed by generative AI engines based on simple descriptions given by humans. Customer support is also on the list of potential activities to be completely disrupted by generative AI by creating human-like avatars who could occasionally replace real people in this job. Even education can make disruptive, good use of this technology for content creation and answer search. The discussion of whether AI will completely replace most of the currently existing jobs is debatable. In the short term, it seems the job market is starting to require some sort of AI knowledge from candidates for an increasingly larger scope of positions. According to BCG research, 80% of the US workforce uses GenAI, with 68% of enterprise employees experimenting with GenAI at least once and 48% of workers using GenAI tools regularly. Today, hiring an AI engineer in Silicon Valley with a PhD and five or so years of experience in a Big Tech often requires a starting salary of one million dollars per year.
As a matter of curiosity, a study conducted by McKinsey indicates the potential that generative AI must generate economic gains in multiple functions, as shown in the graph below. Six of these functions alone (customer operations, marketing, sales, software engineering for IT, software engineering for produc development and R&D) account for approximately 75% of the annual value that generative AI can create. Additionally, financial sectors are increasingly experimenting with Generative AI, with 54% of CFOs reporting their organizations are already exploring its use.
As demonstrated in the study, customer success is one of the main activities that can be significantly enhanced through Generative AI. Hyro, one of our portfolio companies, for example, develops conversational AI assistants that enable customers to interact intuitively with information through voice command or by chat, generating actionable insights for the company. Brex, another portfolio company of ours, has recently launched a GenAI product that automates payables, guaranteeing accuracy. Findem helps companies automate their hiring process, while ClickUp helps manage teams, track employee development, and streamline the entire onboarding process. However, this may change soon as companies aim to develop their expert-level use cases of Generative AI. They don’t want to limit this technology to customer service or small problems; instead, they want to transform their internal operations with Generative AI and solve their hardest problems. To achieve this, they need a secure and reliable internal platform, that’s where Articul8 comes in.
Articul8, our third investment in Fund IV, is a Generative AI company that addresses the issue of security and data leakage in large enterprises when implementing Generative AI. The startup helps organizations securely create their own GenAI platforms, ensuring all data remains within the company and preventing data leakage.
Interestingly, Generative AI can improve productivity and, in some instances, even creativity. A study conducted by some of the most important universities in the US, involving over 3,000 companies worldwide, showed that individual creativity increased when Generative AI was used. However, collective creativity declined as ideas became more similar and less diverse. When GenAI tools are provided to experts, creativity typically increases as they know how to use them effectively and understand their limitations. Conversely, giving these tools to less experienced individuals can be detrimental, as they may not know what to trust and may rely too heavily on the tool itself, leading to quick but flawed outputs. This result highlights the importance of having experts to implement this new technology in companies. While Generative AI can be highly beneficial, improper use can harm a company’s projects.
Unreal humans
Deepfakes are one of the generative AI applications attracting the most attention, given its useful tools for several types of business and the controversies behind its improper use. This variant creates almost perfect human-like avatars animated with human expressions, more recently reaching such a level of realism that it became nearly impossible to distinguish between an authentic human image or animation and a deep fake avatar.
Despite the benefits of this type of technology, deepfakes have also been raising several ethical concerns. Empowering the general public with a tool of such capacity and ease of use without necessarily the consent of those whose identity would be used for any purpose is worrisome. Still, it has already naturally propelled the creation of the first tools to distinguish deep fakes from real pictures or footage. Last year, a picture generated by AI of Donald Trump being arrested was released, and despite how quickly its authenticity was contested, the episode generated enormous speculation. Despite the jaw-dropping realism and unimaginably wide scope of uses that deepfake may bring, it may also behave as a huge propeller of misinformation, something we have already been struggling with for a while, long before deepfakes were possible. Although we invested in one of the pioneers of this market before it pivoted its cybersecurity core to one of the most accurate and refined generative AI engines, our experience in this segment has been quite positive, to say the least D-ID started as a solution provider to make pictures unrecognizable for computers without visibly changing them through human eyes. More recently, the company experienced tremendous success after creating one of the first and more sophisticated deep fake tools, which had nothing to do with its initial business but became its core, given the success of what was initially a mere experimental initiative. D-ID allows users to transform photos into videos, create human-like avatars with the same aspects as real people through Generative AI, and even embed pre-recorded voices from anyone into the animations, a solution that brings immense advantages to multiple media-heavy industries, including education, cinema, and marketing.
Omnipresence
In the past decades, we have been increasingly swarmed by digital technology, which has often been replacing physical objects or solutions to what one would consider tangible problems. Through credit card machines we no longer need to wait for the attendant to get back to us with change, and kids have been gradually losing interest in toys as tablets and videogames provide them with a much more intense playing experience. But, as innovation becomes relevant to society, it also attracts the attention of regulators and governments.
Innovation always surge before regulations gets updated, as it would be impossible to adapt regulation to technological progress that has not yet taken place. However, the speed at which we’ve been innovating recently has been so intense that regulators have faced substantial issues in maintaining the same pace to ensure the proper limits are established for the latest technological progress. Back then, authorities didn’t take long to regulate credit cards or digital banking. Not long ago, social media started presenting some difficult obstacles to regulators, who took some time to understand what limits were reasonable to impose on Facebook, YouTube, and other similar companies. Nowadays, however, authorities have barely figured out how to deal with crypto assets transactions when generative AI suddenly boomed. Our technology has been evolving at such a rapid pace that not too long ago, many experts, including Elon Musk signed a letter to temporarily cease the development of AI to prevent that these technological advancements get out of control and potentially become harmful.
Amid this scenario, the European Union seems to lead the regions that are most rapidly evolving in properly regulating digital innovations, followed by the US. As establishing rules naturally sets limits to innovation and consequently decreases how quickly it evolves, the US has recently taken a step back in this regard, acting more cautiously to impose limits that would contradict its historic constant incitement to innovate.
However, regardless of which limits were already imposed by the European Union, the US, or any other region, one aspect is worthconsidering: unlikealmost everything humanity has already experienced, technology, especially digital , is nearly omnipresent, which means the set of rules adopted by one region may not be enough to efficiently set the limits to a certain technology in case other regions do not adopt it as well. As an example, while Switzerland is currently far ahead in defining rules for transactions of crypto assets, this segment is still arguably poorly regulated across the globe, given how far back other regions are in this respect.
Not all roses
Despite the numerous benefits and positive impact of generative AI in various areas, it still faces challenges that require attention and must continue to improve to reachits full potential.
Together with the regulatory concerns about generative AI, ethical dilemmas are natural concerns given this technology’s capacity to distort reality and potentially generate significant privacy concerns. As previously discussed, deep fakes can play a substantial role in disseminating misinformation and possibly incriminating innocents.
This creative economy stands out as one of the most positively and negatively impacted sectors. While Gen AI can assist creators in more efficient content creation, it can also pose harm to the industry. A notable example was a song released by an unknown individual with fake vocals from The Weeknd and Drake, garnering hundreds of thousands of plays before it was taken down. This incident caused concern for artists, creating an episode otherwise inconceivable before the existence of Gen AI. Interestingly, despite it might appear as an obvious copyright infringement, generative art does not currently qualify for copyright protection, as copyright is an intellectual property right that safeguards original work created by humans, not by software without a direct actual human author. Here is where ethics intersects with regulations.
Another significant challenge for this technology consists of data privacy, particularly for companies. Generative AI is continually trained with data, and every input could be used for generating new content in the future if a certain pattern is identified, potentially causing the leakage of sensitive data. Due to this concern, several companies have prohibited the use of ChatGPT and other types of generative AI. Samsung is among the companies that have taken this stance, prompted by the leakage of sensitive data after some employees used ChatGPT. This has led corporations to reconsider adopting generative AI, hindering the utilization of these tools in processes that could otherwise become more efficient. To address these problems, solutions are designed to enable organizations to use their own private generative AI models, trained with specific data and focused on solving particular problems.
Lastly, access to hardware and its associated costs pose a significant problem for the generative AI market. The training and running of generative AI models demands substantial computational power, with GPUs playing a crucial role in accelerating data analysis and processing. The heightened demand and rapid market growth have caused the hardware industry to lag, resulting in a shortage of chips and decelerating the development of Gen AI models. Currently, Nvidia is the largest AI hardware provider, holding 60–70% of the global GPU market. In 4Q23, the company reported GPU data center sales of US$18.4 billion, reflecting a remarkable 409% year-over-year growth. Despite the industry’s not-so-new status, the rapid growth has also created a scarcity of skilled professionals, as this technology requires advanced expertise in high-performance computing and deep learning. The million-dollar question here is how to manage AI’s energy demand. GPUs require an enormous amount of energy for training models, and it is estimated that the energy required for AI tasks is accelerating with an annual growth rate between 26% and 36%. By 2028, AI could be using more power than the entire country of Iceland did in 2021. In a time when sustainability is one of the world’s main issues, and AI is one of the most groundbreaking technologies ever created, aligning both is essential for the future of humanity.
Another tech bubble?
The hype behind Gen AI contaminated not only society, generally speaking, but also investors. As expected, new startups related to this particular technology are constantly introduced to the market. Although investments in this area are still a fraction of the total investments in AI, they have already reached a substantial size and continue to grow rapidly, as does the valuation of many of these startups.
In 2023, Venture Capital firms invested heavily in Generative AI, reaching more than $21 billion in investments, compared to the $4.3 billion investments in 2022. Despite this relevant number of investments, 67% of generative AI startups are still in the early stage, and 16% have not raised outside equity yet. Still, investors seem to rush to guarantee some participation in the generative AI segment.
Although this may seem like a new tech bubble, according to Bloomberg, generative AI is expected to be a $1.3 trillion market by 2032 at a 43% CAGR. Meta, Nvidia, Microsoft, Alphabet, and Amazon will be at the center of training for large language models. Among the different segments, generative AI interfaces are the largest market, representing 77% of funding in this space.
Even with the recent boom of the Generative AI market, some companies, still in the early stage, have achieved impressive deal sizes, even in the current market conditions. Here is a list of the top 10 Generative AI companies by deal size and their noteworthy contributions in 2023 (Pitchbook):
But venture capitalists are not the only ones after Generative AI. Companies such as Google, Amazon, and Apple are racing to become part of this market. Not by coincidence, one of the major deals last year was Microsoft´s $10 billion investment in OpenAI, the startup behind Chat-GPT and Dall-E. News of companies implementing generative AI in their processes is recurrent as industries use it to make processes more efficient and, in many cases, to avoid human mistakes.
What lies ahead
It is difficult to predict what generative AI will become, maybe as difficult as it was to predict its creation. Many surprises are yet to come, especially concerning how we interact with machines and how machines are fed. Today, public data represents just 5% of all existing data, with the remaining 95% still owned by companies. Generative AI has shown several signs it will be more than just an evolution, and rather, a revolution. In this regard, Bill Gates himself said “This is the most important tech advance in decades”, and there is still a long way for us to go.
Innovation is always happening, but megacycles like the one we are experiencing now are rare events. Interestingly, despite the transformative potential of the tools we have already created with generative AI, such as Chat GPT, Dall-E, and much more, its boundaries and practical value may lie beyond what we can dream of with the references we have today.
As a matter of curiosity, in case you have a few more minutes available, take the opportunity to check out this link of generative AI tools for different sectors created by Mindset Ventures’ team. It showcases the diverse applications and innovative solutions within the field.
