What’s Behind the New Wave of Generative AI Apps?
Paddy Flood, Global Sector Specialist, Public Equities:
There are multiple factors contributing to the emergence of generative AI, including:
1. Improved architecture: There are many different architectural approaches to AI, but in 2017 Google introduced a new architecture based on transformers. This architecture is an essential building block for the large language models (LLMs) we see today as it, among other things, means that models can contextualize whole questions and conversations (as opposed to words or phrases in isolation) and be trained faster.
2. Enhanced computing power: Semiconductors have become smaller and more powerful, allowing tasks to be completed faster and more efficiently. In addition, cloud computing has taken off and has enabled companies to outsource their IT infrastructure to third parties. Without this, companies would have had to invest in expensive AI-related infrastructure, potentially slowing generative AI’s adoption.
3. Data: The increased availability and usability of data, a key part of LLMs, is another reason. The world continues to generate this data, but as cloud computing advances, it becomes easier to access and store it.
4. AI at the edge: Finally, we now have techniques for deploying AI at the edge. This means the AI computations are done on the device where the data is created, rather than on a distant data center. This is crucial for applications such as autonomous driving where data instructions must be acted on immediately with no latency or delay.
What Kinds of Companies Operate in the Generative AI Segment?
Ankur Dubey, Investment Director, Private Equity:
To understand what kinds of companies are at work in the AI universe, we need to understand the technologies needed to build a generative AI application—the technology stack (FIGURE 3). There are four layers:
1. Compute Layer: Generative AI systems require large amounts of computing power and storage capacity to train and run the models. Hardware (semiconductor chips) provides the computing power and cloud platforms like Amazon Web Services, Microsoft Azure, or Google Cloud Platform provide services such as virtual machines and storage.
2. Foundational Model Layer: Foundation models are systems with broad capabilities that can be adapted to a range of different, more specific purposes. This is arguably the most important layer of the generative AI stack. These foundation models are large statistical models built using sophisticated machine learning algorithms that generate human-like responses derived from large volumes of data on which they're trained. Foundation models are split into closed- and open-source models. Closed-source software is proprietary—only the company that owns it can modify it. Alternatively, open source means the source code is publicly available and programmers can change it.
3. Infrastructure Layer: These are the tooling or infrastructure companies for apps that don’t use proprietary foundational models. Such apps need the infrastructure companies to help them fully utilize the technology available at the foundational level. Apps with proprietary models (e.g., ChatGPT) don’t need to rely on third parties in the infrastructure or the layers in foundational models.
4. Application Layer: This is the software that allows users to interact with the underlying AI technology. This can include OpenAI’s ChatGPT product or an internally built solution such as Schroders’ in-house AI product, named Genie.