Part 1: Generative AI 101: Why you should care
Thanks to generative AI, we are on the precipice of a global-shift in technological capabilities. Most investors will agree that we are in an opportunistic investing period and therefore are spending significant time and resources understanding how to better evaluate these opportunities. Most will also agree that there is no thesis that will likely hold exactly true in the long term given the nascent, dynamic nature of the space. For that reason, there are many compelling questions around the underlying technology, evaluation methodology, and areas of importance that we will explore in this three-part series.
What is generative AI?
At a high level, generative artificial intelligence (AI) is a type of computer algorithm that is able to generate net-new content based on a pre-programmed data model. Generative AI outputs can be in the form of sound, computer code, images, text, animations, 3D models and other types of content.
The data models generative AI leverages to create these outputs are referred to as “large language models” (LLMs). LLMs are massive sets of data with often more than 1 billion parameters that consist of content from books, articles, websites, computer code and various other sources. Large language models are then “trained” to learn the statistical relationships between words, phrases, numbers, code and sentences, allowing them to generate coherent and contextually relevant responses when given a prompt.
What does the generative AI tech stack look like?
Generally speaking, the AI tech stack consists of three different layers that all serve different functions. The layers include: the model layer, the application layer and the operating system and infrastructure layer.
1. Model layer — the underlying models that form the foundation of the technology. There are many different types of generative AI models, but they are often categorized as either general models, specific models, or hyperlocal models.
General AI Models: Also often known as open source models, these are LLMs that are released under a license where the copyright holder grants users the rights to use, study, change, and distribute the model to anyone and for any purpose.
Some of the most well known general AI models include:
- GPT-3: Text-to-text outputs. Best known as the model that ChatGPT leverages to provide answers to all of our questions.
- Stable Diffusion: Text-to-image outputs.
- Whisper: Speech-to-text outputs.
Specific AI models: Often built on-top of general models, specific AI models are trained on more narrow, more specialized data, such as tweets, ads, song lyrics or images, which should allow them to outperform general models in their industry.
Hyperlocal AI models: Trained on exclusively local-proprietary data with the hopes of outperforming the general models for specified use cases. Examples of hyperlocal training data include ingesting code in the particular style of an engineering team or tracking design images suited to a specific aesthetic.
2. Application layer — This is the interface and workflow built to leverage the model layer, which enables end-users to leverage, interact with, and collaborate with the AI easily. Applications serve as essential tools, making AI models accessible and easy to use for both businesses and consumers.
3. Operating Systems and Infrastructure layer — This layer often does not get the same attention, but it is just as important. The infrastructure layer refers to the tools or platforms used by developers to access and utilize LLMs to create applications. Many also refer to this layer as the picks and shovels.
What are the implications of this technology?
Now that we have discussed some of the basic technicalities of generative AI, it is important to understand why this technology is so important. In short, every single industry is going to be disrupted by AI. Some early application layer use cases across both blue collar and white-collar industries have shown their promise through rapid adoption and high praise.
For example, in a recent issue of Neal’s Deals, a Ground Up newsletter on early-stage venture news, we covered how countless prominent global law firms have publicly announced their plans to integrate AI into their day-to-day. This is due to the technology’s ability to engage in sophisticated writing and research faster and with greater accuracy than humans. For that reason, many AI companies like CoCounsel, Luminance, Harvey, TermScout and Latch have launched products to assist with litigation preparation, drafting motions, analyzing trial transcripts and developing contracts.
Other sectors, like the music industry, are not as thrilled about the AI takeover, given that humans will soon not be able to tell the difference between AI and human-generated music. They are now grappling with the reality that anyone can create new music with an existing artist’s voice. Universal Music Group, the music label representing Drake and The Weeknd, has been urging streaming platforms to block AI generated content as they have “fundamental legal and ethical responsibility to prevent the use of their services in ways that harm artists.”
Market Maps
There is no shortage of building occurring in each of the three layers of the generative AI tech stack. At Ground Up Ventures, the majority of the opportunities we evaluate are within the application layer. Regardless, the ~50 company market map below is indicative of a lot of the building that is going on in early-stage generative AI.
Application Layer Market Map — Ground Up Ventures
It has become increasingly challenging to evaluate these opportunities and understand where value will accrue, given the ever-shifting use cases and updates to the technology. In our next post, we will discuss Ground Up’s perspective on how to evaluate generative AI startups and which areas of the technology and business model are most important to double click into.
