Over the last year, I’ve been following and writing about Generative AI.
That’s because Generative AI has been impossible to ignore. It’s magical. Companies like Jasper recently raised $125m to write college-level essays, Stable Diffusion raised $101m to be the next Pixar, and OpenAI, the leader in the Generative AI space is rumored to be raising another round at a $20b valuation.
This sector isn’t new to me though. Prior to becoming an investor at Expanding Capital, I led technology partnerships at Snaps, a leading Conversational AI platform, acquired by Quiq. I also founded Humin, a Conversational AI-based CRM that was acquired by Tinder.
Conversational AI always fascinated me because it enables human-like conversations between computers and humans.
Conversational AI is a precursor to Generative AI.
When I was at Snaps, we built within a deterministic artificial intelligence construct, meaning specific input and output. We were totally a “left brain” system. The system was trained on a set of data, then learned to produce the desired output for any given input. This model is what was expected of AI for the last 70 years; use things like specific data set, sensors, or finite parameters, and deliver an expected result.
But generative artificial intelligence is all “right brain” which is based on a probabilistic model that can produce a multitude of outputs. Right Brain AI feels like an exponential explosion of expectations. It’s amazing to me that an AI can create art, write poems, and make music.
One critical insight learned at Snaps was that the Conversational AI engine isn’t what made our platform unique; we leveraged open-source NLP engines like Rasa or Google’s Dialogflow. What made the platform differentiated was our context-setting engine, conversational design structure, labeling & training tools, and analytics. In other words, the value was all of the vertical-specific tooling built around an off-the-shelf NLP engine.
This is similar to what’s happening with Generative AI. Snaps was able to scale by building on top of a “ubiquitous” NLP engine. Now as an investor, I’m excited about the novel companies and also existing incumbents that will leverage Generative AI models.
To get my bearings in this sector, I’ve researched all the companies and open-sourced a comprehensive list here. In this post, I also outline my framework on model & company creation and share a TAM analysis on a few niche sectors that will create $1T worth of value.
Why it matters?
Generative AI, sometimes referred to as synthetic media, is an umbrella term for a collection of algorithms that generate new data and content.
The way that I’ve been thinking about this technology is Right Brain AI, a term I first heard here. I love this term.
The “left brain” has always been known for complex calculations and analysis; its the logical side of the brain. The “right brain” is visual, creative and is correlated more with images than math; it’s the intuitive and free-thinking side of the brain.
I’m so excited about Right Brain AI because it will enable everyone to be more creative. I also believe it will unlock enormous economic value because anyone can partner with a machine now in creative ways; this is a radical amplification of what humans can do. In the future, every company will have to have Generative AI strategy.
Generative AI might be where the next Spotify, Hubspot, or Marketo emerges.
So how did we get here?
The term “generative” was first used in 1965 by Marvin Minsky when he described a neural network called SNARC. Minsky’s SNARC was able to learn from experience through trial-and-error and predict future events with accuracy. I loved this 1998 interview of Marvin in the NYTimes. But it wasn’t until the 1990’s that neural networks were used for generative purposes when researchers started training them on large amounts of text data (e.g., books). The first commercial application was released by Parry Aftab in 1990 called “Stories That Learn” which was designed to create personalized stories based on users’ preferences.
In 2006, Google launched their first generative-like system called “Deep Dream” which was designed to create dream-like images using deep neural networks. In 2015, a paper by Google Research (Attention is All You Need) described a new neural network architecture for natural language understanding called transformers.
The Inflection Point
Prior to today, Generative AI was underwhelming, expensive, and in closed beta. But now, there are many Generative AI companies, some with open-sourced models available for developers to build on. As this space is developing, I envision a company-building and investing framework with three elements.
The Model Layer
The foundation of Generative AI is the model layer. There are two kinds of models that Generative AI is built upon. The first is a large language model (LLM). These can be trained on petabytes of data, which is the equivalent of 20 million tall filing cabinets or 500 billion pages of standard printed text. These models are used to generate something based on a few prompts, like a blog post.
Fine-tuned models go deeper and improve a model’s ability to perform a task. Fine-tuned models are good for difficult tasks with lots of training data, like generating protein synthesis or writing code.
While the model layer is the foundation, I believe the model layer will become ubiquitous, and the upside as a late stage investor is mostly gone. Google, Facebook, Amazon, Microsoft, Tencent have all developed models, other companies will open-source their own. Costs will go down and the models will be provided free, or cheap. They will become a utility larger platforms will give away in exchange for cloud hosting services.
Here are some of the leading companies in the model layer:
- Text: OpenAI GPT3, AI21, can write essays, do content moderation and power chatbots.
- Images: OpenAI Dalle-2, Stable Diffusion, Midjourney, and Google’s Imagen all create images from a text prompt.
- Code: Auto-completion for code is here. OpenAI Codex & Github partnered to launch Copilot, as do Kite, Tabnine and Replit.
- Video: Runway, Facebook’s Make-a-Video, Microsoft's X-Clip, and Phenaki all produce videos via text input.
- Music: Audiogen, AudioLM, and OpenAI are the layer for synthetic music.
- Speech: Facebook GSLM, OpenAI Whisper, and Play.HT can all generate realistic voices
- Science: Nvidia BioNeMo LLM Service and Salesforce ProGen both create protein structures.
The Application Layer
The application layer is where new companies will emerge by leveraging an off-the-shelf model. There’s incredible innovation and company creation occurring in the application layer already.
The caveat here, of course, is anyone can access the same API’s from the model layer! For companies that want to build a monopoly or competitive product, they’ll win by being vertical-focused, having a superior product, exploiting first-mover advantage, and integrating deeply with APIs already used in their niche.
Here are some emerging companies or early projects in the application layer solving specialized use cases that I’m excited about:
- Copywriting: The most covered and funded use case. Companies like Jasper.ai & Copy.ai write drafts, marketing copy.
- Advertising: Omneky is Omnicom in a box. Their product creates ads, writes copy, generates images and runs the media. Then it uses computer vision to optimize the ads.
- Customer service: Quickchat is taking Conversational AI to the next level.
- Sales: Tools like Kalendar, Lavender are challenging Hubspot. These tools create email copy and automate the sending. “SDR as a Service”
- Dating: Swiping removed the friction from matching on dating apps. Now, the friction lies in what to say. Keys analyzes the conversation and creates copy to easily start a dialog.
- Website development: Forget no-code, Debuild generates a website through text.
- Gaming: Latitude is a choose-your-own adventure game. Modbox generates NPCs.
- Data: Veezoo is a data analysis tool. Data analysis might be the most overlooked application of this technology
Image & Video:
- Interior Design: Want to know what your new office will look like? Take a picture of a blank interior and InteriorAI will design and furnish a beautiful room.
- Training videos: Onboarding in a remote world is hard. Synthesia, Tavus, Colssyan are helping companies with marketing and training videos.
- Fraud: Deepfakes will be a problem. Companies like Brighter AI annd Deepware detect fraud
- Songs: Soundful creates royalty-free music. We might see the next Spotify form music created by AI; no royalties.
- Voice synthesis: Murfi.ai, Resemble.ai and wellsaid create new voices to be used in podcasts, ads, and more.
The Augmentation Layer
Finally, these are existing companies that get supercharged when combined with Generative AI models. Businesses that realize Generative AI’s potential could find themselves with a commanding lead in their respective fields. In other words, in a Generative AI-enabled world, the biggest winners might be the incumbents. Incumbents can enhance their existing moat, preventing those in the application layer from creating clones and competing directly. For example:
- Translation: Duolingo is already the leader in language translation. Now they are using GPT-3 to provide grammar corrections. After conducting an internal study, Duolingo concluded that the use of GPT-3 led to an improvement in second-language writing skills.
- Search: Algolia uses GPT-3 in their Algolia Answers product to offer relevant, lightning-fast semantic search for their customers.
- Stock images: Shutterstock partnered with OpenAI Dalle-2 for stock photos.
- Code completion: Github uses GPT-3 in their Copilot feature, which is an AI programmer that helps users write code faster and with less work. Github Copilot can do that by applying to context in your editor and it synthesizes whole lines and entire functions of code.
- Web development: Mutiny & Unbounce are using GPT3 to create more performant web pages.
- Medical summaries: Curai is a health tech company working on expanding healthcare access and delivering the best possible care to everyone. They run a virtual primary care clinic, and leverage AI and machine learning to help their doctors work more quickly and effectively. Curai uses GPT-3 to generate training data for medical summarization.
Accounting & Legal:
- Keepertax helps freelancers automatically find tax-deductible expenses by using GPT-3 to interpret data from their bank statements into usable transaction information.
The landscape of Right Brain AI
Generative AI could easily be close to a $1 Trillion market opportunity. Below, I explore its impact on marketing, legal, and web & mobile development markets.
Historically to run a marketing campaign, a brand would need a team of around 10 people, including a creative director, a copywriter, a photographer, a media buyer, etc. With Generative AI companies are streamlining the creative workflow, creating original content, 10x faster with a click of a button, including an infinitive amount of highly targeted and personalized ads for products and services optimized across media channels, as well as data analysis and insights, designing websites, building landing pages. One good example is Omneky’s platform, which connects into each of the digital ad platforms, i.e. Meta, Google, LinkedIn, TikTok, programmatic, and ConnectedTV, and uses data and AI to generate ad creative. But if creativity is democratized, how will a platform like Omneky become the next Omnicom? I asked Hikari Senju, the CEO. He said, “These incumbents can’t have 1000 Don Drapers under one roof. New creative will be powered by data. in a world of infinite content, the value will accrue to the people who own the database, like how many people saw, clicked, defensibility. Omneky will leverage platform network effects, and the digital product will get better with more users.
The potential is simply huge. Putting in context, in 2021 a total of $250b was spent on marketing services in the United States, including sales promotion, telemarketing, event sponsorship, direct mail, directories and PR. I believe Generative AI has the potential to disrupt a big part of it, with the most impact on advertising agencies. Physical and digital advertising agencies combined generated total revenue of approximately $100b in 2022 and employed a total of 370,000 employees. Globally, the market is around 5x larger.
The total addressable market for legal services globally is set to rally after the pandemic, going from $952m in 2021 to $1.4t in 2030. The adoption of technologies like Artificial intelligence (AI) and natural language processing (NLP), and robotic process automation (RPA) has already been transforming the market, providing legal services at a much cheaper, efficient and with higher quality standards. In 2020, the global legal tech market was valued at $17.6b, with the market expected to increase to $25b in 2025.
I believe that Generative AI will play an important role in improving effectiveness across the industry but also for non-lawyers and business owners who need to understand contracts. This creates a total addressable market for legal tech services way beyond current market projections.
Another way to look at TAM is by total number of employees. Just in the United States, the number of employees in legal occupations is 1.18m in 2021, with an average annual wage of $113,000. This represents more than $130b in wages. Higher productivity driven by the application of Generative AI will have a relevant impact on this figure going forward.
Examples of use cases include using Generative AI technology to automatically draft legal documents that will follow case law and precedents, use intelligent search engine that overcomes the limitations of keyword search to identify legal concepts, translating legal text to plain English, extracting and classifying data from any legal document, automatically filling standard legal documents, among others. Casetext is a company that is applying this technology and can automate critical legal research tasks and fundamental elements of litigation, which has the potential to save attorneys a significant number of hours in research.
Web & Mobile Development TAM
Finally, Generative AI will have a tremendous impact is web and mobile development. The evolution of AI to be able to generate functional code has given rise to no-code platforms that lowered the barrier to entry for web development, making it possible for anyone to create a website or an app. Debuild for example is a Generative AI-powered low-code tool that helps you build web apps incredibly fast and it is creating an autonomous system that can write software at the level of skilled engineers.
Today in the United States there are a total of 400,000 employees in this $185b sector, with a higher concentration on app vs. web development. Globally this market is around 4x larger, and with the increase of e-commerce penetration and adoption of smartphones demand for apps is expected to pose a high growth in the coming years, making the market even bigger. In 2021 a total of 221 billion apps were downloaded worldwide, this number is expected to reach 307B in 2025, a CAGR of 8.7%.
Just the beginning
The future of Generative AI doesn’t end here. The applications and examples that I described above are just a small sliver of what’s possible in this new probabilistic world.
I recently attended TransformX, a conference about AI. Greg Brockman, CEO of OpenAI, and ScaleAI CEO, Alexandr Wang brought up a point about this technology and where it is in the hype cycle. They pointed out that after all, maybe Web3 isn’t about decentralization, maybe Web3 is AI.
Whether the future of the internet is one that’s decentralized or one enabled by AI, Right Brain AI has the potential to empower people and will change the way we interact with machines.