Generative AI

Generative AI is the shiny new thing in tech. Creating content is cheap and fast, and people can’t get enough of it.

AI content developer, Jasper AI, raised $125 million at a $1.5 billion valuation. Shutterstock, the stock image platform, plans to let users generate images using DALL-E. Pieter Levels launched a generative AI project that made $10K in a single day.

Dall-E image for “A person sitting on a computer typing prompts into an AI application that writes social media posts for Instagram”


Anyone can build a generative AI product, but not everyone can build a generative AI business. As technology becomes accessible, it becomes harder to build a defensible business - we’re going to talk about how you can build on.

Generative AI uses artificial intelligence to produce content. You give the software an input (e.g. text) and you receive content that could be text (Jasper), images ( or some other form of content.

This is a massive over-simplification but it works for the purpose of this piece. If you’re interested in the specifics of how this happens, I recommend reading this about Dall-E (generates images based on text input you provide).

It’s useful to think about generative AI as three different layers:

  • Training the model: companies like Open AI or Stable Diffusion develop models by using very large datasets. They expose their model as an API, which means any one can submit certain inputs and receive the output (image, code, text etc).
  • Tuning the model: a company takes one of the models above and tunes it for a specific use case. For example, Jasper trains OpenAI’s GPT3 model to write copy based on inputs provided to the model.
  • Deploying to the end user: the company then uses the tuned model within its application for the end user. In the example below, Copy AI helps you generate product descriptions using a few inputs.
  • Copy AI helps you generate product descriptions using AI


Training models is expensive. You need vast amounts of data and computing power. My view is that this area will be dominated by a few companies like Open AI, Stable Diffusion and so on.

Tuning and deploying to the end user is where most of the action will happen. The market map below (h/t Sequoia) provides a good overview of generative AI landscape.


There will be many useful products that come out of this space. My optimism for this space comes down to the following:

  1. AI is more accessible than ever.
  2. The cost of producing content becomes very cheap. With the right tool and right prompts, you can generate content quickly. No one wants to spend time on repetitive tasks like product descriptions.
  3. It gives you the ability to do things you could not do previously. For example, you have a creative idea but aren’t the best visual artist — DALL-E is here to help. Note however that you still need to be able to provide the right prompts.

If you’re building in this space, there are two very important questions to ask yourself:

  1. What stops a new company from entering the market and winning my customers?
  2. What stops an incumbent from integrating AI into their product?

Building on my piece a few weeks ago about moats in B2B SaaS companies, you can build defensibility by focussing on the following:

The use case

The use case that a product solves for is going to be very important.

Elad Gil published an excellent essay on why value might accrue to startups rather than incumbents in this wave of AI. In doing so, he recommends looking at opportunities that are:

  • Highly repetitive
  • Highly paid
  • Imperfect fidelity is acceptable
  • A digitised workflow does not exist yet

Repeatability is required because that is where AI can save the most time. Highly paid because this means saving time actually leads to lower costs, a clear win for the user. Imperfect fidelity because everyone accepts that no AI algorithm can provide a perfect outcome yet (see the image header for this post if you have any doubts). Human intervention is required in every case. And finally, use cases where a digitised workflow has previously been impossible but is now possible because of AI, are ripe for disruption.

Choosing the right use case is important. The best products will use this technology to provide an outcome that is 10x better rather than 1 - 2x better.

Data as a moat

The companies that capture the most value will use data as a moat. I don’t think every company will enjoy this benefit. In fact, many won’t.

Products that build a data moat will need to achieve at least one of these two things:

Value increases with time for every user

If the value of your product increases with time spent, you have a natural moat.

With Slack, I’d find it difficult to move away because I lose my message history. Try and find the equivalent for your generative AI product.

Users need to recognise that they will lose some amount of value if they switch away. For example, the model you use could be tuned for that specific customer — and if they moved, they’d need to start from scratch. Suddenly you have a cost to switch and start to develop a moat.

Value increases with the number of users

If the product is able to offer a better experience because you have many users, you’ve got a moat.

This is a data product’s network effects. Take Github Copilot, they have great data on what code works (because it’s committed to Github). Across many many users they can see what users like, and what they don’t like. With more users on the platform, Copilot becomes better. I write better code because Copilot provides suggestions using code from other developers. Tough to build a competing product when you’ve got Copilot that’s been used and optimised using millions of users.

Go vertical, not horizontal

Products that go vertical will capture a lot of value.

This won’t be because of AI alone. It might be because they are integrated deeply into a specific vertical or niche. The generative AI element may be a small, but important part of the product.

Prompt engineering

Most generative AI products depend on the inputs provided to the product.

Products that help users improve their prompts to achieve their desired outcome will capture a lot of value. Eventually, the best ones might not require a prompt at all, though this seems pretty far away.

A corollary here is building in sequence with the tech. Today, it’s generally agreed that generative AI is not good enough to “write for you”. I don’t know if it will ever be. The best products will focus on bringing value to the user with the least amount of friction. They will approach this dynamically and innovate as the tech improves.

In other words, you want a viral loop: user inputs → AI generates → user corrects → you capture this feedback and improve your product.

To close

Generative AI is here to stay. If you’re building in the space, expect lots of competition. When you think about incumbents, it’s useful to ask how or why they cannot build a better solution. Incumbents have a dataset that is several times bigger than yours, but they may not move as fast as you do. Separately, think about why a new entrant cant copy you.

On the flip side, don’t let the thought of moats and defensibility stop you. It takes time to build a moat and no one has it from day zero. Think about how you might develop one over time, and get started.