Foundation Models: How Startups Win With Generative AI

Foundation Model - How Startups Can Win With Generative AI

Last month, we saw a whirlwind of activity surrounding Microsoft and Google’s big reveals — Microsoft’s Sydney in Bing and Google’s Bard. While it’s still early days, these products have demonstrated significant advancements in generative AI and provided the world glimpses of what it can enable for the enterprise. Despite the high pace of activity from well-capitalized Big Tech giants, we believe there are several ways startups can create value and utilize generative AI to build meaningful solutions. Three angles of attack are — consumption-based business models, generative native features, and user-centric experiences.

Disrupt Business Models

In our article “Who’s Making the Elephant Dance, we called out how Microsoft has leaned into generative features to reinvent Bing. Microsoft not only re-envisioned search but also disrupted the existing search-based business model. Microsoft took a calculated risk — leveraging its lower market share and diversified business to change the gross margin of search, creating an experience that, so far, Google can’t afford to match at scale. Because the generative models (sometimes called large language models or foundation models) are so large, generative features are expensive to run in production. The increased inference cost drastically increases COGS and reduces gross margin, making incumbents/leaders less likely to adopt them.

Use a Consumption Pricing Model

Startups can leverage that same type of underdog positioning to disrupt SaaS applications of yesterday with consumption-based business models that price in the variable inference costs. Companies like Jasper and RunwayML have implemented such pricing models, increasing prices based on the number of words and images generated, respectively.

Embrace Generative Features

As our colleague Soma Somasegar wrote in January, even before the Microsoft and Google announcements, there has been a rise in broad-based generative native applications, such as ChatGPT, Jasper, Runway, and Lensa, getting a lot of buzz, usage, and media coverage. The hype over generative AI was and continues to be palpable, and with that, consumer excitement has never been higher. That eagerness of users to try out tools is something startups can use to their advantage.

Startups don’t have to have a team of technical experts to build generative AI from scratch. They can utilize foundation models (GPT-3, Co:here, Stable Diffusion) that are already available to enhance their platforms with generative features that will make their users’ lives easier. Some of our portfolio companies have become generative-enhanced — Coda recently announced an OpenAI-enabled virtual assistant called Coda AI that summarizes meeting notes and transcripts. And Lexion launched a GPT-3-enabled word add-on to help draft, execute, and redline contracts faster.

Utilize best-in-class tooling

When it comes to tooling, we talked a lot about that in a previous post that you can read here. But it is a quickly evolving space, and orchestration tools like Fixie, Langchain, Llama Index, and Promptable make incorporating large language models and data into your products easier, so startups should embrace them. And, at the model layer, companies like OpenAI and Co:here are making these models easier to work with. For those with the technical know-how, consider working directly with open-source models.

Create a moat

Startups need to build a moat for their products by improving the underlying generative AI they’re using. To do this, you can: modify your prompts, fine-tune on domain-specific data, or change the underlying model. But how do you ensure the product is improving? Evaluate model outputs and see which ones are better. We have seen successful startups follow this pattern: First, launch with a baseline model, monitor how users respond and what they choose, and track what they’re clicking on and accepting. Then, update the model, experiment, and see how usage changes. Use this pattern to create a flywheel to improve your product — and store that proprietary data, creating a moat.

Re-envision User Experiences

When it comes to existing incumbents in any sector, they are often constrained by their own user experience and can only innovate at the margins. They may also not want to reinvent their UX for fear of disrupting the workflows of existing users. That leaves the door open for startups to step in with these innovative new generative features that users are looking for and excited about.

Runway, which offers a content creation suite and is one of our portfolio companies, recently introduced a new generative model Gen-1, which lets users create new videos out of existing videos with natural language and images. That’s a user experience that Adobe Premiere and Final Cut might want to take note of.

Empower the human expert

Today, state-of-the-art AI systems cannot automate complete workflows. While foundation models are almost there and will get there eventually, startups need to build their UX assuming they’ll have a human in the loop. This generative AI should be assisting and empowering the use of “experts,” not replacing them. For example, Jasper does not automate writing a blog post; it gives the writer suggestions that make them more productive. Tome does not build an entire presentation; it provides the user with options they can play around with. Yoodli is not trying to replace a speech coach; it is making a speech coach more productive.

Get in Touch

We are thrilled to see the innovation this generative AI craze has and will continue to spark. At Madrona, we have significant experience in this space — Aseem led growth for Azure and worked on the early team that led to the investment into OpenAI, and Palak worked on the Microsoft Turing project, the first effort to productize foundation models across the Microsoft portfolio. We love rolling up our sleeves and partnering with companies interested in integrating generative models into their offerings and leveraging them in new and creative ways. If you’re building generative-native or generative-enhanced applications, please get in touch at [email protected] and [email protected].

Related Insights

    Our View on the Foundation Model Stack
    Who’s Making the Elephant Dance: Microsoft and OpenAI Get a Leg Up on Google
    Generative AI — Overhyped or Underhyped?

Related Insights

    Our View on the Foundation Model Stack
    Who’s Making the Elephant Dance: Microsoft and OpenAI Get a Leg Up on Google
    Generative AI — Overhyped or Underhyped?