AI infrastructure gives way to AI-Native Product Design

My most recent LP update...does Multi-Modal == Single Winner?

I wrote the below in my most recent update to our investors at?Factorial?and a few responded that it could be worth sharing it a bit more publicly. I’d love to hear if you have comments or perspectives on it. For context, over the past year, we have spent a lot of time investing in companies building out the infrastructure for NLU Models (natural language understanding), such as Flower Labs , Nomic AI , Patronus AI , Modal , Adaptive ML ,? Modelbit ,? Substrate , and LanceDB .

With the proliferation of this infrastructure, we’re now starting to see real AI-native product design, led by teams at portfolio companies like Pika and Factory .

Note: subscribe to the original substack where I published this here.

Here are the observations I shared, I'd love your perspective as well...

Many people still think about AI as ChatGPT. A year ago that was even more true, and there was still a lot to build in terms of infrastructure and there was broad value to create for developers from vector database management to tooling that enabled continuous training of AI models. Today, many of those infrastructure products now have strong traction with developers, so new products will likely have a higher bar to adoption. When it comes to new pre seed and seed stage companies, I’m still excited about certain areas within the AI developer tool / infrastructure world, but am mindful of the increasing competition and friction in the category broadly.

AI Infrastructure enables AI-native products

One learning from the past year or so is that generative models appear to work across all media types where they’ve been tested (words, images, movies, etc). There is also emerging experimentation around inputting the models themselves to output other models (a pretty cool idea).

Does Multi-Modal == Single Winner?

This phenomenon that models work for all different types of inputs leads some to believe that since you can train a model as easily on language as you can on images, the biggest model training companies (read: OpenAI) will dominate across many different use cases -- from chat to video to healthcare applications. Said differently: the ability to train models outside of a company's initial domain is easy therefore more specialized companies won’t succeed. This seems unlikely to me.

Because AI development is so new, the entire mode of how we interact with them is still up for grabs. As an example, GPT3 has been around since 2021 or so, originally as what they called the "Playground" which was a blank text box that completed whatever sentence you wrote in it. However it wasn’t until they reframed the product as a chat interface that it really took off.

AI-Native Product Design

Interfaces will likely need to be re-thought — some are calling this AI-Native Product Design. This approach will need to contemplate both how we receive information and how we give feedback, creating a "conversation" with the model about its own output. A chat user interface paradigm is certainly one approach but it's hard to believe that the end state for all AI interfaces is chat.

Chat is often inefficient -- for example, it's often useful to point to something. What is the digital equivalent of that? How do we communicate time if we're working on a video? In the same way that the mouse unlocked a new design paradigm for the computer, I think we'll see similar innovations that make sense only in the context of the AI models we want to apply and manipulate.

If interfaces were only about perceiving data, maybe that wouldn't matter. But when user feedback can fine-tune models, or even just give light feedback and direction on a current output, it means that the interface and the model itself go hand-in-hand. New models are released frequently, often as open source, and depending on the product those can be quickly incorporated as a sort of “product feature” if an AI-first team is thinking of it that way.

We’re already starting to see this pattern emerge: team builds a core enabling model for its use case, then can quickly iterate on ancillary models which serve as features to the core product. The sum of these parts is an AI-native product.

We’re in the earliest stages of discovering what successful AI-Native Product Design looks like. Given how different so many of these AI modalities may be (text, images, code, video, etc), and their use cases, it seems probable that we’ll want specialized ways of interacting with each for the best user experience.

It’s an exciting time to be investing in deeply technical companies that are reinventing how we understand and interact with information.

-- Matt

You can find me on Twitter, LinkedIn, & Substack. Here's the original substack post.