Three Ways to Build Lasting AI Competitive Advantages: To Wrap or Not to Wrap

In evaluating an AI product, a question frequently asked is whether a foundational model will eventually offer similar AI responses as your products. Often the question is posed as whether your product is a "wrapper on top of a foundational large language model (LLM) platform."

In responding to such a question, it can be effective to reframe the discussion to the sources of enduring competitive AI advantage in your product.

Why Now?

With the vast capability of LLMs, in many ways it has never been easier to build an AI application. The base corpus of knowledge within computing systems and its accessibility has dramatically increased.  Indeed, fine tuning existing foundational models with reinforcement learning from human feedback (RLHF) can dramatically increase the core competency of your product's platform without the same investments of building a foundational model. Yet at the same time, foundational models continually improve including in areas which may be on your product's roadmap.

Three Ways to Build Lasting AI Competitive Advantage

1.      Focus on Collecting 1st Party Data and Building a Proprietary Data Moat

While capturing data directly from your customers and feeding it to an LLM can generate a quick response, storing the data along with the context including multimodal inputs can help keep your application more capable across foundational model platforms and in more focused contexts.  Second, making sure to use enterprise ready versions of LLMs can help ensure that your data is used for improvements only to your models versus the generally accessible foundational model. Third, managing which information is behind paywalls and firewalls can limit exposing valuable data sets to foundational models.

2.      Chain and Leverage Multiple Models

By appropriately leveraging multiple LLM models, you can build your product to be more capable than a single foundational model. Additionally, this approach can separate aspects of your data from any one specific LLM. This is an area of much debate still. The recently released GPT 4o is an example of an integrated multimodal model though in several other instances video, text, and voice models are all often separately trained models that are used together in an output.

3.      Build Workflows & Agents

Products that deliver a compelling end to end experience can help drive customers to continue to use your application and further differentiate the product from an LLM platform. The product user interface should make your customers more efficient versus an LLM chatbot or conversational experience.

This is a purposely short article to be a conversation starter on a given topic. Tau Ventures is an early stage AI focused venture capital firm. More at https://www.tauventures.com.