A proprietary dataset

“If a company has proprietary data on top of which it builds an AI programme to interpret the data, then the company’s AI-enabled proposition has defensible value.” — Sulabh Soral, Chief AI Officer at Deloitte

Data is typically considered a moat for any business, something which has up to this point established the dominance of a small number of Big Tech firms. With regards to AI, data remains a huge differentiating factor—specifically, the dataset on top of which you build your foundational model, and your strategy for protecting that data. 

There are a number of key factors to consider within this. First of all, does your dataset give you unique insights that no one else has? This may be related to your unique scale or penetration into a certain industry, or it may be that the dataset you have is particularly difficult to gather. 

A ‘wrapper’ on top of a foundational model, such as ChatGPT, does not have this data defensibility since there is nothing stopping someone else building the same tool quickly and efficiently without having to invest into gathering data. These business must find defensibility elsewhere.

Built AI is one startup leveraging the benefit of a proprietary data model, having created a detailed digital twin of the world’s real estate market to help drive decisions in the commercial property industry. Similarly, Qumata offers a new standard for life and health insurance underwriting through its vast proprietary health data set that can accurately calculate risks for any individual and over 800 medical conditions. 

It’s also important to consider how the value of this dataset may evolve. On the individual level, it may be this improves over time for each user (for instance, with a model that is trained or tuned to the behaviour of each user). On a model level, it may grow as the number of users grows (a classic network effect). Reflect on what it is that you are building and whether you can build this defensibility into it. Not all businesses will have this opportunity: models that understand and predict behaviour can constantly fine tune, as behaviour changes over time; whereas speech recognition has a ceiling of how much it can improve.

Feedback loops (where data is fed back in and improves the model) using AI generated outputs, moreover, don’t really behave in the same way that they might with other data sets. On the contrary, some have raised alarms around the risks of training a model on model-generated content, leading to potential model collapse. 

Build deep tech at the core of your IP

Another aspect of defensibility is the technology itself—is your product or IP of unique value that means it's doing something that no one else can do? 

There are a number of paths towards this. Building your own foundational model means that your tech has a key differentiation point that is hard to replicate. This is particularly true if you are developing new models with a difference (e.g. filling gaps from other models). Mistral, for instance, has an open source lens to its model; Poolside, meanwhile, aims to build a tool that can write software using natural language. Along the lines of going vertical, not horizontal, founders may explore opportunities to build foundational models for different domains. 

Another route to defensibility is building an AI system that automates the discovery of solutions in complex deep tech spaces, such as biotech or material sciences. Take Peptone, a startup using AI to discover new drugs for complex disordered proteins that typically might take decades to discover. Or there’s Materials Nexus, a startup accelerating the R&D process for ‘next-gen’ materials using AI and quantum computing. 

If you are building in a complex, deeptech space, your team brings enormous value to your proposition. Having domain expertise, particularly in a scientific space, adds weight behind your business, which will help enormously when it comes not just to your engineering process but also to sales, fundraising, and so on. 

Look beyond generative AI

Amid the hype around generative models, it may sound glaringly obvious but founders may be better served taking the path less travelled. 

Consider what it is that generative AI does, and more importantly, what it can’t do. There are still enormous problems to be solved, and a number of strong machine learning techniques that can help solve these.

Broadly, there are two types of model—generative and discriminative. Generative models take huge swathes of data and information and apply them to generate new data points: for instance, large language models that are capable of producing text (ChatGPT) or diffusion models used to generate images (Stable Diffusion). 

Discriminative models, meanwhile, are able to distinguish between the different data points and draw boundaries between them using probability estimates and maximum likelihood. There are numerous opportunities here that generative AI cannot touch—including regression models, classification, forecasting, recommendations, computer vision, and so on—where building a proprietary ML model can set you apart from the competition. 

Over the next few months, as some of the hype around Gen AI dies, expect to see the next wave of AI companies coming along—those applying innovative technology to interesting problems will be best positioned.