A better way to build an AI business
New AI capabilities tend to evolve through three main phases: from pure research to usable technology demos, and finally into productized businesses. Once they reach this last phase, they seem to diverge along one of two paths:
- Large, often very well-funded products that try to solve many problems for many people using a single, powerful model.
- Vertical-oriented solutions that connect multiple models, but focus on a narrow user, problem, or market segment.
This article makes a case that the latter will prove to be a better, more profitable model for a significant number of emerging AI startups.
Research and demo phase
At first, a new AI technology is something that we read research papers on and see the results of, but aren’t able to play with ourselves - especially if we don’t have the technical know-how or resources to set it up ourselves. Dall-E is a well-known example of a project that was recently in this phase when they were sharing images that they were generating but not providing access for anyone else to do so.
Before long, you start to see public demos (often an official one from the researchers) and single-purpose products that are wrapped in more user-friendly interfaces, making it possible for anyone to play with them.
- Dall-E now gives anyone (at least after being on a waiting list) easy access to providing prompts and generating images through their web portal.
- DreamStudio for Stable Diffusion provides a similar interface, but Stable Diffusion goes further in that you are also able to download and run the model yourself - as long as you are fairly technical.
After the research and demo phases, many projects then form a commercial entity that tries to productize their tech. Sometimes this evolves directly from the team that created the demo in the previous phase, but in the case of public/open source models, it can be anyone. One big example of a company that is productizing a single model is Microsoft with GitHub Copilot, which can generate code based on its training set.
What’s common with these sorts of companies is that they are trying to solve a broad swathe of problems with a single model, and although they can do amazing things, it can sometimes feel like they are a hammer in search of a nail - they have built something powerful, but applying it consistently and generally is quite hard. It can seem like they do a lot of different things, but none of them deeply enough to solve any single use case.
An alternative approach to going wide with your model/business is to integrate multiple AI technologies and research areas into a single product, focused not on general applicability but on solving problems associated with some quite specific niche.
This approach is being made possible thanks to the emergence of powerful models with quite permissive licenses, such as Stable Diffusion; you get an almost open source development experience, where you can integrate various components (models in this case) to build a complete, novel application.
Companies like DhiWise and Locofy are applying similar models as GitHub, but are doing it within more specific use case verticals (React and Flutter apps for the former, and converting designs into code for the latter). Synura is also taking this approach to integrating multiple research areas into a single application, focused squarely on making it easier to plan and create videos with AI assistance.
A better way
Focusing on a specific kind of user or a specific class of problem allows these companies to create deeper, more compelling solutions that paper over some of the rougher edges that you run into when applying similar AI tech to more general problems. The model doesn’t need to do everything anymore, it just needs to be part of a suite of well-integrated tools.
This combination of tailored, problem-specific solutions with AI assistance is a much more compelling product, solves real problems for real people today, and avoids being blocked by difficult edge cases in applying a model more generally. We should expect to see a lot of these kinds of companies appear and thrive soon, especially as we see more and more open models and innovation happening.